Review of Seattle’s New Water Demand Model
Dr. Richard Palmer,
Austin Polebitski, Lee Traynham,
Kathleen King, and Ben Enfield
Department of Civil and
Environmental Engineering
University of Washington
Box 352700
Seattle WA 98195-2700
For
King County Department of
Natural Resources and Parks
July 07, 2006
Table of Contents
Introduction ................................................................................................... 2
Why forecast? ................................................................................................ 3
Long-Term Water Demand Forecasting Methods .................................... 4
The SPU Regional Demand Forecast Model .............................................. 7
Model Processes and Assumptions ............................................................ 11
Process 1: Weather Adjusted Base Year; Consumption by Sector and Service Area11
Process 2: Base Year Flow Factors by Sector and Service Area ............................. 13
Process 3: Forecast Flow Factors by Sector and Service Area Adjusted for Income Effect
................................................................................................................................... 13
Process 4: Intermediate Forecast of Retail Demand by Sector and Service Area ... 15
Process 5: Sector Forecast Aggregated to Larger Service Areas (LSA): Seattle, Non-CWA,
and CWA .................................................................................................................. 16
Process 6: Forecast of Net Retail Demand Aggregated Across Sectors .................. 18
Process 7: Forecast of Total System Demand.......................................................... 19
Model Assumptions .................................................................................................. 21
Observations and Recommendations ........................................................ 25
References and other Useful Documents .................................................. 27
Appendix 1 – Metropolitan Water Demand Forecasting Models ............ 1
Tacoma Water Demand Model ................................................................................... 3
Everett’s Water Demand Forecasting Model .............................................................. 7
Portland, OR Water Demand Model ......................................................................... 10
Washington Metropolitan Area Water Demand Model ............................................ 14
San Francisco, CA Water Demand Model ................................................................ 17
Appendix 2 – Overview of UrbanSim ......................................................... 1
List of Figures and Tables
Figure 1 – Past Water Demands, Actual Water Demands and Population Served. ........................ 3
Figure 2: Seattle Demands – Past and Present (1980 Study) .......................................................... 8
Figure 3: Seattle Total System Demands – Past and Present (2001 Study).................................... 9
Figure 4: Seattle Total System Demands – Past and Present (Current Demand Study) ............... 10
Figure 5: Seven Water Demand Forecast Model Processes ......................................................... 12
Figure 6: Forecast Factor Generation Process for Residential and Non-Residential Demands ... 14
Table 1: Savings from Conservation Potential Assessment Model .............................................. 17
Table 2: Seattle Demand Model Sheets Associated With Each Process ...................................... 20
Figure 7: Seattle Demand Forecast Model.................................................................................... 21
Table 3: Numerical Assumptions within the Seattle Demand Model........................................... 22
Table 4: Sample Calculation of Water Demand ........................................................................... 24
Table of Contents – Page i
Introduction
This paper reviews Seattle Public Utilities’ (SPU) recently developed water demand
model. This model projects future water demand within SPU’s service area, which currently
includes the City of Seattle, the Cities of Bellevue, Bothell, Duvall, Edmonds, Kirkland, Mercer
Island, Redmond, Renton, and Tukwila, and a number of water districts. Regional forecasts of
future water demand provide a basis upon which to evaluate the adequacy of available regional
water supply. Although neither future water demands nor future water supply availability is
known with certainty, these forecasts do provide a framework of upon which informed decisions
are made.
Forecasts of water demand in the Puget Sound, and the methods employed to generate the
demands, are of particular relevance today. Local water supply utilities, resource agencies,
county and city governments, and tribal representatives agreed in 2006 to participate in a
voluntary planning process (noted as the Regional Planning Process). This process will identify
current water supply and demand issues and evaluate the region’s ability to deal with its wide
variety of system opportunities and challenges. Of particular interest in these evaluations is
estimating the region’s future water demand and water supply, the potential impacts that climate
change will have on the region, the role that water reclamation might play in bridging the gap
between supply and demand, and estimating the regional waste water requirements. Forecasts of
future water demand will help define the region’s perceived needs and will significantly
influence the preferred alternatives identified in the planning process.
Interest in the forecasts of water demand also arise because of the region’s inability to
accurately forecast water demands in the past. Figure 1 presents a summary of three past water
demands produced by SPU, the actual water demand that occurred, and the population for the
region’s service area. Forecasts made in the 1970’s, 1980’s and 1990’s greatly over-estimated
the actual demand for water, primarily due to unanticipated reductions in per capita water
demand. Rather than per capita water demand increasing over time, as had been forecasted, the
per capita water consumption of SPU’s customers has decreased. Our ability to forecast water
demands in the future is a function of not only estimating future regional population, but how
this population will respond to changes in the price of water, the extent to which water efficient
appliances enter the home, and changes in regional preferences in using water.
This report is divided into five sections. The next section provides more background in
why forecasts are created and how they are used. Next, general background is provided that
describes the basic approaches that have been used to estimate water demand. To provide
context, the general approach used in the SPU model is presented, together will descriptions of
the demand models used by Everett and Tacoma. Next, the new SPU model is reviewed in detail
by describing the model’s components and assumptions. Finally, strengths, concerns and
limitations of the model are identified.
As suggested, accurate water demand forecasting is an essential ingredient in water
resources planning. This challenge is one faced by the profession, not just the Pacific Northwest.
As the value of water has increased over time, more attention has been placed on how best to
estimate future water demands and what actions can be implemented to manage demand. In a
recent literature review of water demand forecasting techniques Arbués et al. (2003) noted that
“there has been a change of approach to water management,” one that is less supply-side driven
and which focuses on the factors that impact water demand. In their article, citing over 110
refereed publications, Arbués et al. (2003) note that the most important factors influencing water
Review of Seattle Demand Model – Page 2
1,600
350
1,400
300
1,200
250
1,000
200
800
150
600
1960 Water Plan
1980 Water Plan
1990 Water Plan
Actual Water Demand
Population
100
50
0
1970
1975
1980
1985
1990
1995
2000
2005
400
Population Served, in 1000's
MGD (Million gallons per day)
400
200
0
2010
Year
Figure 1 – Past Water Demands, Actual Water Demands and Population Served.
demand include residential population, industrial use, housing characteristics, frequency of
billing, price, income, weather variables, rate structure, and indoor versus outdoor water use.
Howe and Linaweaver (1967) first demonstrated that water demand is not simply a function of
the number of residential customers and industrial water users. Their work, like many other
studies, detailed the impacts of price on demand and began the formalization of mathematical
approaches to forecast water demands (Carver and Boland 1980). Since the 1960s, water
demand forecasting has emerged as a distinct field, with mathematical models of water demand
replacing simple engineering extrapolations of demand over time (Baumann, Boland, and
Hanemann 1997).
Why forecast?
Accurate forecasts help ensure that an appropriate amount of water supply is available
when it is needed and that it will be provided in an efficient and timely fashion. Utilities forecast
water demand for a variety of reasons. These reasons include: 1) creating an accurate estimate of
future water demands for purposes of supply reliability, 2) estimating the cost of providing water
in the future, and 3) establishing current and future rate structures to balance costs and revenues.
All of these applications are best served by an accurate forecast. Accurate forecasts
allow utilities to provide safe and ample water at minimum cost. Overestimates of future water
demands can result in premature capital investments and stranded resources. Historically, it has
been common to overestimate water demands (Gleick et. al. 2004). Recent studies have
demonstrated this tendency throughout the U.S. and Figure 1 illustrates this phenomenon for the
Review of Seattle Demand Model – Page 3
City of Seattle. During the 1950’s and 1960’s, it was common in the US to use a declining block
structure (the more water used, the lower the unit price) and average cost pricing. Both of these
practices typically under valued water and encouraged excessive use of water.
Because underestimating future water demands can create unnecessary shortfalls and
limit the growth of communities, there appears in the past to be a significant professional bias to
overestimate water demands. Underestimating demand and/or overestimating supply can result
in the need for voluntary or mandatory restrictions and in the most unfortunate of cases, can
create fire and public health challenges. The American Water Works Association (AWWA,
1996) provides guidelines for generating water demand forecasts. It should be noted that the
AWWA also recommends a very high level of supply reliability (98%). The tendency to error
on the side of forecasting demands higher than those that occur may be, in part, a professional
response to ensure that demand does not overtake supply and thus violate the 98% reliability
level.
Long-Term Water Demand Forecasting Methods
Overview
A forecast of future water demands is essential in water resources planning—not only for
preparing water supplies to meet anticipated future demands, but also for understanding what
influences water demand. An accurate forecast indicates the quantity of water needed for the
future; however, demand forecasting is an evolving art, and there is no single forecast method
that is appropriate for all settings. Many forecasting methods have evolved to address
demography, economies, and social attitudes of a region. Methods for forecasting water
demands have been evolving since mid-1960s (Gottlieb, 1963; Howe and Linaweaver, 1967).
Prior to that, water demand was modeled as a rudimentary function of the number of residential
and industrial water users.
There are two distinct types of water demands: long-term and short-term. Short-term
demand forecasts often project water demand within a single year, season, or even week. Longterm forecasts, however, commonly look at timeframes of ten to fifty years. Forecasts in excess
of fifty years are typically created for general guidance, as the uncertainties associated with such
forecasts are great. This demand forecasting review focuses on water demand forecasting
techniques appropriate for long-term scenarios (approximately 30 years).
Consideration of water conservation is essential because of its implications on future
water use. Water conservation, defined by Baumann et al. (1980) as “the socially beneficial
reduction of water use or water loss,” has played an important role in reducing water demands in
recent decades. However, water conservation effects are now reflected in recent (since the early
1990s) historic water-use data, adding another layer of complexity to water demand analysis.
The nature of water conservation makes it difficult to predict; the timing and extent of
conservation are a function of policy, public education, and attitude (political/social feasibility),
in addition to the “hardening” of demand as water use efficiency increases. Household water use
is related to more durable appliances such as toilets, dishwashers, and laundry machines. Even if
advances in technology, changes in plumbing codes, the price of water, or public awareness
encourages increased water conservation, the delay in installing water-saving appliances may
cause a time-lag, making the task of simulating the casual relationship more difficult.
Review of Seattle Demand Model – Page 4
Forecasting Methods
Demand forecasting methods currently range from the relatively simple to the highly
complex. The most common forecasting methods include:
(i)
Trending
Trending, or trend analysis, is a simple method that forecasts water demand by fitting a
trend line to historic data and extrapolating this trend into the future. This method may be
appropriate for small utilities for short-term projections, but can grossly over estimate water
demands for long-term forecasts.
(ii)
Per Capita
Per capita demand forecasting requires data on current per capita water usage and applies
it to a population forecast. The change in demand is proportional to a change in population.
This method does not consider changes in public perception of water, economy, technology, or
other changes.
(iii) Sectoral Disaggregation
Forecasting by sectoral disaggregation is similar to the per capita method, but per capita
water use factors are disaggregated into smaller subsets of customers to increase the resolution of
the forecast. This requires a population forecast for each customer subset, as well as past usage
data for each subset. However, like the per capita forecasting method, sectoral disaggregation
may fail to account for changes in lifestyles, economy, technology, etc.
(iv) Econometric
Econometric models establish a statistical relationship between water demand and several
determinants (commonly including: price, income, number of household, weather, employment).
Historic data for all the determinants, as well as past water usage, is required. Independently
generated forecasts for any of the determinants (e.g. for weather, climate change forecasts) are
incorporated to reflect how the changes in determinants will alter demand.
(v)
Variable Flow
The variable flow method utilizes water use factors (consumption per household or
employee), like the sectoral aggregation method, but modifies the water use factors over time to
account for changes in price and conservation measures. Water use factors can be obtained from
past billing data; however the importance of weather-normalizing the data beforehand is noted.
After the water use factors are derived from weather-normalized data, the factors are adjusted for
price and income effects over time, integrated with conservation planning data, and multiplied by
demographic projections to create the water demand forecast.
(vi) End Use
End use models are a more detailed approach to water demand forecasting. End use
models identify the quantity of water used for specific activities (e.g. flushing the toilet), and by
applying this together with the population data and frequency of the water-consuming activity, a
water demand forecast is formed. End use models are helpful for evaluating the impact of
Review of Seattle Demand Model – Page 5
conservation programs, but are not widely used to forecast water demands because they are so
data-intensive.
Discussion
Various forms of econometric models are frequently applied, but they require substantial
data for calibration. The incorporation of price variables and conservation in econometric
models is a field of current research. It is often difficult to separate price effects and
conservation effects. Similarly, trying to quantify the impacts of conservation programs
throughout historic datasets is a complicated undertaking. Most econometric models do not
account for the social behavior of individual consumers and instead attempt to account for
changes in social attitudes related to water with an aggregated approach.
Forecasts using econometric methods are more accurate when the demand equation is
site-specific. This is an advantage because the method is flexible enough to be used in a variety
of locations. However, econometric models require more effort since a standard demand
equation does not exist. There are commercial models available (e.g. IWR-MAIN), but these
now recommend that clients estimate their water use equations. Although the detail of
econometric models may increase the accuracy of forecasts, the transparency of the models has
been criticized in the past. Furthermore, the complexity of econometric models may imply more
certainty than actually exists.
Variable flow models have become more popular recently as a result of some of the
complications of econometric modeling. The Outlook (Central Puget Sound Supply Forum
2001) utilized the variable flow factor approach for its forecast of water demands in SeattleEverett-Tacoma, which covered more than 158 water utilities. The cities of Tacoma and Everett
both independently employ the variable flow factor approach in their water demand forecasts.
Variable flow methods require less data but may obtain results comparable to those obtained by
the econometric method. Although variable flow models increase transparency, the loss of
detailed data is also a source of criticism. Seattle has recently switched from an econometric
model to a variable flow water demand forecast model. They are currently using @Risk (a
Microsoft Excel add-in) to create their variable flow model. The U.S. Forest Service has also
utilized the variable flow factor approach to projected future water demands at a national scale.
Trending, per-capita, sectoral disaggregation, and end use methods are not widely used
for regional long-range water demand forecasts. While end use models are typically avoided
because they require data that may be too detailed and variable for a regionally-scoped project,
the other methods are seen as markedly simplistic for modern day long-term demand forecasting.
Nevertheless, some of the more basic approaches are still useful for small utilities for short-range
forecasts.
Summary
When selecting a demand forecasting technique, it is important to consider the tradeoffs
between performance, transparency, and data requirements. Some Puget Sound utilities have
selected the variable flow approach believing the forecast results are of comparable accuracy and
that the transparency of the forecast is increased. It is important to recognize that many
approaches exist, and exploring several forecasting techniques for a common site can help
determine what method produces the best results with information available from the region.
Later in this document, we examine the demand forecasting methods used in the nearby cities of
Review of Seattle Demand Model – Page 6
Everett and Tacoma. In Appendix 3 the demand forecasting review is expanded to include the
cities of Portland, OR; Washington D.C.; and Vancouver, B.C.
The SPU Regional Demand Forecast Model
Seattle Public Utilities has created a new water demand model to forecast its future water
needs (Flory 2006). For many years, Seattle used an econometric model. A few years ago,
modifications to that model began for a variety of reasons including the perception that the
model was overly complicated, difficult to calibrate and was hard to explain to the public. After
evolving for some time, the econometric model transformed to what has been described as a
“Variable Flow Factor” model, which is a fixed flow model that addresses individual demand
groups, but which is modified by a series of other factors, including price, income, and
conservation. Here we examine Seattle’s historic demand forecasting models as well as Tacoma
and Everett’s current water demand forecasting model.
History of Demand Modeling In Seattle
Recent water demand modeling by the City of Seattle has primarily used econometric
methodologies. The model developed for the 1985 Comprehensive Plan relied on input from the
1980 Census of Population and water consumption disaggregated into customer classes for the
year 1980. To forecast water demand, the Puget Sound Council of Governments’ (PSCOG)
population forecasts were used. Separate models were developed for each of four classes; single
family residences, multi-family residences, industry and commercial, and government and
education sectors. Consumption within each group was described by regression equations that
typically consisted of the price of water and basic demographics. Information included income
of single family households above and below some specified income lines, number of
households in area, number of multiple residence households, incremental water price paid by
multi or industry sector, and employment type in sector for industry, commercial, government,
and education. The report generated a water demand and calculated price elasticity. The
forecast did not prove accurate—due to the unexpected effects of the 1987 and 1992 droughts
and conservation efforts. It was estimated that by the year 2000 consumption for the Seattle
system would be approaching 200 MGD, and that by 2005 the Seattle system demand would be
210-215 MGD. Actual demands for those years were 147 MGD and 130 MGD respectively.
Figure 2 shows Seattle’s forecasted demand for the 1980 demand study.
Review of Seattle Demand Model – Page 7
Figure 2: Seattle Demands – Past and Present (1980 Study)
Water demand updates were made every few years, with more in-depth studies being
completed with every new Comprehensive Plan. The 1994 Comprehensive Plan relied on the
Seattle Water Departments Water Supply Plan of 1993. The Seattle Water Department
continued to use an econometric model for their demand forecasting that was a jointly developed
effort between the Seattle Water Department and the East King County purveyors. The same
user-classes were used as in the 1985 model (single family residential, multi-family residential,
commercial/industrial, and governmental /institutional). The econometric model was developed
and calibrated using data from 1975-1987. Some key variables were water and sewer rates,
weather, income levels, number of single or multi-family homes, and employment levels. With
this model, price elasticity of demand played a larger role in the forecast. Different price
elasticities were estimated for each sector. For most sectors, the elasticities were disaggregated
to peak and non-peak usage, and, in some sectors, outdoor and indoor water usage. This
framework was consistent in its responses to the peak outdoor consumption to changes in price.
The data used in the model for the forecast came from PSRC and from Seattle City Light. PRSC
provided future population and employment from 1990-2020 and Seattle City Light provided
economic and demographic forecasts that were prepared by Conway and Associates. Different
scenarios were evaluated, with two scenarios selected for more in-depth analysis: (1) no new
wholesale customers, and (2) incremental demand from eight requesting utilities. Conservation
was also included within this model. A demand forecast not including the conservation or code
savings predicted that in 2000 demand would be ~205 MGD and in 2010 it would be 242 MGD.
With conservation included, in 2000 demand would be ~175 with constant real rates and ~185
with marginal cost rates. Actual consumption in 2000 was approximately 147 MGD.
Review of Seattle Demand Model – Page 8
The 1997 model utilized the same basic framework as the 1993 Water Supply Plan model.
It was recalibrated with 1994 consumption levels and included more recent data, including 1996,
to compute the variable inputs. The 1997 model also contained new conservation estimates. A
new model was developed in 1999, which was an extension of the 1993/1997 model. Seattle
adopted an aggressive conservation plan (1% per year) for the period 2000-2010. The 1999
model assumed that the 1% reduction is met every year. The 1999 model forecasted water use of
150 MGD for 2000, 146 for 2005, and 144 MGD for 2010. Once again, actual demands for the
2000 and 2005 years were 147 MGD and 130 MGD, respectively. Even with conservation
included, the model did not predict the large drop in consumption over a relatively short period
of time. Figure 3 depicts 1997 and 1999 water demand forecasts made by the City of Seattle for
the 2001 Water System Plan Update. Figure 4 contains Seattle’s latest demand compared with
its 2004 predicted demand. The 2004 demand is the last demand using the econometric model.
Figure 3: Seattle Total System Demands – Past and Present (2001 Study)
Review of Seattle Demand Model – Page 9
Preliminary Draft 2006 Official Water Demand Forecast
200
200
180
180
Firm Yield
Jan 2004 Official
Forecast**
160
Annual Average MGD
140
160
140
Actual
Demand*
New Demand Forecast**
120
120
100
100
80
60
40
80
The gray area represents the
added uncertainty involved in
extrapolating beyond 2030.
* Actual billed consumption plus average non-revenue
60
water (12 mgd)
** Both forecasts include the Environmental Block.
40
20
20
0
1995
0
2000
2005
2010
2015
2020
2025
2030
2035
2040
2045
2050
2055
2060
Figure 4: Seattle Total System Demands – Past and Present (Current Demand Study)
Review of Seattle Demand Model – Page 10
Model Processes and Assumptions
SPU’s forecasting process is implemented in seven basic processes in the Water Demand
Model within a series of Excel Spreadsheets. An overview of the spreadsheets used to facilitate
each of the calculations is also provided in this section. Figure 5 (Flory 2006), presents a basic
framework for the model. The seven process boxes have been numbered (by these authors) to
help explain the model. A complete diagram depicting the links between processes (Figure 7)
and a table associating each process with a model sheet (Table 2) can be found at the end of the
section. In general, the model concepts are very simple, with the model being implemented in a
series of Excel Spreadsheets.
Process 1: Weather Adjusted Base Year; Consumption by Sector and Service Area
In Process 1, a weather-adjusted base-year is calculated. The period of 1992–2003 is
used to determine the average winter demand (November – February). Non-winter demands are
determined by comparing monthly averages to the winter values, resulting in a series of monthly
factors. Weather data are not explicitly used to account for data variability. This process,
described in detail below, occurs outside of the demand model and utilizes data from SPU’s
WaterBIRD database. The resulting average annual weather adjusted demands generated in the
file, WtrBIRDccf92-04Corrected.xls, serve as key inputs to the Seattle demand model.
WaterBIRD External File (WtrBIRDccf92-04Corrected.xls)
A weather adjustment factor is calculated that represents the percent difference between
winter and non-winter months. Winter is defined in this model to be the 120 days within the
non-consecutive periods of January 1st to March 16th and from November 16th to December 31st.
Using demand data from 1992 through 2003, four seasons are defined: summer, spring, winter
(winter1 (1/1 – 3/16), winter2 (11/16-12-31)), and autumn. The period 1992-2003 is used for
averaging the weather factors. The weather adjusted demand in CCF (hundreds of cubic feet) is
calculated as:
Winter1 CCF  Winter2 CCF *
Days per Year
* 1  Weather Factor 
Days of Winter per year
The final weather adjusted values are the base-use demand for each sector (winter usage = base
demand) determined over a 12 year period of varying summer usage thus giving an average
‘weather adjusted’ value that would on average be expected for any year.
Review of Seattle Demand Model – Page 11
WATER DEMAND FORECAST MODEL STRUCTURE
1
Historical Retail Consumption
by Sector and Service Area
Weather-Adjusted Base Year
Consumption
by Sector and Service Area
Weather Data
Base Year Demographics by
Sector and Service Area
2
Base Year Flow Factors
by Sector and Service Area
Applied to SF
& MF Sectors
3
Projected Income Growth
Income Elasticity
Forecast of Flow Factors
by Sector and Service Area
Adjusted for Income Effect
Demographic Forecasts
to 2030 by Sector and Service Area
Extrapolated to 2060
4
Intermediate Forecast of
Retail Demand
by Sector and Service Area
5
Sector Forecasts Aggregated to
Large Service Areas (LSA):
Seattle, Non-CWA and CWA
Projected Growth
in Real Rates
Price Elasticity
Code Savings
Price/Code/Program
Overlap Function
6
Retail &
Non-CWA
Program Savings
Forecast of Net Retail Demand Aggregated Across Sectors
Seattle
Non-CWA
T&D
Non-Revenue Water
CWA
Distribution System
Non-Revenue Water
Other Sources of Supply
NUD Block net
of NUD Demand
Transmission System
Non-Revenue Water
Total Seattle
Demand
7
New Wholesale
Customers
Total Non-CWA
Demand
Forecast of Total System Demand
Figure 5: Seven Water Demand Forecast Model Processes
Review of Seattle Demand Model – Page 12
Total CWA
Demand
CWA Block
Process 2: Base Year Flow Factors by Sector and Service Area
In Process 2, base year flow factors are calculated by each sector (single family
residential, multifamily residential, manufacturing non-residential, and non-manufacturing nonresidential) and service area (Seattle-inside city limits, Seattle-outside city limits, and Individual
wholesale customers). Stated simply, this process takes the weather adjusted demands for each
sector and service area and divides them by the appropriate factors (number of households or
employees) to create the base year flow factors. Within the demand model, this process takes
place in the sheets “Seattle Inside,” “Seattle Outside,” and in each of the wholesale city and
water district sheets. See “Process 4” for a detailed description of these worksheets.
Process 3: Forecast Flow Factors by Sector and Service Area Adjusted for Income Effect
In Process 3, the flow factors are forecasted. This is accomplished by adjusting the base
flow factors by an estimated income elasticity and projected income growth. (Elasticity is
defined as the % change in demand divided by the % change in income). This term is used to
modify only the single family residential and multifamily residential sectors. This assumes that
an increase in income will result in an increase in water use. Within the demand model,
calculation of the forecast flow factors takes place in the sheets “Seattle Inside,” “Seattle
Outside,” and in each of the wholesale city and water district sheets. See “Process 4” for a
detailed description of these worksheets. The “income effect” factor is calculated in the
independent variables worksheet (IndpVar). This sheet effectively serves as the model’s control
panel where the values of key input variables can be easily modified. Figure 6 depicts the
generation of the forecast flow factors for both residential and non-residential demand.
Independent Variables (IndpVar)
The IndpVar sheet contains many variables that are used elsewhere in the spreadsheet.
The variables that define the “price elasticity” and “income elasticity” are found in this sheet.
The first section deals with the price and income effect on water demand. The value of
the income effect variable is 0.27, which is the percent increase in demand for a percent increase
in income. Past studies of Seattle’s demand suggested an income elasticity of 0.50 but the other
value was chosen from a literature review. The price elasticity is defined in Process 5, below.
The annual growth rate of income comes from the U.S. Bureau of Economic Analysis (Dick
Conway & Associates, 2006). The “Non-Residential” variable allows an increase in the price
effect for the non residential water uses.
Non-Revenue losses of water are also defined in this sheet. The Wholesale Distribution
Non-revenue (value from annual surveys of wholesale customers 1994-2004) determines the
additional water needed by the wholesale customers. The system non-revenue section shows the
percentage of total water consumed that is expected to be lost in the “acquisition and
transmission” of water to the Seattle customers.
Review of Seattle Demand Model – Page 13
Figure 6: Forecast Factor Generation Process for Residential and Non-Residential Demands
Review of Seattle Demand Model – Page 14
A central feature of the model is the incorporation of PSRC’s population forecasts.
PSRC provides “High,” “Medium,” and “Low” Growth Scenarios. This is represented in the
model as the percentage of the 2000 population that is added or subtracted from the initial
growth forecast. Cell I17 in this sheet selects between these three scenarios. The percent of the
forecasted growth is allocated between single family, multi-family, and non-residential accounts
with the use of “dampening” factors, which allow the independent adjustment of the growth
scenarios for single family and commercial growth. For example, a value of 50% for the “Single
Family Hi-Lo Dampening Factor” and a “Growth Scenario 2060” value of 144.6% would result
in single family home growth of 22.3% above the median prediction by 2060.
The Weyerhaeuser Project sheet allows the addition of water supply from the
Weyerhaeuser Project in Northshore and Woodinville. The incremental demand of new
purveyors sets the future increases of the Ames Lake, Bend and Sallal demands
The all High/All Low values section allows quick control of most of the important values
in the independent variables tab.
Process 4: Intermediate Forecast of Retail Demand by Sector and Service Area
In Process 4, an intermediate forecast of retail demand is generated by using the PSRC
demographic forecasts to calculate future demand to the year 2030 based on the variable demand
factors, and then to extrapolate those demands to 2060. This process takes place in the sheets
“Seattle Inside,” “Seattle Outside,” and in each of the wholesale city and water district sheets.
Individual City Demands (Seattle_Inside, Seattle_Outside, Bellevue thru WD125)
The “Seattle_Inside” and “Seattle_Outside” sheets forecast the weather and income
adjusted retail water demand inside and outside of Seattle’s city limits, respectively. Each of the
other specific city and water district sheets (Bellevue through WD125; 22 sheets total) forecast
weather and income adjusted wholesale water demand. This last group can be divided into three
blocks: CWA, which includes “Bellevue” to “Tukwila,” Non-CWA, consisting of “Bothell” to
“Woodinville,” and District Demands, which are the sheets that begin with “WD.” A summary
of the wholesale calculations is found in the sheets “WhlsSum,” “WhlsAs1,” and “Non-CWA,”
all of which are described below. However, only the individual city and water district sheets
(and not the summary sheets) produce results that are utilized elsewhere in the model.
Inputs to the city demand worksheets include annual weather adjusted demand by sector
(1992 – 2003), annual household and employee data (1990 – 2003), a growth factor for single
and multifamily households, and a single value to capture income effect. These data are
obtained from SPU’s WaterBIRD output (DSccf92-03.xls) and purveyor data
(PvRetailByMnth94-03.xls), PSRC’s demographic records (TAZ02.xls), and the demand model
sheet “IndpVar.” These sheets are further described below.
For each city, water demand is forecasted for three user groups: single family, multifamily, and non-residential. Household and employee population are forecasted by interpolating
between the adjusted 2003 values and PSRC’s 2030 forecasted values (with small adjustments
made for the relatively slow growth rate early in the period due to economic recession). The
results are then extrapolated to produce a population forecast through 2060. Beginning with the
2003 data, future per household consumption is forecast by adjusting for income effect, resulting
in consistently increasing residential consumption levels. Per employee consumption is assumed
to remain constant at 2003 levels. For the residential sector, water demands are forecasted by
Review of Seattle Demand Model – Page 15
multiplying the number of anticipated households by the forecasted weather and income adjusted
per household usage. For the non-residential sector, the number of anticipated employees is
multiplied by the 2003 weather and income adjusted per capita usage. The annual “weather and
income adjusted demand” needed by each group is then summed in a totals column, which is
referenced by aggregate sheets (see Process 6).
Wholesale Summary (WhlsSum) & Wholesale Combined (WhlsAs1)
Although both of these sheets aggregate the wholesale customers’ populations and
demands, they accomplish the task in different ways. The wholesale summary sheet adds the
values of each of the populations and demands for each wholesale customer sheet and uses the
summation as the basis for the sheet’s calculations. The WhlsAs1 takes the summation of the
wholesaler input data from TAZ02.xls and then performs the calculations. This serves as a
check that the wholesale sheets account for all of the data in the TAZ sheet.
Non-CWA (Non-CWA)
The Non-CWA sheet is a summation of the single family, multifamily, employment and
population for each of the following blocks of water use: CWA, non-CWA, non-CWA + Seattle
and total.
Northshore and Woodinville Other Sources of Supply (Northshore; Woodinville)
The “Northshore” and “Woodinville” sheets are similar to the other wholesale customer
sheets except that they have an additional section that determines the effect of price, code and
conservation for single family, multifamily and commercial uses. In addition, these sheets
calculate the net water used, considers non-revenue uses and then subtracts the other sources of
supply.
Other Sources of Supply (OSOS)
The “OSOS” worksheet forecasts the total anticipated CWA and Non-CWA demand
supplied by other (non-SPU) sources. Other sources include Redmond, Skyway Wells, Highline
Wells, Olympic View, Water District 90, Water District 125, Northshore Utility District, and
Woodinville. Direct communications with purveyors and the 2003 Survey of Wholesale
Customers suggest that the current 5 MGD demand supplied by other sources will increase to
approximately 8 MGD by 2030.
Process 5: Sector Forecast Aggregated to Larger Service Areas (LSA): Seattle, Non-CWA,
and CWA
In Process 5 the forecasts for the sectors are aggregated into service area jurisdictions
(Seattle Service area, Cascade Water Authority area, and a Non-Cascade Service area). This
aggregation is necessary to define specific service areas and to address the different fashions in
which conservation is addressed. Before entering Process 6, these forecasts are adjusted for two
issues, price elasticity and programmatic conservation actions. Price elasticity is defined here as
the % decrease in demand for a % increase in water price. Different water prices can be tested
for each jurisdiction and the impact of price on demand determined. The savings from program
changes and code changes are also calculated before Process 6. Price and conservation impact
factors are determined in the “Price” and “Cnsrvtn” sheets, respectively. These values are
Review of Seattle Demand Model – Page 16
applied to the forecast aggregate demands in the “Seattle Total” and “Wholesale Total” sheets,
which are described in Process 6.
Price (Price)
The price worksheet determines the anticipated impact of price on water demand. As
inputs, the sheet requires the price elasticities for each sector—single family (SF), multifamily
(MF), and non-residential (NR)—as well as the anticipated annual growth in real prices, which is
assumed constant at 1%. From these values, water demand as a fraction of 2005 demand is
calculated for each sector. Finally, the impact of price on water demand is determined as a
percentage decrease in 2005 demand for each sector.
Seattle’s water rates vary between sectors (retail and wholesale, residential and nonresidential), seasons, and consumption levels. As such, the demand model utilizes a system
average price of water. This is the ratio of total annual system revenue to annual billed demand,
excluding CWA portions. The average annual change in system price is presumed to be a
constant 1%, and is based on the assumption that revenue requirements will double by 2060
(after adjusting for inflation).
Assumptions for price elasticity values are founded in a literature review by SPU staff
and Seattle’s 1992 econometric model. Due to the complexity of the rate structure, overlapping
conservation impacts, and other factors, middle range values were chosen. The model’s base
price elasticities of water demand for single family, multifamily, and nonresidential groups are 0.20, -0.10, and -0.225, respectively.
Conservation (Cnsrvtn)
This worksheet determines the impact of code and programmatic savings that are applied
to Seattle’s retail and non-CWA customers. Code savings consist of anticipated reductions in
demand due to federal and state plumbing code changes that result in the incorporation of new
and more efficient technologies. The code savings as percent reductions in unadjusted demand
forecasts have been calculated externally by the Conservation Potential Assessment Model
(CPA) through 2030.
Table 1: Savings from Conservation Potential Assessment Model
2010
2020
2030
Single Family
-3.5%
-7.4%
-9.4%
Multifamily
-5.1%
-10.9%
-13.9%
Non-Residential
-3.7%
-7.1%
-8.7%
Source: “Inputs and Assumptions for the Water Demand Forecast Model”
Programmatic savings are reductions brought about by specific SPU campaigns. SPU’s
“Everyone Can Save” program and Initiative-631 savings are expected to result in 6.8 MGD of
savings from 2005 to 2010. This target reduction amount is distributed proportionately amongst
the retail and Non-CWA single family, multifamily, and non-residential sectors. In January 2006,
the Seattle Regional Water System Operating Board determined that 15 MGD of combined price
1
This 2001 initiative increased water rates for Seattle's largest residential and commercial consumers, and the extra
revenue was then used to fund water conservation retrofits for low-income residences. Additionally, as opposed to
selling the conserved water to developing suburbs, the program specified that it remain in rivers and streams as an
"environmental block" to support salmon populations
Review of Seattle Demand Model – Page 17
and programmatic savings should occur between 2011 and 2030. With this in mind, the change
in price impact for the retail and non-CWA groups from 2010 to 2030 is calculated. This value,
less the assumed 50% overlap between price and conservation impacts, is subtracted from the
target goal of 15 MGD to produce the required amount of programmatic savings. Again, the
target reduction amount is distributed proportionately amongst the retail and Non-CWA single
family, multifamily, and non-residential groups.
Process 6: Forecast of Net Retail Demand Aggregated Across Sectors
In Process 6, forecasts of net retail demand are made. This involves adjusting the
aggregated forecasts by such factors as the demand for non-revenue water, water lost in
distribution, and other sources of supply. This results in total water demands for Seattle, CWA
and the Non-CWA areas.
At this stage, the weather, income, price, and conservation adjusted intermediate
forecasts are aggregated over each service area: Seattle, CWA, and Non-CWA. Forecasts of
non-revenue demands (from “IndpVar”) are added to the demand forecasts to produce net
demand for each sector. This process takes place in the “Seattle Total” and “Wholesale Total”
sheets.
Seattle Total (Seattle Total)
In “Seattle Total,” Seattle’s retail demand forecasts are further refined. Inputs include
“weather and income adjusted demand” by sector, price impact, code impact, and program
impacts. This data comes from the “Seattle Inside,” “Seattle Outside,” “Price,” and
“Conservation” worksheets elsewhere in the model.
Net conservation impact is the sum of code and programmatic impacts, less the overlap
of these conservation effects with price impact. In the base scenario, price is assumed to overlap
with conservation measures by 50%. Justification for this percentage is not provided. The price
and net conservation impacts effectively reduce the “weather and income adjusted demand”, and
once removed produce Seattle’s net retail consumption forecasts for the single family,
multifamily, and non-residential user groups. These annual net billed consumption values for
each user-group are summed and added to Seattle’s portion of non-revenue water to produce a
final forecast of total retail water demand through 2060. With the covering of in-city reservoirs,
non-revenue water is assumed to decrease between 2000 and 2015 from 12 to 9 MGD. However,
after 2020, increased leaks due to infrastructure aging will result in an assumed 1 MGD increase
in non-revenue water every 10 years.
Wholesale Total (Wholesale Total)
In “Wholesale Total,” the wholesale demand forecasts are further refined. Inputs include
“weather and income adjusted demand” by sector, price impact, code impact, program impacts,
and other sources of supply. These data come from the various city and district wholesale
worksheets, and the “Price,” “Conservation,” and “OSOS” worksheets elsewhere in the model.
Net conservation impact is the sum of code and programmatic impacts, less the overlap
of these conservation effects with price impact. As a base, price is assumed to overlap with
conservation measures by 50%. The price and net conservation impacts effectively reduce the
“weather and income adjusted demand”, and once removed produce CWA and Non-CWA net
retail consumption forecasts for the single family, multifamily, and non-residential user groups.
These annual net billed consumption values for each user-group are summed and added to
Review of Seattle Demand Model – Page 18
transmission and distribution non-revenue demands, assumed to be a constant 1% and 6% of
retail water demand in the wholesale service area, respectively. Finally, other sources of supply
are removed to produce a forecast of total CWA and Non-CWA (including incremental demand
of potential new purveyors) wholesale water demand for single family, multifamily, and nonresidential groups through 2060.
Process 7: Forecast of Total System Demand
In Process 7, the total system demand is calculated from the individual demands from
Process 6, plus any anticipated new wholesale customers and the Cascade Water Alliance water
block. This process takes place in the “System Total” worksheet, while results can be seen in
both the “Scenario” and “Table” sheets.
System Total (System Total)
“System Total” contains the calculations and values used in “Scenarios.” “System Total”
gathers values from all other worksheets to generate scenario values. It also contains data
pertaining to the effects of conservation, that has been included. These data probably originate
from adjusting High/Low conditions and Conservation numbers within ‘”IndpVar.” ‘”Table’”
and “Scenarios” reference ‘”System Total” to display final forecasted demand output for the
model. “Scenarios” and “Table” could be consolidated into the ‘System Total’ sheet.
“System Total” uses aggregate total data from the “Northshore,” “Wholesale Total,”
“Seattle Total,” and “InpdVar” sheets to generate the four scenarios described in ‘Scenario’.
Table (Table)
The “Table” sheet is a aggregation of all the sheets in the forecast model. It provides
final forecast values for water demand. This sheet is the source of the figure in the demand
forecast documentation. The official forecast is computed assuming that non-CWA demand
increases, CWA chooses a block reduction schedule, new areas such as Sallal have an increased
demand, non-revenue water slowly increases overtime as leaks and other losses increase, and
that the environmental block demand remains at four MGD into the future.
The main inputs to “Table” are from aggregate demand sheets. These sheets provide the
total demand for the specific regions that correspond to the larger categories of CWA, non-CWA,
etc, and for the non-revenue and environmental demands.
Scenarios (Scenarios)
“Scenarios” is similar to “Table” in that it is a compilation of other sheets in the demand
forecasting model. It takes values from “System Total” to generate the scenarios. It is different
in that it divides the forecast into four scenarios. These scenarios are listed below:
a) Actual Water – This scenario is one in which Seattle, non-CWA, and CWA water demand
increases, and CWA does not enter into a block reduction agreement with the SPU.
b) CWA-Block – The CWA-Block scenario consists of Seattle and non-CWA demand
increasing overtime, but CWA entering a block reduction agreement with SPU. The final
forecast is much less in this scenario than in the previous scenario.
Review of Seattle Demand Model – Page 19
c) NUD-Block – This scenario is the CWA-Block scenario but with an increased NUD-Block
now, in which NUD increases purchased amount from 5.9 to 8.6 MGD now and continue to
purchase 8.6 MGD over the next 55 years.
d) Environmental Block – The environmental block is the same as the NUD-Block but includes
the environmental block, which drops to 4 MGD and stays constant at 4 MGD through 2060.
Within each scenario a high and low forecast was made about the baseline case. These high and
low scenarios were pasted as values into the sheet. It is assumed that these values were
generated from either @Risk or were pasted in from PRSC High and Low estimates controlled in
the “IndpVar” sheet.
Table 2: Seattle Demand Model Sheets Associated With Each Process
Process
Model Worksheets
WaterBird (external file)
Conservation
IndpVar
Price
System Total
Seattle Total/
Wholesale Total
Seattle Inside/Outside
Purveyors
(Cities and Water Districts)
1
X
2
3
4
5
6
7
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
Process requires input from these sheets or files
Process takes place in these sheets or files
Results of the process are output to these sheets (if not the sheet in which the process occurs)
Review of Seattle Demand Model – Page 20
Figure 7: Seattle Demand Forecast Model
Model Assumptions
The model uses PSRC 2004 Traffic Analysis Zones (TAZ) level forecasts for
demographic variables. A straight line extrapolation is used to forecast beyond 2030. Table 3
contains most numerical assumptions used in Seattle’s water demand forecasting model.
Non-Revenue Water
Seattle’s demand model assumes that non-revenue water will decrease from 12 MGD to 9 MGD
between 2000 and 2015 as in-city reservoirs are covered. After 2020, the model assumes a 1
MGD per decade increase in non-revenue water as infrastructure ages and more leaks occur
throughout the system. Wholesale customer distribution system non-revenue water is assumed
to be a constant 6% of retail water demand in the wholesale service area over the forecast period.
This is added to the forecast of wholesale customers’ retail demand.
Environmental Block Assumptions
An environmental block is added to the demand in accordance with the I-63 Ordinance.
The environmental block increases demand by 6 MGD from 2006-2009, to 9 MGD from 2010
through 2014, and to 12 MGD in 2015 and thereafter. In 2010 it is assumed that Seattle will
obtain 5-8 MGD of this block elsewhere and so from 2011 on only 4 MGD is actually added to
the total demand.
Review of Seattle Demand Model – Page 21
Table 3: Numerical Assumptions within the Seattle Demand Model
Average annual rates of growth for SPU’s
entire service area over the period 2010
through 2030
Population
0.80%
per year
Single Family
Households
0.40%
per year
Multifamily Households
1.90%
per year
Employment
1.10%
per year
Annual growth in real
mean household
income
Elasticity of residential
demand to changes in
real (inflation adjusted)
household income
Price Elasticities
Single Family
Households
Multifamily Households
Non-Residential
Average system price
increase
1.70%
per year
0.27
-0.20
-0.10
-0.23
1%
per year in real terms
Code savings are expressed as percent reductions in the
unadjusted forecasts of consumption by sector, i.e., prior to
adjusting demand for the impacts of income growth, price
increases, and programmatic conservation.
2010
2020
2030
Single
Family
Multifamily
NonResidential
-3.50%
-7.40%
-9.40%
-5.10%
-10.90%
-13.90%
-3.70%
-7.10%
-8.70%
Beyond 2030, cumulative code savings are assumed to
continue growing but at a rapidly decreasing rate
Review of Seattle Demand Model – Page 22
Conservation Savings
The Seattle Demand model assumes that the remainder of their programmatic targets are
met by 2010, this and other programs equate to a 6.3 MGD reduction from 2005-2010 in demand.
The model also assumes that from 2011-2030 another 15 MGD reduction will occur given future
price- induced and programmatic savings. These savings only apply to Seattle demand and
current non-CWA wholesale customers, and that after 2030 no more savings will apply. Because
of overlap between the price induced savings and the programmatic savings, it is assumed that
“the price effect overlaps with code and programmatic savings as long as the total amount of
overlap represents less than half of total code and programmatic conservation (as is the case over
the forecast period). However, if the price effect exceeds combined code and programmatic
conservation, the amount of overlap is capped at 50%.” (personal communications, Bruce Flory
2006)
Demand from Whole customers and CWA-Block Contract Assumptions
Future demand that wholesale customers (both CWA and non-CWA) expect to receive from
other suppliers is subtracted out of their demand from the Seattle system. Currently 5 MGD is
subtracted from wholesale demand, this is projected to increase to 8 MGD by 2030.
Seattle’s demand model incorporates their recent agreement with Cascade Water Alliance
(CWA). The model assumes that SPU will provide 30.3 MGD through 2023 to CWA, after 2023
the demand is expected to decrease by 5 MGD immediately in 2024, and by another 5 MGD in
2030. After 2030 a 5 MGD decrease in demand every 5 years is assumed by the model through
2045.
Demand for Northshore Utility District (NUD) consists of a fixed block of 8.6 MGD.
The demand model assumes that this amount is constant through 2060. NUD currently only uses
about 5 MGD of their block but expects to reach 8.6 MGD demand in 2060. New wholesale
customers such as North Bend, Ames Lake, and Sallal are modeled based on local projections.
In total it is estimated that their demand will be 1.4 MGD in 2040. It is assumed in the model
that Edmonds will continue to purchase water from Alderwood.
Weather Assumptions
Seattle’s demand model uses an average weather adjusted value of demand to account for
variability due to weather effects. An analysis of daily consumption data from 1982 to present
shows a maximum variability of +/- 5% in demand. Because the model uses average weather
adjusted demand data, it is assumed that +/- 5% of the demand value will yield future minimum
and maximum demand values for any given year. There is no assumption made about future
climate changes other than that future variability will remain within the +/- 5 % range given.
Review of Seattle Demand Model – Page 23
Year
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2015
2020
2025
2030
2060
Weather
Adjusted
Demand
CCF
MGD
11,958,026
11,348,881
11,413,932
11,478,975
11,541,495
11,650,480
11,691,240
11,437,463
11,567,843
11,059,027
10,818,663
10,571,274
10,338,223
10,128,461
24.4
23.3
23.4
23.5
23.6
23.9
24.0
23.4
23.6
22.7
22.2
21.7
21.1
20.8
PSRC
Single
Family
Households
131,418
131,826
132,236
132,647
133,059
133,472
133,886
134,302
134,719
135,138
135,557
135,611
135,664
135,718
135,970
136,222
136,474
136,726
136,978
137,230
137,481
138,741
140,001
141,261
142,520
150,078
Income
Effect
GPD per
SF Hhld
Seattle
Inside
Forecast
Seattle
Outside
Forecast
Single
Family
Price
Impact
Code
Formula
Program
Impact
Net
Consv
Impact
Net
Single
Family
Gallons
per Day
Gallons
per Day
MGD
MGD
MGD
MGD
MGD
MGD
MGD
MGD
0.70
0.70
0.71
0.71
0.71
0.73
0.75
0.76
0.78
0.90
184.8
175.3
175.8
176.2
176.2
177.8
177.8
173.4
174.4
167.1
163.4
159.6
155.4
152.4
153.1
153.8
154.5
155.2
155.9
159.5
163.2
167.0
170.9
196.0
23.4
23.5
23.6
23.9
24.0
23.4
23.6
22.7
22.2
21.7
21.1
20.8
20.9
21.0
21.2
21.3
21.4
22.1
22.8
23.6
24.4
29.4
3.0
3.0
3.1
3.0
3.1
3.0
3.0
3.0
2.9
2.7
2.7
2.6
2.6
2.7
2.7
2.7
2.7
2.8
2.9
3.0
3.1
3.6
26.4
26.5
26.7
26.9
27.0
26.4
26.7
25.7
25.0
24.4
23.8
23.4
23.5
23.7
23.8
24.0
24.1
24.9
25.7
26.6
27.4
33.1
0.0
-0.5
-0.9
-1.4
-1.8
-2.3
-3.4
-4.4
-5.3
-6.1
-5.4
26.4
26.5
26.7
26.9
27.0
26.4
26.7
25.7
25.0
24.4
23.8
23.4
23.0
22.7
22.3
22.0
21.6
21.0
20.5
20.2
20.0
24.2
0.0
0.0
-0.1
-0.1
-0.2
-0.2
-0.5
-0.8
-1.0
-1.3
-3.4
0.0
-0.2
-0.3
-0.5
-0.7
-0.8
-1.4
-1.8
-2.1
-2.3
-2.7
0.0
-0.3
-0.6
-0.9
-1.3
-1.6
-2.3
-3.0
-3.7
-4.5
-4.5
Table 4: Sample Calculation of Water Demand
A sample calculation of how the Seattle Model forecasts demand is shown above in Table 4 and
described below in six steps. The demand calculated pertains to the single family residential
sector, with the last column being the net single family demand for the City of Seattle (a
combination of inside and outside Seattle city limits).
Calculation Steps:
1. Start with weather adjusted demand data in CCF, convert to MGD. Use PSRC data
through 2003 for base case of growth in specific sector.
2. Interpolate between 2003 and 2060 to obtain household values for the in between years.
3. Calculate GPD per sector unit by extrapolating from base year (2004) to 2060 using the
Income Effect as the driving variable, i.e. future GPD per unit is forecasted by
multiplying the previous years GPD per unit by 1 plus the income effect. Income effect
is made up of elasticity multiplied by the annual growth rate.
4. For the city of Seattle, the values of Seattle Inside and Outside are combined for each
sector, though the calculations are identical for each region, including those outside of the
Seattle area.
5. The impacts of price, programmatic and code savings are totaled and used to create the
Net Conservation Effect (NCE) value. The Net Conservation effect is calculated as
follows: NCE = [Code Impacts + Program Impacts] – Maximum of [(Code Impacts +
Program Impacts) x 50% or 50% x Price Impacts]
6. The final sector demand is the summation of sector unit demand (i.e. Single Family Unit
demand) + Price Impacts + Net Conservation
Review of Seattle Demand Model – Page 24
Observations and Recommendations
The model uses the basic concept of “variable water demand.” The idea of “variable
water demand” for the residential sector implies that the model begins by calculating the per
capita use in a reference year (like 2003) and then is modified by other variables. The process
for forecasting municipal water demand is straightforward, uncomplicated, and simple to apply.
Aside from forecasts of future population, the model uses just four basic variables, per capita
water use in a reference year, price elasticity, income elasticity, and conservation.
Although model simplicity is often a virtue, the current version of this model relies
almost solely on two factors, population forecasts and conservation measures. There are several
other potentially important factors that are not currently modeled, including changing outside
irrigation patterns associated with changing landscaping norms, changing outside irrigation
patterns associated with decreasing lot sizes, changing water use associated with different
climatic patterns, technological changes in appliances, water code changes, and general changes
in the public’s conservation ethic. While it is currently unknown how important these factors are
in accurately estimating water demand, it appears likely that anticipated changes in these
variables might substantially affect demand estimates.
Income and Price Elasticities
In this model, the two other important explanatory variables for municipal water demand,
price elasticity and income elasticity, negated one another. Price elasticity decreases the future
demand for water as the price of water increases. Income elasticity increases the future demand
for water as regional income increases. Because of the values chosen for these variables, the
future net effect of these factors is a small increase in the demand. The values for single family,
multi-family, and non-residential price elasticities and income elasticity are not based on
calculations for data specific to Seattle and its suburbs, but were selected from published
literature values. They were selected based upon “professional judgment.”
Conservation
Water conservation is an extremely important variable and is treated in the model as a
target. The values for the conservation target have a major impact on the future demand for
water and these values are defined explicitly in the Seattle Water Demand model. It is assumed
that these values are derived from the Conservation Potential Assessment but a review of this
document is beyond the scope of this project. In the model, there is no explanation of the
functional relationship and feedback associated with the conservation component. Conservation
is “self-realizing”; if a target is given to the model, it will achieve that target regardless of what it
takes to reach that target. Because of this, Seattle’s water demand is forecasted to stay relatively
constant due to the impacts of conservation and the decreases in demand associated with the
Cascade Water Alliance contract.
Seattle’s demand remains below the 150 mgd level in the future due to an anticipated
reduction of 6.8 mgd due to programmatic conservation between 2005 and 2010, a 15 mgd
reduction due to price and programmatic reductions between 2011 and 2030, and the removal of
25 mgd of Cascade Water Alliance water demand due to contract agreements. It is important to
note that the demand being reported by Seattle is only for their anticipated demand and does not
reflect the increased demand encountered by other regional utilities.
Review of Seattle Demand Model – Page 25
The conservation component of the model is the most significant explanatory feature, yet
remains the least transparent. Because of the importance of this component of the model, and the
lack of documentation of its working within the model, more documentation is needed to explain
the assumptions included in the conservation component.
Industrial Water Use
The industrial water demand forecast is very simple. Like municipal water demand, the
demand forecasts for industrial water are very simple to apply. Aside from PSRC forecasts of
future manufacturing and non-manufacturing employees, the model uses just four basic
variables: average per employee water use in a reference year (2005), a factor describing nonmanufacturing per employee water use relative to average per employee water use, price
elasticity, and conservation. The average amount of water used per employee is determined by
dividing total industrial demand by the total number of employees in 2005 for a specified region.
Demand per non-manufacturing employee is then calculated as an SPU specified (based on
growth forecasts) portion of average demand per employee. In turn, this produces the demand
per manufacturing employee, as each group’s per employee demand multiplied by its number of
employees must result in the total industrial demand for 2005. It is assumed these 2005 per
employee demands remain constant through 2060, and annual industrial demands are simply the
product of group per employee demand and forecasted employees, adjusted for price and
conservation impacts.
Climate Change
It is becoming increasingly important for long-term demand forecasts to include seasonal
variation within the model to examine the impacts of climate change on water demand. Climate
change will result in increased summer temperatures that may increase the demand for outdoor
water use and may result in changes in patterns of water use. The current model does not have a
convenient way to incorporate these potential changes, or other seasonal type changes due to its
annual time-step. The model assumes that seasonal ratios for demand will stay constant into the
future or remain within the current system variability.
Validation
Seattle’s new water demand forecasts apparently have been released without the model
being validated. It is important that the model be validated before it is considered as a possible
regional model. This can be accomplished through demonstrating replication of historic
demands. Although the model is simple and requires few variables to be applied, its
transferability to other jurisdictions has not been demonstrated.
Review of Seattle Demand Model – Page 26
References and other Useful Documents
American Water Works Association. 2002. Water System Security: A Field Guide, Denver, CO:
American Water Works Association.
American Water Works Association. 2001. Emergency Planning for Water Utilities, M19,
Denver, Colorado: American Water Works Association.
American Water Works Association. 2001. Water Resources Planning: Manual of Water
Supply Practices. M50, ed. Denver, CO: American Water Works Association.
Arbués, F., M.A. Garcia-Valiñas, and R. Martinez-Espiñeria. 2003. Estimation of Residential
Water Demand: A State of the Art Review. Journal of Social-Economics, 32. 81-102.
Athanasiadis, I.N., A.K. Mentes, P.A. Mitkas and Y.A. Mylopoulos. 2005. A Hybrid AgentBased Model for Estimating ResidentialWater Demand. SIMULATION, 81 (3): 175-187.
Baumann, D.D., J.J. Boland., and W.M. Hanemann. 1998. Urban Water Demand Management
and Planning. New York, NY: McGraw-Hill.
Billings, R.B. and C.V. Jones. 1996. Forecasting Urban Water Demand. Denver, CO:
American Water Works Association.
Boland, J.J. 1997. Assessing Urban Water Use and The Role of Water Conservation Measures
Under Climate Uncertainty, Climatic Change, 37: 157-176.
Brookshire, D., H.S. Burness, J.M. Chermak and K. Krause. 2002. Western Urban Water
Demand. University of New Mexico, Department of Economics.
Brown, T. C. 2000. Projecting U.S. Freshwater Withdrawals. Journal of Water Resources
Research, 36 (3): 769–780.
Carver, P.H., and J.J. Boland. 1980. Short Run and Long Run Effects of Price on Municipal
Water Use. Water Resources Research. 16(4): 609-615.
Dandy, G., T. Nguyen and C. Davies. 1997. Estimating residential water demand in the presence
of free allowances. Land Economics, 73 (1): 125-139.
DeKay, C.F. 1985. The evolution of water demand forecasting, Journal AWWA, (October): 5461.
Foster, H.S., and B.R. Beattie. 1979. Urban Residential Demand for Water in the United States.
Land Economics, 55 (1): 43-58.
Froukh, M.L. 2001. Decision-Support System for Domestic Water Demand Forecasting and
Management. Water Resources Management, 15: 363–382.
Review of Seattle Demand Model – Page 27
Gottlieb, M. 1963. Urban domestic demand or water in the United States. Land Economics, 39
(2): 204–210.
Howe, C.W. and F.P. Linaweaver. 1967. The impact of price on residential water demand and its
relationship to system design and price structure. Water Resources Research, 3 (1): 13–32.
IWR-MAIN. http://www.iwrmain.com/
Kiefer, J.C. and G.A. Porter. 2000. Development of Probabilistic Water Demand Forecast for the
San Diego County Water Authority. Planning and Management Consultants, Ltd.
Kindler, J. and C.S. Russell, eds. 1984. Modeling Water Demands. Orlando, FL: Academic
Press.
Ludlum, M. 2004. Seattle Public Utilities: 2004 Water Demand Forecast Update, Phase 1,
Literature Review.
Nauges, C., A. Thomas. 2003. Long-run Study of Residential Water Consumption.
Environmental and Resource Economics, 26 (1): 25 - 43.
Palmer, R. N. and K. V. Lundberg. “Integrated Water Resource Planning.”
Seattle Public Utilities, 2000. Seattle Public Utilities: 2001 Water System Plan Update
Seattle Water Department, 1985. Volume II of the 1985 COMPLAN: Long-term Water Demand
Forecasts Rate Impact Analysis
Seatle Water Department, 1980. Seattle Comprehensive Regional Water Plan (COMPLAN)
Tacoma Water Demand Forecast Study (2004), Integrated Utilities Group, Inc.
The National Academy of Sciences. 2002. Regression Models of Water Use. Estimating Water
Use in the United States: A New Paradigm for the National Water-Use Information Program.
Available through: http://www.nap.edu/openbook/0309084830/html/100.html
U.S. Army Corps of Engineers. 1988. IWR-MAIN Water Use Forecasting System. V. 5.1 (June).
Carbondale, IL: Planning and Management Consultants, Ltd.
Western Water. 2004. Demand Forecasting Report: Water Demands and Wastewater Flows.
Review of Seattle Demand Model – Page 28
Appendix 1 – Metropolitan Water Demand Forecasting
Models
In our report, a review of demand forecasting techniques is provided for the cities of
Everett and Tacoma. Both of these models use a variable flow approach. However, to get a
broader perspective of modeling techniques used across the country, a thorough review of
forecasting methods for Tacoma, WA; Everett, WA; Portland, OR; Washington D.C.; and San
Francisco, CA. was performed.
The table below summarizes the type of demand model, the explanatory variables and the
strengths and weaknesses of each model.
Appendix 1 – Page 1
Utility
Seattle
Tacoma
Everett
Portland
Washington DC
San Francisco
Population
Served
Water Demand
Forecasting
Method
Agency that
Constructed the
Model
Explanatory Variables
Seattle Public
Utilities
Strengths
Weaknesses
Income and Price
elasticities, population/per
capita water use, and
conservation
Simple, not as data
intensive, and transparent
Simple industrial water
component, does not
address weather
directly, conservation
may drive final values
Reliance on a single
coefficient for the
model
1.35 M
Variable Flow
300,000
Variable Flow
(also called a unituse or
“requirements”
model)
Integrated Utilities
Group, Inc.
Price elasticities, inflation,
future water price, per
capita water use, population
Simpler and more
accurate (from historic
comparisons) than
previous econometric
model, considers climate
change
500,000
Variable Flow
(based on Central
Puget Sound Water
Supply Outlook)
The City of
Everett, Central
Puget Sound
Forum
Population, water use
factors (derived from per
capita water use)
Not as data-intensive as
other options, more
transparent
Uncertainties in
determining future
water use factors
787,000
Structural TimeSeries (Confluence
integration model)
Regional Water
Providers
Consortium,
Portland Bureau of
Water Works
Considers several factors
that drive trend
Mathematically
complex
4.1 M
Sectoral
Disaggregation
Interstate
Commission on the
Potomac River
Basin
Models daily demand and
long term demand.
Replicates past well,
climate change easily
incorporable
Mathematically
complex and requires
large datasets.
1.62 M
End Use (Demand
Side Management
Least-Cost
Planning Decision
Support System)
San Francisco
Public Utilities
Commission,
Maddaus (2003)
Bottom-up approach helps
to plan conservation and
provides more detail
Large data requirement,
high level of detail for
large scope of study
Population, seasonal
changes, weather
(precipitation and air
temperature), and long-term
cyclical trend
Weather variables
(precipitation, temperature,
period of dry days, etc), and
day of the week. Long term
forecasting done through
unit-use factors
Population, employment,
efficiency of water fixtures,
range of water fixtures used
Appendix 1 – Page 2
Tacoma Water Demand Model
Background
In 1999, R. W. Beck developed a long-term water demand forecast based on a multiple
regression analysis for the city of Tacoma. The statistical models generated forecasted demand
by customer class through 2050. However, the model was limited by historical data inputs
proved to be insufficient, and produced unreliable statistical relationships.
In 2003, Tacoma Water of Tacoma Public Utilities sponsored the development of a new
water demand forecast model. An initial effort by Integrated Utilities Group, Inc. (IUG) resulted
in a model that utilized an econometric approach, and relied on 48 multivariate regression
equations to relate daily demands to various controlling factors. However, actual historical
demand values did not support the model’s suggested results. As an alternative, the forecasting
process was simplified and two new models were produced to forecast water demands for the
Tacoma service area. The “Retail Demand Forecast” was developed by IUG. This model uses a
statistical approach to predict future water use for the residential, government, commercial, and
small industry sectors through the year 2040. A “Qualitative Demand Forecast” was produced
by Economic and Engineering Services, Inc. (EES) to forecast wholesale, large industrial, and
potential future customer demands.
Retail Demand Forecast
IUG refers to its “Retail Demand Forecast” as a “requirements model,” to suggest that the
model forecasts the requirements of water users. Also known as a unit-use model, this singlecoefficient model forecasts gross retail water demands as the product of a forecast of water use
per account and the forecasted number of future accounts. Within the Retail Demand Forecast
model, annual short-term forecasts were generated through 2010. Long-term forecasts were then
produced for 2020, 2030, and 2040. Water demand forecasts were determined for the single
family residential, multifamily residential, commercial and governmental, and parks and
irrigation sectors. The model includes a risk assessment component that uses Monte Carlo
Simulation to determine the degree of uncertainty in the long-term forecasts generated. Each of
the basic model processes are further described below.
1.)
2.)
3.)
4.)
5.)
6.)
Tacoma Retail Demand Model Processes
Forecast the number of accounts by rate code
Fit distributions to historical consumption (CCF per day)
Calibrate forecast model
Run Monte Carlo Simulations
Develop peak-day and peak-season forecasts
Develop price effect adjustment
The Retail Demand Forecast makes use of data from Tacoma Water, the Puget Sound
Regional Council (PSRC), and other sources. As such, two different scenarios were developed
and are differentiated by whether the forecasted increase in customer accounts is based on 2002
PSRC growth assumptions or on the historical growth rates calculated by Tacoma Water.
The 2002 PSRC data includes forecasts of population, number of households, and
employment levels within 938 Traffic Analysis Zones (TAZ) for 2010, 2020, and 2030. Traffic
Analysis Zones are geographical units defined for state or local transportation planning. Each
Appendix 1 – Page 3
zone varies in size but typically contain less than 3,000 people. While 68 of the TAZs are
completely inside of the Tacoma service area, 67 are split by the service area boundary. To
account for this, the demand forecasts were pro-rated such that the percentage of TAZ area
located within the Tacoma service area was used to determine the portion of TAZ’s population,
employees, and households actually served by Tacoma Water. Based on maps of Pierce County
and the Tacoma Water Service area, IUG determined that area was not an appropriate
representative of population for 23 TAZs. The proration for these TAZs was further adjusted to
reflect their varying patterns of developmental density. It was assumed that this, growth rate
forecast had a triangular distribution, which is somewhat similar to a standard normal
distribution. Because PSRC growth forecasts have historically been high for inside-city
customers in the Tacoma area, the water demand forecasts based on this data represent the
“high” scenario.
The second scenario is based on Tacoma Water’s historical growth rates by class as
reported in the “Summary of Water Revenue” documents. Using data from 1991 to 2002 and
adjusting for Tacoma Water’s acquisitions, distribution functions for the growth rate of each
class were determined. Finally, these growth rates were adjusted to incorporate the PSRC’s
long-term growth forecasts. For inside-city customers historical growth rates are used for 2003
through 2020, while the historical growth rates are adjusted by the proportionate reduction in
PSRC forecasts for the period 2020 through 2040. Growth rates used for outside-city customers
differ only in that 2010 through 2020 is a transition period in which rates are adjusted to
gradually achieve a 50/50 mixture of outside-and inside-city growth rates. The demand forecast
scenario based on these adjusted historical growth rates is considered to produce the “expected”
water demand values.
To determine daily water use per account, IUG queried Tacoma Water’s customer billing
database, which contained 10 years worth of historical information, by rate code. The data was
used to estimate the statistical distributions of daily water use per account for each class. The
resulting daily per account demands, were then multiplied by the forecasted number of accounts
previously calculated to produce forecasted demand values for each sector.
The Retail Demand Model was calibrated by “backcasting” demand for 1997 through
2002. The actual demand values and model results were not quite identical, and the difference
has been attributed to inconsistencies between Tacoma Water’s database and Tacoma Water’s
official “Summary of Water Revenues” reports. An adjustment factor was introduced to account
for the discrepancy, and was applied to the initial IUG forecasts.
A Monte Carlo analysis was used to capture uncertainty in the Retail Demand Model.
For each independent variable in the forecast model, statistical probability distribution functions
were assigned, as discussed in the relevant sections above. A computer simulation then
randomly sampled from these probability distributions and calculated water demand. By
repeating this process thousands of times, a range of forecast values and their associated
probability of occurrence was produced.
After forecasting the unadjusted, average-day sector demands, peak-day and peak-season
factors were calculated using customer billing data, and then used to generate peak-day and
peak-season forecasts. Peak day forecasts (PDFs) were calculated as the demand during the
peak-day of the year divided by the average annual daily demand. The PDF used in the model is
a 5-year averages based on Tacoma Water’s daily production reports for 1998 through 2002, and
has a value of 1.56 for all sectors. Peak-season was defined as June1 through September 30, and
peak season factors (PSFs) were determined by comparing daily demands (in CCF per account)
Appendix 1 – Page 4
based on average annual consumption with those based on consumption during the peak season.
The PSFs were calculated using historical data from 1993 to 2002 for a sample of accounts with
billing cycles falling neatly inline with the peak-season period. Calculated PSF values are shown
below.
Peak Season Factors for Tacoma Retail Demand Model
Sector
Inside City
Outside City
Single-Family Residential
1.37
1.44
Multifamily Residential
1.14
1.30
Commercial/Governmental
1.18
1.22
Parks/Irrigation
2.05
1.66
Finally, demand values were adjusted to account for the impact of water price by
considering price elasticity, inflation, and future water price. Using a 1997 study by the
American Water Works Association Research Foundation (AWWARF) called “Long-Term
Effects of Conservation Rates,” IUG selected a range of reasonable price elasticities for each
class. For the high-demand scenario based on PSRC data, price elasticities at the lower end of
the range were selected. Average price elasticities were chosen for the expected scenario. The
values for both scenarios are specified below.
Price Elasticities for Tacoma Retail Demand Model
Customer Class
PSRC
Adjusted
Scenario
Historical
Scenario
Residential
-0.200
-0.250
Multifamily
-0.050
-0.125
Commercial & Governmental
-0.100
-0.150
Parks & Irrigation
-0.350
-0.450
Given an assumed annual inflation increase of 3%, the assumed changes in nominal price of
water are as follows:
Changes in Nominal Price of Water
Annual Increase
During Years
8.7%
2004 – 2006
7.1%
2007 – 2008
6.2%
2009 – 2010
3.0%
All Years Thereafter
The assumed price elasticities were then applied to these expected price changes to
determine the impact of price on water demands, and the forecasts were adjusted accordingly.
Qualitative Demand Forecast
EES was responsible for forecasting water demands for Tacoma’s wholesale and large
industrial customers. Wholesale customers consist of current wholesale customers, customers
with contracts for future wholesale purchases, and all potential future customers. The
information for this model came from the “Piece County Coordinated Water System Plan”
(1999) and the “Outlook Report” (2000), as well as information on existing contracts and
demands. In recognition of the uncertainty involved in forecasting future wholesale demands,
Appendix 1 – Page 5
both a high and low forecasts were produced and correspond to the high and low retail demand
forecasts developed by IUG.
Forecast Adjustments
Since January 1, 2000, Tacoma Water has been committed to reducing water demands by
10% over a 10-year period. This reduction is required by a memorandum of agreement (MOA)
with the Washington State Departments of Ecology and Health as part of the Secondary Supply
Project. After controlling for savings from plumbing code changes, water rebate programs, and
wholesale customers (as estimated by EES), IUG determined the amount of additional,
unspecified conservation savings necessary for retail and wholesale customers in order to meet
this goal.
Tacoma Water also included an adjustment for climate change based on a study prepared
the City of Portland, Oregon. Additional storage capacity in Hanson Dam is expected to mitigate
the majority of negative supply impacts that would result given a potential decline in snowpack .
However, anticipated warming trends of 1.5°C for the decade 2020, and 2.0°C for the decade
2040 are expected to increase summer water demands. The projected impact of climate change
on the average annual demand is 4% by 2040. The projected impact of climate change on peakseason demand is 8% by 2040. The increased demands were incorporated into the “high”
(PSRC) scenario forecast.
Finally, forecasts were adjusted to account for “unaccounted-for water,” or the difference
between actual water production and metered water sales. This includes water used for system
maintenance, street cleaning, fire fighting, and other unmetered use, as well as that which is lost
through leakage. Historical data provided by Tacoma Water suggested that, on average, 10.6%
of water was attributed to unaccounted-for use between 1998 and 2002. According to IUG, these
numbers are comparable to industry standards.
Appendix 1 – Page 6
Everett’s Water Demand Forecasting Model
Background
The City of Everett utilizes the Regional Water Supply Outlook, developed by the Central Puget
Sound Water Supplier Forum, to forecast demands until 2020. Additional projections extend
until 2050. A high and low forecast is produced in addition to the baseline forecast. The model
used by the City of Everett is a variable flow factor model, which can be broken down into the
following steps (based on demand type):
Baseline (“Medium”) Demand Forecast
1) Calculate water use factors (current and adjusted)
2) Calculate peaking factors
3) Separate large industrial accounts
4) Obtain demographic data and baseline forecasts from PSRC
5) Multiply baseline population forecast by water use factor
High Forecast
1) Use the high population forecast from PSRC
2) Multiply high population forecast by previously calculated water use factors
Low Forecast
1) Use the low population forecast from PSRC
2) Adjust water use factors to reflect enhanced conservation and wastewater reuse
3) Multiply low population forecast by lower water use factors
2020-2050 Forecast
1) Extrapolate demand with expected growth in population
2) Large industrial demands are an exception—assumed to remain constant
Baseline Forecast
To create the baseline forecast, the first step was collecting information about singlefamily, multi-family, non-residential, and non-revenue water demands for a three year period
(1996-1998) from utilities that use Everett water. Corresponding demographic data from the
Puget Sound Regional Council determined the number of single-family and multi-family
households, in addition to the number of employees in each utilities service area for that period.
Next, water use factors are calculated for each water use category (single-family
residential, non-residential, etc.) for each utility. Water use factors are the ratios of household
demand to the number of households for the two different types, or—in the case of nonresidential demand—the ratio of non-residential demand to the number of employees in the
service area. Water use factors are calculated separately for each utility. Water demands and
corresponding employment information for large industrial accounts (which require in excess of
100,000 GPD) are not included in the non-residential portion because the demand model
considers them separately (discussed below). Water use factors represent the average demand
per household or employee from 1996-1998, and project the demand used in the model for 2000.
These factors take into account conservation practices and plumbing code savings active in the
Appendix 1 – Page 7
late 90’s. If a utility had insufficient data to calculate water use factors, factors from a similar
utility were used.
When determining future water demands, the City of Everett expects water use factors to
decline in the future due to increasing implementation of the 1993 State plumbing code changes.
Between 2000 and 2020, water use factors are reduced by 10% for single-family residential, 12%
for multi-family residential and 14% for non-residential sectors. These reductions are based on
analysis done by Seattle Public Utilities.
Peaking factors were calculated by taking the ratio of maximum daily demand to average
daily demand, using survey data from 1996-1998. Maximum daily demands for untreated water
can be estimated accurately using production data. Maximum daily demand for treated water,
however, cannot be directly measured. Treated water consumption is measured on a bimonthly
basis, but a rough estimation can be made by using the maximum daily production (i.e.
maximum treated at the filter plant and withdrawn from distribution storage during a day). This
is acknowledged as a minimum value for maximum daily demand, because many wholesale
customers also have their own storage to help meet maximum demands, and the withdrawal from
private storage is not monitored by the City of Everett. If utilities had insufficient data to
calculate peaking factors, data from similar utilities were used to supplement any available data.
The existing large industrial water demands are assumed to be held constant over the
entirety of the forecast. This assumption is based upon predictions given by large industries
themselves, who expect that increases in production will be offset by advances in efficiency.
Any significant changes in demand are predicted to correspond with future use of reclaimed
water. However, the City of Everett predicts re-development of a former industrial site, and
increases in demand corresponding with this predicted redevelopment are incorporated into the
demand forecast. The demand for the redeveloped industrial site is assumed to require untreated
water.
The Puget Sound Regional Council provided data and forecasts for single-family and
multi-family households, and employment for the years: 1990, 1998, 2000, 2010, and 2020.
Demands are predicted by taking the water use factors and multiplying them by the demographic
forecast provided by the PSRC. For each utility, the five components used to represent total
demand include: single-family residential, multi-family residential, non-residential (excluding
large industries, as discussed above), large industry, and non-revenue demands. Unmetered
accounts (in the single-family residential section) are assumed to use 20% more water than
corresponding metered accounts in 2000, and this discrepancy is reduced gradually until it is
eliminated in 2010 (i.e. the same demand factor will be used for currently metered and
unmetered accounts in 2010). The average percentage of non-revenue water (1996-98) is held
constant throughout the forecast.
Multiplying the projected water use factors by the projected population yields the demand
forecast for each sector of each utility. Adding together the five demand components for each
utility creates a utility-specific baseline total demand forecast for 2000, 2010, and 2020.
High and Low Forecasts
High and low forecasts were also produced. These are formed by examining population
growth and conservation changes over time. High and low projections for population growth are
provided by PSRC, but only for the county as a whole. The county-wide growth forecasts for
PSRC were used to create high and low growth rate projections for the period 2000-2010 and
2010-2020. The high growth rate was approximately 50% higher than the baseline, and the low
Appendix 1 – Page 8
growth rate was approximately 50% lower than the baseline. The low forecast also includes
effects of enhanced conservation program measures and reclaimed wastewater use.
2020-2050 Forecasts
To forecast beyond 2020 (to 2050), the model assumes that growth in demand will
increase proportionally with increase in population. Large industrial demands are an exception,
and assumed to remain constant due to predictions that higher efficiency will offset any increases
in production.
Appendix 1 – Page 9
Portland, OR Water Demand Model
The City of Portland uses an economic model, with time-series features for both shortterm and long-term water demand forecasts. The economic model uses a regression analysis to
define explanatory coefficients that include : population, seasonal changes, weather
(precipitation and air temperature), and long-term cyclical trend variables in addition to indicator
(dummy) variables. Historic production data from the Headworks facility are used as input to
the model. To reflect the effects of seasonal and daily variations, the Portland water demand
forecasting model is run using a daily timestep.
The City of Portland’s model incorporates a relatively small number of explanatory
variables, but some variables are used numerous times at various temporal lags.
Seasonal Variables
Seasonal variables are used to represent the annual pattern of water use, where summer
months have a higher demand than winter months. To create this pattern, a seasonal variable
uses trigonometric functions . The equation for the seasonal variables is defined as follows:
 2 it 
 2 it 
SSit  sin 
 and SCit  cos 

 DIY 
 DIY 
where:
i = the number of cycles per year,
t = the day of the year,
DIY = the number of days in the year (i.e., 365 days or 366 days for leap years).
SS and SC = the seasonal sine and cosine variables
Weather Variables
Weather variables in the demand model reflect the daily variations in water demand.
Because the model incorporates seasonal variables, the weather variables must be adjusted so
that they do not reflect seasonal changes. The weather data (i.e. precipitation and air
temperature) are regressed on a Fourier series with six sine and six cosine components. The
natural logarithm of actual daily precipitation less the seasonal precipitation created through the
seasonal regression analysis is the seasonally adjusted precipitation variable. The resulting
equation is shown below:
6
6


Pdl 0  P0   ˆ   ˆi SSi   ˆ j SC j 
i 1
j 1


where:
i and j = the number of lag days,
α, β and γ = regression coefficients,
SS and SC = the seasonal sine and cosine variables (defined above),
P0 = the natural log of the scaled precipitation data, defined as: ln(DP + 1) with DP is the
daily precipitation (in inches).
Appendix 1 – Page 10
Likewise, the Fourier series combined with precipitation and lag precipitation variables
are regressed to represent temperature dependence on season and precipitation. The natural
logarithm of actual temperature data less the regression of temperature yields the temperature
variables used in the demand model. Performing these modifications minimizes the correlation
between seasonal patterns, precipitation, and air temperature. The resulting equation is shown
below:
6
6


Tdl 0  T0   ˆ   ˆi SSi   ˆ j SC j  ˆ P0  ˆ P1 
i 1
j 1


where:
i and j = the number of lag days,
α, β, γ, δ and λ = regression coefficients,
SS and SC = the seasonal sine and cosine variables (defined above),
T0 = natural log of maximum daily temperature,
P0 and P1 = the natural logarithms of scaled contemporaneous and one day lag daily
precipitation.
Tdl0 is used in the regression model to account for the temperature that is separate from the
seasonal cycle
Demographic Variables
The Portland model uses a population variable to represent economic activity in the
demand forecast. A population variable does not, however, capture the downward trend in water
demand since the mid to late 1980’s. Water demand rose with population during the 1960’s
through the mid-80’s. Yet, water demand fell after that period, despite a growing population.
This complexity is represented in the Portland water demand model by a set of low-frequency
sine and cosine wave variables. These cyclical variables are determined similarly to the seasonal
variables, and follow the form:
 2 it 
 2 it 
LCTSit  sin 
 and LCTCit  cos 

 DIC 
 DIC 
where:
i = the number of cycles per year,
t = the day of the year,
DIC = the number of days in the cycle, 1960-2002 (15,706 days).
Indicator Variables
Indicator variables are incorporated into Portland’s model to replicate changes in demand
that are not produced by the seasonal, weather, and demographic variables. In Portland’s
demand model, indicator variables are used to account for changes in demand due to weekends,
building code changes (formed in 1992), mandatory curtailments, and economic downturn in
2001-2002.
The weekend indicator variable has a value of one during the weekend (Saturday and
Sunday), and a value of zero during the rest of the week. The coefficient for this variable
Appendix 1 – Page 11
indicates the average change in demand, both in magnitude (percent change) and direction
(through sign).
During the summer months of 1992, Portland residents faced mandatory curtailments due
to drought conditions in the region. A variable is created for each month of curtailments,
covering July, August and September of 1992. These three variables in the water demand model
are used to represent the change in demand due to the mandatory curtailments.
Also in 1992, Congress passed a law which changed the building codes, requiring new
and newly remodeled housing units to install fixtures with increased water efficiency. In an
attempt to account for changes caused by the building code changes of 1992, the Portland model
incorporates an explanatory variable which has a value of zero from 1960 to 1991, and a value of
one thereafter. Its coefficient represents the percent change in demand.
Furthermore, two indicator variables are used in the model to account for the effect of the
economic recession facing the region during 2001 and 2002.
Resulting Model
The resulting regression model, with 43 variables, produces an adjusted R2 with a value
of 0.888. The population variable has an “elasticity” of 0.972, implying that a 1% rise in
population yields slightly less than a 1% increase in water demand. This appears to be accurate,
since conservation efforts are captured within the indicator variables. The regression model
indicates that, in general, temperature has a greater proportional affect on water demand than
precipitation.
Creating Forecasts
Portland uses the regression model to generate both short-term and long-term forecasts.
Applying variable forecasts to the regression model (e.g. population forecast) creates the water
demand forecast. For the long-term water demand forecast, Portland’s Regional Planning
Agency (Metro) provides population forecasts. Although the population forecasts are related to
land-use and socioeconomic assumptions, more assumptions regarding conservation and longterm cyclical trends are incorporated into the regression model. Portland does not include
weather variables in their long-term water demand forecast because the city does not have
reliable long-term weather forecasts. The resulting water demand forecast—which incorporates
demand model coefficients, the predetermined variables, and the population forecasts—is
weather-normalized. Weather effects from 1960 to 2002 can be applied to the weathernormalized demand forecast. By looking at the resulting demand forecast, Portland can evaluate
the best, worst, and middle-of-the-road scenarios for peak season and peak event demand.
Forecasting Skill
To evaluate the forecasting skill of their model, Portland calculated the Mean Absolute
Percent Error. Portland found that the demand model performed markedly better after 1980.
The error calculated for the period covering 1960 to 2002 was 7.6%, whereas the error for 1990
to 2002 was down to 5.6%. Also, when examining model results for monthly or annual average
water demand, the modeling error further decreases.
Appendix 1 – Page 12
Source: (Portland) Regional Water Supply Plan Update, Chart 2-1, 2004.
Appendix 1 – Page 13
Washington Metropolitan Area Water Demand Model
Background
Two demand models have been created to forecast demand for the Washington
Metropolitan Area (WMA). A daily demand model is developed to forecast daily demand so
that reservoir efficiency can be maximized. Efficiency of reservoir operation is of great
importance in the WMA system because releases from upstream reservoirs require
approximately five to seven days of travel time and incorrectly timed releases can cause loss of
revenue or an insufficient supply of water to the WMA system. A long-term demand forecasting
model forecasts future water demand for a number of scenarios, including climate change
impacted scenarios.
Daily Demand Model
The WMA daily demand is primarily dictated by weather type variables and day of the
week. To model daily demand, long-term demand is “de-trended”. A dataset is created that
represents the daily demand factors and a regression can be performed on the dataset. The longterm demand factors are determined in later steps.
The dataset is de-trended using a liner regression for each WMA water supplier; the
residuals from the linear regression are what are used in the multiple regression models.
Twenty-two variables are used to determine daily demand. The types of variables in the
regression model are mostly weather related. These include variables like maximum daily
temperature, daily precipitation, lagged daily precipitation and temperature variables, and days
without precipitation. Other variables include day of the week and the Palmer Drought Severity
index. A multiple regression was performed for each water supplier using a backward stepwise
linear method for determining which variables explained the most variance. Each supplier had
similar variables that best explained the data, though none were exactly the same. After
constructing the multiple regression models, an auto-regressive integrated moving average
(ARIMA) model is constructed to account for the autocorrelation in the regression model error
term. The model overall has good explaining power, though performs better in the summer
months than in the fall months due to more response to weather type variables during the
summer than during the fall.
Long-term Demand Model
For long-term water demand forecasting three groups of water use categories are created:
single family home use, multi-family home use, and employee water use. Dwelling unit ratios
are developed for each region within the service area. The ratios are equal to the number of
single family households divided by the number of multi-family households.
Historic and forecast population, number of households, employment, and other needed
values are obtained from the Census 2000, Metropolitan Washington Council of Governments
(MWCOG) Round 6.1 and 6.4a Cooperative Forecasts. Unit use values for each sector (single
family, multi-family, and employee) are determined through use of the dwelling unit ratios,
MWCOG housing and employment data, and water consumption billed by regional utilities. The
unit use values are the main input into the long-term demand model.
Appendix 1 – Page 14
Conservation and unmetered water-use values were accounted for, both were
conservative estimates of current values. Long-term demand is calculated by assuming little
change in gallon per day household consumption, and applying the unit use values to the
MWCOG forecasts. MWCOG forecasts population growth, number of households, employment,
and other region specific data through 2025. The water demand forecast extrapolates the end
value through 2045. Low, intermediate and high forecasts are generated to simulate differing
MWCOG scenarios. A high forecast during a drought year is also forecast.
Overall the model is robust, accounting for long-term demand factors in the long-term
model and accounting for short-term driving factors in the daily demand portion of the model.
This allows versatility as well, being able to examine drought years in the future more accurately
because the weather variables have more effect in the daily model. Below are figures depicting
the current forecast compared to previous forecasts for water demand, and a model validation run.
The model validation run shows how well the model incorporates the day to day weather
variability within the forecast.
Appendix 1 – Page 15
Appendix 1 – Page 16
San Francisco, CA Water Demand Model
The San Francisco Public Utilities Commission (SFPUC) and its 28 wholesale customers
created an “end-use” demand forecasting model, the Demand Side Management Least-Cost
Support System (DSS). Demand is forecasted in two steps: (1) establishing base-year conditions
and (2) forecasting future water demand. Both steps are described in more detail the sections
below.
Water demands are forecasted on a wholesale customer scale, and represent the total
demands for each wholesale customer. Many wholesale customers supplement water received
from SFPUC with other water supply sources, however the forecast for water demanded from
SFPUC alone is reported separately. The DSS model reports the total demand for each
wholesale customer.
Establishing Base-Year Conditions
Although analysis began in 2003, 2001 was selected as the base-year for the DSS model.
This is because 2001 had relatively “normal” climatic conditions; did not show effects of
economic recession as much as 2002; and was a complete record, unlike 2003. To establish
base-year conditions, SFPUC gathered existing data from outside sources. Baseline demand
models required the following sets of data: (a) customer-billing data, (b) demographic data, (c)
water use data by customer billing category, (d) fixture replacement rate data, (e) indoor/outdoor
use data, (f) users per account, (g) water usage by end-use.
(a) Customer-billing records from 2001 were provided by the wholesale customers and
compiled by account category to represent the amount of consumed water. Account
categories included: single-family residential, multi-family residential, commercial,
industrial, institutional, and irrigational.
(b) Demographic data includes estimates of housing, employment, and population.
Sources of information include the Association of Bay Area Governments (ABAG)
and the U.S. Census Bureau.
(c) Standard values for water use by customer billing category comes from nationally
published information from the American Water Works Association Research
Foundation (AWWARF), which acts as a reference for normal ranges for water use
by end use for each billing category.
(d) Fixture use and replacement rates for the areas served by SFPUC are estimated by the
U.S. Census Bureau, AWWARF, Alameda County Water District, California Urban
Water Conservation Council (CUWCC), and East Bay Municipal Utility District.
(e) Indoor/outdoor water use separated by looking at plots of water use for a series of
years. The low points in the water use trend correspond to winter months, when
outdoor water use is minimal. Therefore, these low values for each year define the
indoor water use, and anything in excess is assumed to be for outdoor use.
(f) Average users per account are determined by examining the number of accounts
relative to current population and employment data. Sources of data include ABAG,
California Department of Finance, and the U.S. Census Bureau.
(g) Recent studies, such as AWWARF’s Residential End Uses of Water (Mayer et al.
1999), provided data regarding water consumption by end use.
Appendix 1 – Page 17
After collecting the necessary data a fixture model was created that forecasts water usage
based on fixture conditions for a specific year. The fixture model accounts for improvements
and replacements of water-using fixtures made during a year for existing accounts, as well as
fixture conditions for new water accounts. Forecasts from the fixture model are used in the
water demand forecast, discussed in more detail in the next section on forecasting future water
demand.
The final component of the base-year forecast is the calibration process. Calibrating the
model determines whether the end use and fixture models, and verifies that the service area
population and per-capita/employee water use are close to measured data. Every customer
billing category is calibrated (where applicable) before future water demand forecasts are
produced. Calibrating water use disaggregates production into its end uses, and then
reaggregates the end uses with the average number of persons to determine total water
production.
Forecasting Future Water Demand
After the base-year conditions are calibrated, the process of forecasting is possible. To
forecast future water demand, forecasts of population and account growth are necessary. The
account growth rate projections assume that the average number of individuals per account
(residential or non-residential, respectively) does not change; therefore, account growth is
directly related to the predicted growth rate for population and employment within each
customer-billing category.
A population projection is selected for each SFPUC wholesale costumer The population
growth rates are applied to the base-year (2001) population data to forecast population until 2030.
Similarly, employment growth rates for each customer’s service area are applied to the base-year
employment data to create employment forecasts for use in the non-residential water demand
forecast. There is only one available source for employment forecasts, the Association of Bay
Area Governments (ABAG).
After the population and growth forecasts are determined, they are applied to the DSS
model and base-year (2001) information to forecast water demand in five-year increments until
2030. The DSS model creates water demand forecasts with and without considering plumbing
code changes. For forecasts that incorporate plumbing code changes, assumptions for fixture
replacement rates were made, and are shown in the charts below. These replacement rates are
used in the fixture model, which is incorporated in the DSS model. SFPUC used fixture
replacement rates estimated by the California Urban Water Conservation Council (CUWCC),
which considers the age of housing, income levels, and fixture saturation study results.
Appendix 1 – Page 18
Additionally, the fixture model incorporates market availability assumptions to determine
what the level of efficiency will be for newly replaced clothes washers. SFPUC used estimates
from the Consortium for Energy Efficiency, shown below.
The final water demand forecast incorporates population and employment forecasts
alongside end-use forecasts. By multiplying the forecasted number of accounts (derived from
the population or employment forecasts) by the end-use forecast per account (from the fixture
model), a demand forecast for each wholesale customer by account category is created.
Summing these values gives the total potable water demand for each year. Recycled water
demand is not projected in the model, but wholesale customers provided recycled water demand
projections where applicable. These demands were added to the potable demand forecast values
to create a total water demand forecast. The result of the water demand forecast is shown in the
graph below.
Source:
SFPUC Wholesale Customer Demand Projections Technical Report (URS 2004)
Appendix 1 – Page 19
Appendix 2 – Overview of UrbanSim
Model Description
This appendix serves two purposes. It introduces an urban development model that is being
pursued by the Puget Sound Regional Council for generating population projections to be used in
regional planning and it also describes a prototype model that makes use of the UrbanSim model
to forecast water demands.
The following description is taken from the UrbanSim website:
(http://www.urbansim.org)
The model implements a perspective on urban development that represents a dynamic process
resulting from the interaction of many actors making decisions within the urban markets for
land, housing, non-residential space and transportation. For example:




Households make choices about whether to move, and if they move, where to locate.
Businesses make similar decisions.
Developers make choices of what properties to develop or redevelop and into what use,
at what density and scale.
Governments make infrastructure investments, and place constraints on development in
the form of land use plans, density constraints, environmentally-sensitive land
restrictions, urban growth boundaries, and many other policies.
By treating urban development as the interaction between market behavior and governmental
actions UrbanSim is designed to maximize reality, thereby increasing its utility for assessing the
impacts of alternative governmental plans and policies related to land use and transportation.
Thus, the model design enhances the strategic planning capabilities of MPOs and other state and
local agencies needing to evaluate growth management policies such as urban growth
boundaries, assess consistency of land use and transportation plans, and address conformity
with respect to air quality implementation plans.
Running the model requires exogenous input information derived from:





Population and employment estimates
Regional economic forecasts
Transportation system plans
Land use plans
Land development policies such as density constraints, environmental constraints, and
development impact fees
The user interacts with UrbanSim to create "scenarios," specifying alternative packages of
forecasts, land use policy assumptions, and other exogenous inputs. The model is then executed
for a given scenario, and the results of one or more scenarios can be examined and compared.
Appendix 2 – Page 1
UrbanSim excels in its flexibility to disaggregate households, businesses, and land use. The
classification detail is a function of the needs of the user and available data, but as currently
structured, its output information includes:








Future year distributions of population
Households by type (e.g. income, age of head, household size, presence of children, and
housing type)
Businesses by type (e.g. industry and number of employees)
Land use by type (user-specified)
Units of housing by type
Square footage of nonresidential space by type
Densities of development by type of land use
Prices of land and improvements by land use
In the area of user-benefits, there is considerable controversy about what the most appropriate
measures are, and therefore there are a variety of measures provided in the evaluation
component. Transportation infrastructure characteristics are input by the user to the travel
demand modeling process. The model does not predict infrastructure characteristics, but can use
such information to predict development. The components exist to add functionality to account
for the costs of infrastructure as part of the evaluation of alternative scenarios.
By developing a model that is behavioral in its approach, the operation of UrbanSim becomes
fairly simple to understand, but is able to capture complex interactions in the markets for land,
development, and transportation. It is a valuable tool for improving the level of understanding of
how a metropolitan region is developing and how various combinations of land use and
transportation policies and investments are likely to shape these trends. Some of the issues of
interest, such as affordable housing, are within the scope of the model to be of use, since it deals
with predicting housing prices, and disaggregates households by income as well as other
characteristics, and can capture the affordability impacts of alternative scenarios. Preservation
of land in green space would be feasible to incorporate within the model by earmarking specified
parcels for green space preservation, which would influence the supply of land, and could be
tested as an attractor for residential or business location. Urban design issues could similarly be
explored, given the parcel-level capacity of the developer module, and the ability to incorporate
a flexible set of terms in the location choice equations for businesses and households. The
specific abilities to test these and other policy issues of interest depend on myriad factors being
considered as this planning tool evolves.
Water Demand Modeling In UrbanSim
In the spring of 2006, a prototype water demand forecasting model was developed using outputs
from UrbanSim. Bi-monthly water demand was generated using City of Seattle residential
sector consumption data. The model developed is currently being expanded to include weather
variables and historic census data. The model will be run into the future, with climate change
Appendix 2 – Page 2
forecasts to asses possible demand impacts. A shorter time-step (weekly) will be used in the
final model and the region modeled will be expanded.
Methodology
Selected Modeling Approach
After examining current water demand forecasting methods, a multiple regression model
based on available land use and demographic information from UrbanSim and PSRC forecasts.
This approach is essentially an econometric model. However, the model’s determinants also
relate it to a variable flow model. Determinants chosen include population, land use, percent
residential, average income per housing unit, and lagged demand.
A multiple regression model is a multivariate statistical technique, which examines the
variable being forecasted (e.g. water demand) and multiple other variables (e.g. population, lot
size). Multiple regression models have the advantage of forecasting water demand while
considering forecasts of its determinants (e.g. population forecast). The equation for a multiple
regression resembles:
y = ß0 + ß1x1 + ß2x2 + ß3x3 + ... + ε
Where:
y = dependent variable, e.g. water demand
x = predictor variable, e.g. population
ß = regression coefficients
ε = residual term
When using a multiple regression model, it is important to minimize the cross-correlation
between predictor variables. If predictor variables are related, a change in one variable will
cause other variables to change, resulting in an exaggerated change in the dependent variable.
Although relationships between predictor variables may be unavoidable, an effort to minimize
collinearity is important for maximizing the prediction power of the regression model.
Utilization of UrbanSim
Regression coefficients can be calculated inside UrbanSim by using population data for
the present year. Additionally, UrbanSim has the capability to create “rolled-back” land use data,
which removes housing, commercial lots, industrial lots, and other parcel features based on when
the buildings were created. However, historic population information is not available on the
gridcell level, which impacts the ability to perform the regression over a longer historic period.
If gridcell population is used as a predictive variable, the regression in UrbanSim can only
include information from 2000 on. By using past consumption data from Seattle Public Utilities
(SPU) for the dependent variables (aggregated to the gridcell level) and information from
UrbanSim for the predictive variables, regression coefficient values can be obtained.
UrbanSim forecasts the regional distribution of population, households by type,
businesses by type, land use by type, units of housing by type, square footage of nonresidential
space by type, densities of development by type of land use, prices of land and improvements by
land use. UrbanSim forecasts can be used as input in regression models as predictive variable
forecasts. Therefore, to create a water demand forecast, the regression coefficients (derived in
Appendix 2 – Page 3
UrbanSim using current information for the predictive variables) are used with UrbanSim
forecasts of the predictive variables.
Model Verification
Once developed, the water demand forecast is compared to other water demand forecasts
for the City of Seattle. Results from SPU’s current forecasting method, in addition to previous
forecasting methods, can be weighed against outputs from the water demand module
incorporated into UrbanSim. This comparison provides an indication of the reasonableness of
the forecast. It should be noted that water demand forecasts traditionally over-predict
exaggerated future water demands.
Data Collection and Processing
To create water demand forecasts, observed data is needed. The City of Seattle has for
many years kept a thorough database of bimonthly consumption data at the parcel level. This
was used to create the consumption record in UrbanSim.
Dependent Variable - Water Demand
To accurately forecast water demand, historic data are typically used to calibrate and
validate the model. For model calibration bimonthly billing and consumption data from 1992 –
2001 was used.
Access Databases of Historic Consumption Data by Parcel
Microsoft Access databases of historic consumption data were available. These data
were reformatted for inclusion into the UrbanSim SQL database. The database was first related
to a parcel database because the SPU account information only had account id’s. By relating the
accounts to the parcel it was possible to later relate the parcels to the gridcells that are used in the
UrbanSim simulation. After the accounts were related to the parcels, each parcel was divided
into sector demand. Four sectors were created from the original database; these were
commercial single and multiple register parcels, and residential single and multiple register
parcels. After the sectors were separated into their separate databases, the data was prepped and
then read into the MYSQL database by sector.
Conversion to MYSQL Database with Consumption by Gridcell
After the data were read into the MYSQL database, the process of relating the parcel to
the gridcell could occur. Through this relating process, a fraction of demand from each parcel
was assigned to a gridcell, because not all the parcels lie exactly within the gridcell layout. After
the parcel data was joined to the gridcell through this fractionation process, the data was ready to
be read into UrbanSim. Slow run times were eliminated by writing the appropriate data per
regression run to a float file before reading into UrbanSim to allow faster access.
Independent Variables: Weather
Weather is a key predictor for peak consumption when forecasting demand on a daily
timestep. This is typically attributed to outdoor lawn watering, car washing, and other outdoor
water use though some of the increase is also increased measured as in-house usage because of
the temperature rise.
Appendix 2 – Page 4
Historic Weather Data
Historic weather data from 1991-2005 was tabulated and formatted for use in the
regression process. Maximum and mean bimonthly temperatures as well as total precipitation
values were created from daily time-series records that were obtained from the National Climatic
Data Center database. To be able to use the weather data as a predictor variable, a timeseries of
the UrbanSim data must be created. For coefficient determination, the timeseries created from
the unrolled data would have to be sampled over the period available and coefficients determined
from that. Because we are currently able to only use cross-sectional census data, weather will
have to be semi-omitted from the study. This is because the weather variable will act as a
constant, picking up variability that is not actually variability explained by the weather variables
because we are not able to create a timeseries of data for all UrbanSim variables back to the
1990’s.
Forecasts
If weather is included in the water demand module in the future, climate change impacts
on water demand could be researched. With temperature forecasted to increase 2-5 degrees F in
the next 100 years (NCDC, 2006), water demand patterns could become significantly different
because of increased peak demand as days grow warmer for longer periods of time and shifted
seasonal precipitation and temperature could potentially shift when water is demanded as well as
amount demanded. Climate change impacts coupled with changed land-use patterns could also
compound effects of one-another causing pattern shifts as well.
Gridcell Data
Historic (Unrolled)
To be able to use the weather data, historic parcel data, and lagged water demand as
predictor variables, historic parcel data had to be generated through a process of “unrolling”
UrbanSim back to 1990. The unrolling process involves eliminating parcels based on built year.
By removing parcels of certain built year, the data for each parcel is closer to representing the
past than if the parcels were left as is with the 2000 year data in place. After the data was
unrolled, variables could be predicted by sampling over all the years so that inter-year variability
could be explained by the coefficients. This also allows weather to act as a predictor variable
because multi-year data is being sampled and therefore allowing a trend to be predicted. One
thing the rollback does not do is generate population, income, or jobs into the past. Instead that
would have to be a separate effort that is outside the scope of this project but would provide data
so that a multi-year regression could be performed.
Forecasts
The gridcell data used for forecasting future years is generated by UrbanSim. As time
evolves in the UrbanSim model, data is sent from UrbanSim to the water demand model. The
coefficients that have been determined previously are used to predict water demand at each
gridcell. The gridcell water demand is then summed across the Seattle region for each billing
period to give bi-monthly and total annual demand.
Progress
Appendix 2 – Page 5
In this research, there were several area of progress and some challenges.
Regression
Throughout the course of the quarter important progress was made despite some
problems regarding data limitations and computational constraints. Historic water consumption
data was integrated into a MYSQL database, parcel scale data was converted to gridcell level
detail and water demand was calculated based on different combinations of predictor variables.
Several combinations of predictor variables were tested and a maximized R2 value of
0.22 was determined. The number of predictor variables were minimized by keeping predictors
that increased the R2 value by more than a percent and had a significant t-value. The following
gridcell variables proved to be useful for increasing the power of our regression equation:
Constant value
Average income per housing unit
Natural logarithm of the total land value
Natural logarithm of the total population within walking distance of a gridcell
Percent residential
Demand lag (the gridcell’s demand during the same bimonthly period of the previous year)
All
variables
highly
however,
remained
January-February
Adjusted R-Squared: 0.219192752919
----------------------------------------------Coeff_names
estimate
SE
constant
-602.526
14.8282
average_income_per_unit 0.000505
5.69E-05
land_value
33.6819
1.01359
population
18.8016
1.19527
percent_residential
1.37066 0.0349323
t-values
-40.6339
-8.86989
33.2301
15.7299
39.2377
of the
used were
significant;
the R2 value
low. Water
consumption is subject to the dynamics of human behavior, which is difficult to predict in a
model. Additionally, modeling water consumption on such a small (150 meters by 150 meters)
grid might overstate the difficulties of modeling water demand on a regional scale. In other
words, using the regression equation on regionally-scaled data might provide better R2 values,
not necessarily better results though.
Forecasting
The initial forecasts cover the years 2001 to 2005, enabling comparison of our results to
actual consumption values and to forecasts from SPU. After creating forecasts which used
demand lag as a predictor variable, it was realized that water consumption was increasing
significantly every year. Therefore it was decided to run forecasts without demand lag in order
to see if the results would become more realistic.
The results of our forecasts are below. Although the prediction power and accuracy of
the forecasts are not sufficiently accurate for water resources planning based on the five year test
forecasts, the progress made thus far gives a strong foundation for creating a more robust model.
Appendix 2 – Page 6
Results and Conclusions
Based on initial results, UrbanSim does a good job of identifying the seasonality of water
demand without having a weather variable for prediction. This is because the water demand
model is divided into bi-monthly periods for the regressing of the coefficients. This allows the
model coefficients to have varying magnitude for each season and so the model detects the
changing trends for each season. The demand lag variable proved to be a poor predictor because
it forces the use of last years demand forecast to effect the next years. This seems to be causing
the demand to “blowout” and increase consumption by exorbitant amounts for each year
forecasted. The model with the lag demand variable was increasing consumption by about 10
MGD a year, whereas the model without the lagged demand variable was increasing by only 2
MGD over 5 years, which is much more realistic.
Demand Forecast With Lag
175.00
Consumption (MGD)
155.00
135.00
2001
2002
115.00
2003
2004
95.00
2005
75.00
55.00
1
2
3
4
5
6
Bi-monthly Period
Demand Forecast Without Lag
Consumption (MGD)
145.00
2001
125.00
2002
2003
2004
2005
105.00
85.00
1
2
3
4
5
6
Bi-montly Period
Challenges
In attempting to develop a functional water demand model in the Open Platform for
Urban Simulation (OPUS) environment, several problems were encountered. It has been noted
Appendix 2 – Page 7
that, with respect to urban simulation, “Early simulation efforts were hampered by limitations in
data, computational capacity, modeling techniques, behavioral theory, and did not meet the
optimistic expectations set out for them” (Waddell, 2006). Despite advances in several of these
areas, significant limitations still abound.
Data Limitations
Historic Data
The first significant hindrance to the project was a lack of sufficient data. To develop
coefficients for a regression model, it is necessary to have accurate historical data for all
variables. However, historic data prior to the year 2000 was not readily available for the
proposed independent variables. As described above, the historic values had to be “unrolled”
from more recent data, and the accuracy of the results of this procedure is highly uncertain. As
such, coefficients were derived using the five years worth of available data (2000 – 2005).
However, doing so eliminated the opportunity to calibrate the model by comparing predicted
water demands to actual values during these years.
Employee Data
Because water is consumed on a per-person basis, the population forecasts is often a key
variable for demand prediction. Unfortunately, historic values for the number of employees in a
gridcell were not available for this study. A regression model developed for commercial water
demand would have lacked information for a significant explanatory variable, severely limiting
the model’s explanatory power. Because of this, resources were focused on generating a model
to predict residential demands.
Database Size
The large size of the input databases quickly became an additional concern. Containing
ten years worth of consumption data distributed by parcels, some of the databases consisted of
millions of records. This led to difficulties when the data was prepared and processed in order to
develop MYSQL databases.
Complex Coding
Time Series Data
The need to perform a regression on time series data presented further difficulties.
Unlike the sample models presented in class, the water demand model was to be developed using
ten years worth of historical data, instead of just one. Because of the seasonal nature of water
demand, it was necessary that the model be run at a smaller, bimonthly time-step, as opposed to
the one year time-step. It was not immediately clear how these deviations might be incorporated
into the model code. While the model currently runs at the bimonthly time-step, regression
coefficients are generated using one year’s worth of data only. Unfortunately, this precludes the
use of weather data, which is constant across all gridcells for a given bimonthly period and year.
Behavioral Theory
Despite advances in behavioral theory, very little is known about the way in which
individuals make water consumption choices. For example, demand elasticities are a measure of
Appendix 2 – Page 8
how much demand varies with a given change in price. However, there is no agreed upon value
for the demand elasticity of water and various studies has produced a wide range of possible
values. Because individuals tend to behave differently across time, geographies, social classes,
etc., a great deal of uncertainty persists when forecasting water demand.
Further Investigation
Address Challenges
To develop a fully functional water demand forecasting model for a region, such as the
City of Seattle, the challenges described above need be addressed to the extent possible.
Datasets should contain at least 15 years worth of historic values and information for all
demographic, economic, and climate variables so that a reliable regression model can be
produced. This data may be available from the PSRC and/or other local agencies, but it may
require significant processing before it can be incorporated into the model.
Enhance the Model
With capable coders and sufficient data, the model can be greatly improved upon such
that it may meet the goals initially laid out for this project. Regression coefficients can be
calculated using data from multiple years, weather data can be incorporated, and a model for
commercial demand can be developed.
In the future, the demand model can be further enhanced in several ways. First,
additional explanatory variables can be examined and potentially integrated into the current code.
Variables of interest that are not currently available might describe water price, conservation
efforts, income distribution, lot sizes, drought events, economic recession, and more. Some of
these variables can be created within the Opus environment using available UrbanSim data.
Others can be generated outside of the model and then incorporated as input tables.
SPU supplies water not just to inner-city communities, but to surrounding areas as well.
The model can be adapted to predict demands in these areas as well. This task will likely prove
challenging. Water delivered to outlying regions is generally wholesale, and, in some cases,
historic data for both dependent and independent variables may not be available at the gridcell
scale.
Finally, conducting a more extensive statistical analysis of the data may prove to be
advantageous. Input data can be examined for outliers, correlated variables, and the need for
dataset transformations by performing exploratory data analyses. Accounting for the findings of
these exercises may result in more consistent and accurate results.
Appendix 2 – Page 9