The value of real-­‐-me data in controlling electric loads for What%might%the%low%infrastructure%case%%
%%look%like?%
demand response MW%
substa)on%power%
consump)on%
)me%
distribu)on%
substa)on%bus%
MW%
central%
controller%
TCL%power%
consump)on%
MW%
market%signal%
load%state%
es)mator%
)me%
)me%
broadcast%
control%
Johanna Mathieu, Mechanical Engineering Duncan Callaway, Energy & Resources Group University of California, Berkeley 11/10/11%
thermosta)cally%
controlled%loads%
(TCLs)%
J.%Mathieu,%UC%Berkeley%
Carnegie Mellon Conference on the Electricity Industry: March 12-­‐14, 2012 10%
As more wind and solar are added to the grid there is more need for load following and regula-on. [Makarov et al., “Opera-onal Impacts of Wind Genera-on on California Power Systems,” 2009]
• These services could be provided by new generators, energy storage, and/or demand response. • We usually think of using LARGE loads for Demand Response (DR). • In our work, we simulate small residen-al loads. Why small?? -­‐ more reliable -­‐ simple local controls 3/13/12 [source: CAISO] -­‐ spa-ally distributed -­‐ con-nuous, not discrete, control response Mathieu & Callaway, UC Berkeley 2 Thermosta-cally Controlled Loads (TCLs) • Refrigerators, water heaters, air condi-oners, electric space heaters, etc. • Hystere-c ON/OFF control (dead-­‐band) • Store thermal energy like ba`eries store chemical energy 3/13/12 TCLs"
Mathieu & Callaway, UC Berkeley 3 Two Op-ons for Revenue 1. Par-cipa-on in ancillary services markets & fast -mescale energy markets GOAL: market signal tracking 2. Energy arbitrage: buy more low, buy less high GOAL: minimize energy costs 3/13/12 Mathieu & Callaway, UC Berkeley 4 Market Signal Tracking central controller distribu-on substa-on bus MW market signal TCL power consump-on -me -me broadcast control 3/13/12 Mathieu & Callaway, UC Berkeley 5 Market Signal Tracking Offline and real-­‐-me info/data from loads central controller distribu-on substa-on bus TCL power consump-on MW market signal load state es-mator -me -me broadcast control 3/13/12 Mathieu & Callaway, UC Berkeley 6 Market Signal Tracking Offline info from loads; Real-­‐-me data from substa-on MW substa-on power consump-on central controller -me distribu-on substa-on bus TCL power consump-on MW market signal load state es-mator -me -me broadcast control 3/13/12 Mathieu & Callaway, UC Berkeley 7 Market Signal Tracking Real-­‐-me data from substa-on MW substa-on power consump-on central controller -me distribu-on substa-on bus TCL power consump-on MW market signal load state es-mator -me -me broadcast control 3/13/12 Mathieu & Callaway, UC Berkeley 8 for cooling devices
(3)
for heating devices
!""#$"%&$'()*+(,-'$./(
Heterogeneous TCL P0&%&$(123(&#%3425-3(6-'$.(
opula-on Model xk+1 = Axk + Buk
y k = Cxk
I(
(4)
ON
Nbin
Nbin-1 Nbin-2 Nbin-3
(5)
state
=
Ri Prate,i
−Ri Prate,i
1
OFF
2
3
4
Nbin Nbin Nbin Nbin
+1
+4
+3
+2
2
2
2
2
Nbin Nbin Nbin
Nbin
-2
-1
-3
2
2
2
2
I(
x -­‐ frac-on of TCLs in each state bin normalized temperature
A -­‐ transi-on matrix u -­‐ input vector that allows us to switch TCLs ON or OFF
!(,%#7-8()#%3425-3(,%&#29('$4:#21$4(&;$(6-8$6$3&(-<()*+4(%#-=3'(
&;$('$%'>1%3'?(
y -­‐ aggregate power (and possibly a measurement of x) @@B@CB@@(
D?(,%&;2$=E(F*(G$#7$.$H(
@A(
Key control assump-ons: • We can only switch TCLs on or off, not change the temperature set point, etc. • We broadcast same control to all TCLs, which switch on or off probabilis-cally based on their current state. • We can’t switch TCLs outside of the dead-­‐band. [Mathieu & Callaway HICSS 2012; Koch, Mathieu, & Callaway PSCC 2011] 3/13/12 Mathieu & Callaway, UC Berkeley 9 Tracking Results: 5-­‐minute market signal One-­‐step look-­‐ahead propor-onal predic-ve controller 1,000 heterogeneous TCLs Scenario 3 -­‐real-­‐-me data: substa-on power -­‐offline parameter es-ma-on Scenario 4 -­‐real-­‐-me data: substa-on power -­‐online parameter es-ma-on kW
Ptotal,real
5% Forecast Error
10% Forecast Error
5% Forecast Error
10% Forecast Error
kW
Scenario 3:
30% Metering
500
0
−500
Scenario 4:
500
0
−500
0
3/13/12 Ptotal,des
500
0
−500
100% Metering
kW
Scenario 2 -­‐real-­‐-me data: on/off state -­‐offline parameter es-ma-on Reference Case
Scenario 2:
500
0
−500
kW
Scenario 1 -­‐real-­‐-me data: full state -­‐offline parameter es-ma-on Scenario 1:
0.5
Hours
Mathieu & Callaway, UC Berkeley 10
0.5
Hours
1
10 Energy Arbitrage (Preliminary Approach & Results) • Three Levels of Control: 1.
2.
3.
Linear program to solve for the op-mal trajectory TCL popula-on controller (currently, a propor-onal controller) a popula-on of individual TCLs, each with its own local controller • Ambient temperature is very important: Energy & Power Capacity of 1,000 Air Condi-oners or Heat Pumps Power capacity
6000
400
5000
4000
300
kW
kWh
Energy capacity
500
AC steady state
AC capacity
Heat pump steady state
Heat pump capacity
3000
200
2000
100
0
1000
0
10
20
30
40
Outdoor air temperature ( °C)
3/13/12 0
0
10
20
30
40
Outdoor air temperature ( °C)
Mathieu & Callaway, UC Berkeley 11 LP Set-­‐up Goal: minimize energy cost J = l(1)p(1)ΔT + l(2)p(2)ΔT + … + l( N )p( N )ΔT
subject to: e(k + 1) = e(k) + ( p(k) − pss(k))ΔT
€
0 ≤ e(k) ≤ emax (k)
0 ≤ p(k) ≤ pmax (k)
€
Itera-ve set-­‐up: –€ re-­‐run LP at each -me-­‐step €
– implement only the first control Price forecast error model: ω (k + 1) = γω (k) + αε (k)
Temperature forecast error standard devia-on: β
3/13/12 Mathieu & Callaway, UC Berkeley €
12 LP Trajectory: 1 hour predic-on horizon, itera-ve 5−minute price
$/MWh
400
200
0
3892
3894
3896
3898
3894
3896
3898
kWh
600
3900
3902
State of charge
3904
3906
3908
3910
3904
3906
3908
3910
400
200
0
3892
3900
3902
controlled consumption
uncontrolled consumption
Power Usage
kW
6000
4000
2000
0
3892
3/13/12 3894
3896
3898
3900
3902
hour of year
3904
Mathieu & Callaway, UC Berkeley 3906
3908
3910
13 Tracking Performance: P-­‐controller (for now) aggregate power consumption (kW)
6000
Iterative LP Trajectory
Actual TCL Population
5000
4000
3000
2000
1000
0
4615
4620
4625
4630
hour of year
Issues: • the resource is over-­‐es-mated in the morning: power and energy capacity is a func-on of current and PAST ambient temperature • the propor-onal controller isn’t great, plus some TCLs are outside of the dead-­‐
band and uncontrollable 3/13/12 Mathieu & Callaway, UC Berkeley 14 Energy Cost Savings Actual savings Air condi-oners 25
savings (%)
Noisy forecasts* 12.2%
9.0%
Heat Pumps
10.2%
6.7%
Water Heaters
41.0%
16.4%
Refrigerators
27.1%
13.6%
Air condi-oners
Savings predicted by LP
Actual savings
20
Perfect forecasts
15
10
* α = 40$/MWh β = 3°C γ = 0.3 5
0
0
10
20
30
40
50
α ($/MWh)
60
70
80
β = 3°C γ = 0.3 3/13/12 Mathieu & Callaway, UC Berkeley 15 Per-­‐TCL Capital and Annual Costs Required to Break Even [Regula-on and load following revenues es-mated with data from Eyer & Corey 2010 (SAND2010-­‐0815 ). Arbitrage revenue es-mated with interval LMP data from California’s Merced node. Details in Mathieu et al. ACEEE 2012 (forthcoming).] 3/13/12 Mathieu & Callaway, UC Berkeley 16 8
GW
GWh
10
6
Water heaters
Refrigerators
30
20
4
How BIG is the resource poten-al? 10
2
0
2000
4000
6000
0
8000
2000
Hours
4000
6000
8000
Hours
Es-mates for most of California (5 Scenario
largest u-li-es) based on RECs and CEC data. B. 2020 Base
GWh
GWh
2012 Resource Dura-on Curve 14
12
12
10
10
8
8
6
6
4
4
2
2
0
0
GW
GW
A. 2012 Base Scenario
Energy Capacity
14
Energy Capacity
2000
2000
4000
6000
4000
Hours 6000
8000
8000
50
50
40
40
30
30
20
20
10
10
0
0
Power Capacity
Power Capacity
Central AC
Heat pumps
Water heaters
Refrigerators
2000
2000
Hours
3/13/12 50
50
Power Capacity
Power Capacity
40
40
GW
GW
GWh
GWh
14
12
12
10
10
8
8
6
6
4
4
2
2
0
0
30
30
20
20
10
10
2000
2000
4000
4000
Hours
Hours
6000
6000
8000
8000
0
0
2000
2000
4000
4000
6000
Hours 6000
Hours
[Mathieu et al. ACEEE 2012 (forthcoming)] C. 2020 Electrification Scenario
Mathieu & Callaway, UC Berkeley 14
12
10
8000
8000
Hours
C. 2020 Electrification Scenario
B. 2020 Base Scenario
Energy Capacity
Energy Capacity
14
2020 Resource Dura-on Curve, assuming increased efficiency and 30% of water/space heaters converted to electric 4000
6000
4000
Hours 6000
Energy Capacity
50
40
Power Capacity
8000
8000
17 Key Takeaways • TCLs can be controlled to follow signals and s-ll provide the service requested by the consumer. • Possible revenue sources are par-cipa-on in energy/ancillary services markets or energy arbitrage. • The revenues may cover the costs. • The resource poten-al is large. 3/13/12 Mathieu & Callaway, UC Berkeley 18 Thank you! Ques-ons? The value of real-­‐-me data in controlling electric loads for demand response Johanna Mathieu jmathieu@berkeley.edu Thanks to Stephan Koch (ETH Zurich), Mark Dyson (UC Berkeley), and the PSERC Future Grid Ini-a-ve. 3/13/12 Mathieu & Callaway, UC Berkeley 19