1 Dealing with Uncertainty in Water Distribution Systems’ Models: a Framework for 2 Real-Time Modeling and Data Assimilation 3 4 Christopher J. Hutton1, Zoran Kapelan2, Lydia Vamvakeridou-Lyroudia3, and Dragan A. SaviΔ2 5 6 Abstract 7 Water Distribution System (WDS) models when applied with real-time data may improve 8 system control, and in doing so, help meet consumer and regulatory demands. Such real-time 9 modelling often overlooks the multiple sources of system uncertainty that cascade into model 10 forecasts, and affect the identification of robust operational solutions. This paper considers key 11 uncertainties in WDS modelling, and reviews promising approaches for uncertainty quantification and 12 reduction in the modelling cascade, from calibration, through data assimilation, to model forecasting. 13 An uncertainty framework exemplifying how such methods may be applied to propagate uncertainty 14 through the real-time control process is outlined. Innovative methods to constrain uncertainty when 15 the time-horizon and data availability limit such thorough analysis are also discussed, alongside 16 challenges that need to be addressed to incorporate uncertain information into the control decision. 17 Further work evaluating the value of these methods in light of computational resources, and the nature 18 of model errors in real WDS is required. Such work is necessary to demonstrate the benefits of 19 considering model and data uncertainty, leading to robust control decisions. 20 Subject Headings: Water Distribution Systems; Uncertainty Principles; Numerical Models; Control 21 22 ________________________________________________________________________________________________________________________________ 1 Associate Research Fellow, College of Engineering, Mathematics and Physical Sciences, University of Exeter, Exeter, Harrison Building, North Park Road, Exeter EX4 4QF, U.K. Email: c.j.hutton@ex.ac.uk 2 Professor, College of Engineering, Mathematics and Physical Sciences, University of Exeter, Exeter, Harrison Building, North Park Road, Exeter EX4 4QF, U.K. 3 Senior Research Fellow, College of Engineering, Mathematics and Physical Sciences, University of Exeter, Exeter, Harrison Building, North Park Road, Exeter EX4 4QF, U.K. 23 Introduction 24 Improved operation of Water Distribution Systems (WDS) is required to meet the (potentially) 25 competing objectives of consumer expectation, regulatory requirements and shareholder satisfaction 26 (Bakker 2003; Ofwat 2005; Ogden and Watson 1999). Unnecessary leakage and energy costs occur in 27 many systems as a result of limited system control, which if improved may lead to better operational 28 performance (Jamieson et al. 2007), and mitigate the need for expensive infrastructure investment - a 29 cost ultimately borne by consumers. 30 Sensor development has increased the availability of online monitoring data (Storey et al. 31 2011), which can inform end-users of the states of a WDS. Actuator settings can then be modified in 32 real-time to improve operational performance and deal with system anomalies. Such an approach is 33 heavily reliant on the skills and experience of the system operator to make the most robust 34 intervention, in response to (often) sparse measurements. The potential for improved operational 35 performance when monitoring data is supplemented with on-line system modelling has been 36 demonstrated in a number of studies (Machell et al. 2010; Preis et al. 2010; Rao and Salomons 2007). 37 However, both the data and models used in such Decision Support Systems (DSS) contain 38 considerable uncertainty (Hutton et al. 2011). A key issue not considered fully is the propagation of 39 uncertainty from model calibration, through data-assimilation to real-time model forecasting. 40 This paper presents a framework considering uncertainty in the key stages of WDS model 41 development and application for Real-Time Control (RTC), and provides a critical review of methods 42 applied to quantify and reduce uncertainty at each of these stages. The paper does not intend to 43 provide an exhaustive review but seeks to critically highlight and classify a range of methods applied 44 in the WDS literature. The review also includes promising methods for dealing with model 45 uncertainty applied in related scientific fields, and considers the key issues governing their application 46 in the context of WDS control. 47 This paper is organised as follows. First, a review of uncertainties in WDS modelling is 48 considered, followed by a framework specifying the cascade of uncertainties that exist in producing a [Type text] 49 model for real-time application. Key methods for dealing with uncertainty in Calibration, and Data 50 Assimilation, are then reviewed. The ideal case of propagating all uncertainties by ensemble 51 representation into model forecasts, and alternative methods given current limitations in data and 52 computational resources is then considered. Finally, conclusions and recommendations for future 53 research are presented. 54 55 Real-time Control Modelling and Uncertainty in WDS 56 Coupling hydraulic system models with on-line system observations – in what is termed a 57 decision support system (DSS) – allows the extrapolation of system measurements both in space, by 58 simulating unmeasured locations, and in time by simulating future system states. Developments in 59 computational power, simulation software (e.g. EPANET2; Rossman 2000), calibration tools (Savic et 60 al. 2009), and sensor development (Storey et al. 2011) have facilitated the application of real-time 61 WDS modelling, typically for two purposes: 62 ο· 63 64 Provide warnings of future system states (Pappenberger et al. 2005), such as pipe burst (Bicik et al. 2011a; Misiunas et al. 2005). ο· Constrain understanding of operational system states, and explore a range of control strategies 65 for optimal operation (Martinez et al. 2007). 66 This paper focuses on modelling uncertainty during normal operating conditions in WDS, 67 although the methods considered are also relevant for anomalous system conditions (Bicik et al. 68 2011a). Although a number of studies have demonstrated the potential improvements of using on-line 69 models for system control (Machell et al. 2010; Preis et al. 2010; Rao and Salomons 2007), both the 70 data and models used in such DSS contain considerable uncertainty (Bicik et al. 2011b). 71 Uncertainty may be divided into two categories (Hall 2003): (1) Aleatory uncertainty, such as 72 the irreducible spatial and temporal variability in water demand, and (2) Epistemic uncertainty, which 73 results from incomplete system knowledge. Numerical models contain known epistemic uncertainty, [Type text] 74 which may be reduced, or tolerated (e.g. skeletonisation), but never fully eliminated. An explicit 75 incorporation of either type of uncertainty into WDS models is typically lacking. At best this 76 represents bad practice; at worst, in the context of control decisions and risk management, misplaced 77 confidence in deterministic predictions may lead to undesired consequences. 78 Methods for quantifying and reducing uncertainty in systems models have been developed in 79 hydroinformatics (Hall, 2003) and related scientific fields (Beven 2006; Kavetski et al. 2006a). To 80 apply such methods for real-time WDS control, sources of uncertainty in WDS modelling first need to 81 Μ (e.g. EPANET2 (Rossman 2000)), containing equations that be considered. A general model, π 82 Μ = {xΜ1 , … , xΜπ } with length p represent the functional relationship between a vector of system states, π 83 Μ π¨ = {xΜ1 , … , xΜπ }, also with length p (e.g. (e.g. Nodal Pressure), given a set of initial system states π 84 Μ = {θΜ1,..., θΜd}, with length d Tank Levels, Pump and Valve Settings), a vector of model parameters π 85 Μ = {dΜ1 , … , dΜπ } with (e.g. Pipe roughness and pump curves), and a time-series of driving conditions, π 86 length n (e.g. water demand): Μ, π Μ=π Μ (π Μπ¨, π Μ) π 87 (1) 88 The circumflex (hat) indicates the uncertain nature of the model variables. Three types of model 89 uncertainty need to be dealt with in WDS models (Hutton et al. 2011), as discussed below. 90 Model Structural Uncertainty refers to errors in the mathematical representation of reality, and is a 91 Μ ≠ π). Examples include: form of epistemic uncertainty, where the model will never equal reality (π 92 ο· Skeletonisation - the removal of pipes not considered essential for system analysis – 93 represents one of the key WDS model structural uncertainties. Skeletonised models may 94 neglect dead ends and high elevation nodes in the network, and adversely affect pressure 95 surges (Boulos et al. 2004), demand satisfaction predictions (Walski et al. 2003), contaminant 96 consequence assessment (Bahadur et al. 2006) and chlorine decay simulation (Menaia et al. 97 2003). Whilst skeletonised models may be hydraulically equivalent to all pipes models for 98 steady state conditions (Jung et al. 2007; Preis et al. 2011), they can perform poorly in [Type text] 99 transient conditions (Jung et al. 2007). Models that include all pipes (e.g. Jacobsen and 100 Kamojjala 2009) may be computationally unfeasible in real-time, while increased data 101 requirements for calibration may outweigh structural uncertainty. 102 ο· Water demand is typically aggregated at junction nodes in WDS models, yet consumers 103 extract water from along pipes within the network. Although head loss corrections to 104 overcome this simplification have been developed (Giustolisi 2010), accurate specification of 105 distributed demand is difficult. 106 ο· Demand driven models may be considered valid for normal operating conditions in well 107 designed and maintained WDS, whereas pressure driven solutions are more appropriate in 108 cases of fire flow, pipe leakage and valve closure (Giustolisi et al. 2008a; Giustolisi et al. 109 2008b). The latter approach requires additional data to determine the relationship between 110 pressure head and flow (Ozger and Mays 2004), which are not usually accurate, and increased 111 computational time, which is not always available for real time computations. 112 113 Parameter Uncertainty reflects uncertainty in equation variables used to represent system 114 components (e.g. pipe roughness). Such uncertainty is aleatory, as parameters can vary over space and 115 time, and epistemic as system discretisation in space and time can result in a failure to reconcile the 116 scale observations with model parameters. Parameter values are often therefore ‘effective’ (Lane 117 Μ ≠ π). 2005) in that they produce the correct prediction, but often have little physical meaning (π 118 ο· Pipe roughness is problematic to identify accurately in WDS models as it cannot be directly 119 measured, and because of pipe deterioration (Boulos et al. 2004; Kleiner and Rajani 2001), 120 roughness changes with pipe age. Roughness values calibrated using junction pressure 121 measurements (Kapelan et al. 2007; Savic et al. 2009) will reflect uncertainties in system 122 specification, roughness pipe grouping, and data uncertainty. 123 ο· Due to wear (Hirschi et al. 1998) pumps typically do not operate at the efficient point supplied 124 by the manufacturer (Walski et al. 2003), and alongside valve settings, may need to be 125 considered in the calibration problem. [Type text] 126 Measurement/Data Uncertainty refers to uncertainty in quantities used to define initial conditions 127 Μ π¨ ≠ π π¨ ), model inputs (π Μ ≠ π), and model state observations (πΜ) utilised in evaluation of model (π 128 predictions (πΜ ≠ π). Such uncertainties result from instrumentation errors, and mismatches between 129 the scale of observations and predictions (e.g., demand lumping and disaggregation). 130 ο· Measurement errors affect the accuracy with which system states may be quantified, both 131 through direct measurement, and indirectly when such measurements are employed to 132 calibrate and evaluate WDS model performance (Bargiela and Hainsworth 1989). 133 Independent quantification of these uncertainties is ideally required such that they may be 134 propagated into calibrated states and parameters, and ultimately model forecasts (see Muste et 135 al. (2012) for a review of methods for instrument error quantification). 136 ο· Aleatory demand uncertainty is large in WDS models, as demand fluctuates over a variety of 137 temporal and spacial scales depending on consumer type (Davidson and Bouchart 2006; 138 Herrera et al. 2010). 139 ο· Epistemic demand uncertainty results from a low density of metered houses, and the difficulty 140 of obtaining such information in real-time. Demand is more readily inferred by calibration to 141 measured pipe flow, water quality and DMA measurements (Branisavljevic et al. 2009; 142 Jonkergouw et al. 2008; Kang and Lansey 2009). However, such approaches may require 143 downscaling to individual network nodes (Kang and Lansey 2009). 144 145 A framework for dealing with uncertainty in real-time control of WDS 146 Μ to make real-time predictions requires that all of the terms on the Employing the model, π 147 right hand side of Equation 1 are reasonably constrained. A general procedure for achieving this is 148 Μ), such as roughness are calibrated off-line, presented in Figure 1. First, model parameters (π 149 Μ π¨ ), such as demand and nodal pressure (Preis et al. 2010), are following which, model states (π 150 estimated by assimilating real-time measurements into the model. Finally, when coupled to a system 151 demand forecast (e.g. Herrera et al. 2010), the model can simulate future system states in response to [Type text] 152 control options/objectives (Rao and Salomons 2007). The flow of data through such model 153 development is typically deterministic. However, to make robust control decisions, model predictions 154 should be made conditional on WDS model and data uncertainty.. 155 Uncertainties present in WDS have been considered extensively in the WDS literature. 156 However, assessing uncertainty in component parts of model development independently can lead to 157 an under-estimation of total uncertainty (Zappa et al. 2011), and miss-placed confidence in model 158 predictions. A number of frameworks for dealing with total model uncertainty have been presented in 159 the research literature, including the BAyesian Total Error Analysis framework (BATEA; Kavetski et 160 al. 2006a; Thyer et al. 2009), Global Assessment of Model Uncertainties (GAMU; Deletic et al. 161 2011), and the Generalized Likelihood Uncertainty Estimation procedure (GLUE; Beven and Binley 162 1992) amongst others (Liu and Gupta 2007; Schoups and Vrugt 2010). Such frameworks have mainly 163 focussed on the key issue of model calibration. However, the full propagation of uncertainty from 164 calibration (Section 4) through data assimilation (Section 5) to model forecasting (Section 6), and the 165 challenges in using such information to inform real-time control (Section 6.6) need to be considered 166 in the specific context of WDS (Figure 1). 167 168 Calibration and Parameter Uncertainty 169 Introduction 170 WDS model parameter calibration, which is reviewed by SaviΔ et al. (2009), is a necessary 171 step prior to applying a model to understand system operation. Model performance is evaluated by 172 Μ = {yΜ1 , … , yΜπ } with length n, to a comparing a time-series of the selected model response variable, π 173 vector of system observations, πΜ = {π¦Μ1 , … , π¦Μπ } to obtain a vector of model residuals: 174 Μ, π Μπ¨, π Μ ) = yΜπ (θ|π Μπ¨, π Μ ) − yΜ π ππ (θ|π π = 1, … , π (2) 175 The typical calibration approach applied to WDS models forces the residual errors as close to 176 zero as possible by adjusting the value of model parameters, either by a non-evolutionary optimisation [Type text] 177 method (e.g. Gradient Descent) or an evolutionary algorithm, (e.g. Genetic Algorithm; Savic et al., 178 2009; Table 1). Implicit in such an approach is the assumption that there are no other forms of 179 Μ = π), the initial states represent the true system uncertainty; that the model equates to reality (π 180 Μ π¨ = π π), the input drivers represent the true drivers of the system (π Μ = π), and that the states (π 181 observations are error free (πΜ = π). As reviewed above, this is not the case ; multiple parameter sets 182 may be found that produce equally likely (or behavioural) model predictions, a form of model 183 parameter equifinality (Beven 2006). 184 In light of parameter equifinality it is more useful and interesting to obtain information on 185 parameter and predictive uncertainty, which for each variable can be expressed in the form of a 186 posterior Probability Density Function (PDF). In WDS models this has typically been achieved post 187 calibration with the First Order Second Moment (FOSM) method (Bush and Uber 1998; Lansey et al. 188 2001). FOSM assumes model linearity, independence and normality of measurement errors and 189 parameter values (Kapelan et al. 2007). Further, it may not be applicable if the posterior parameter 190 distribution deviates from the multi-normal distribution (Vrugt et al. 2003). 191 In a range of modelling disciplines the preferred method to obtaining the posterior PDF is via 192 Bayes’ equation, which can be specified considering the joint inference of both model structure and 193 parameters (Draper 1995): 194 Μ|πΜ, π Μ, π Μ|π Μ,π Μπ¨, π Μ ) ∝ π(πΜ|π Μ,π Μπ¨, π Μ )π(π Μ )π(π Μ) (π 195 The first right hand term represents the likelihood function, the second term the prior parameter 196 distribution, and the third term the prior distribution of possible model structures. In most applications 197 a single model is applied, which collapses this final term to a single set of structural assumptions. 198 Solving Bayes’ equation analytically is typically intractable, and therefore some form of posterior 199 sampling is conducted; methods include random Monte Carlo sampling and Latin hypercube sampling 200 (Kang et al. 2009), and a more advanced family of Markov Chain Monte Carlo (MCMC) methods, 201 which use past information derived from the posterior distribution to inform the nature of further 202 sampling (see Vrugt et al. (2009a) for recent developments and methodological issues). Specification [Type text] (3) 203 of the likelihood function represents the key difference between formal and informal Bayesian 204 approaches (Table 1). 205 206 Formal Bayesian Calibration Approaches 207 The classical formal approach requires specification of a likelihood function which 208 necessarily makes strong assumptions regarding the nature of model (and measurement) errors. The 209 standard approach assumes that residual errors are mutually independent, Gaussian-distributed and 210 homoscedastic, leading to a Gaussian error model (or log likelihood for convenience; Vrugt et al. 211 2009b). Unfortunately residual errors often do not conform to such a simple distribution (Thyer et al. 212 2009), which can lead to bias in the posterior parameter PDF, and predictive distribution (Beven et al. 213 2008). 214 In the implicit formal approach, modifications to the likelihood function are made to account 215 for the different types of error, including adding autoregressive terms to account for correlated errors 216 (Bates and Campbell 2001; Sorooshian and Dracup 1980), Box-Cox transformations to 217 reduce/remove Heteroscedasticity and non-Gaussianity (Box and Cox 1982; Freni and Mannina 218 2010), and methods that apply a mixture of distributions in the error model (Schaefli et al. 2007). 219 More recently heteroscedastic, skew, kurtosis and bias parameters have been included in applied 220 likelihood functions that, alongside model parameters, have to be jointly inferred from the sampling 221 procedure (Schoups and Vrugt 2010). 222 Explicit approaches have also been developed, most notably through the BATEA framework, 223 that seek to represent explicitly different forms of error (Kavetski et al. 2006a). For example, in 224 hydrological application multipliers have been applied to account for measurement uncertainties in 225 input rainfall (Kavetski et al. 2006b). Such representation is via additional “latent parameters” (Thyer 226 et al. 2009; Vrugt et al. 2009b), that have to be inferred jointly with model parameters during the 227 calibration procedure. Such an approach is vulnerable to “ill-posedness” that results from the 228 difficulty of specifying a priori that nature of input and structural errors (Renard et al. 2010) – i.e. [Type text] 229 separation of errors in system skeletonisation and demand estimates. Whilst the implicit approach 230 does not attempt to separate out these sources of uncertainty, practical estimates of parameter and 231 total predictive uncertainty may be more readily derived (Schoups and Vrugt 2010). 232 An integral yet frequently ignored aspect of the formal inference procedure is the application 233 of posterior diagnostic checks to evaluate the error model hypothesises encapsulated in the likelihood 234 function (Thyer et al. 2009). Such checks include Quantile-Quantile (Q-Q) plots, autocorrelation 235 plots, and plots of residual error against output magnitude (Schoups and Vrugt 2010; Thyer et al. 236 2009; Yang et al. 2007). 237 238 Informal Bayesian Calibration Approaches 239 The GLUE procedure (Beven and Binley 1992; Beven and Freer 2001), more recently 240 referred to as an informal Bayesian approach (Smith et al. 2008a), seeks to find “behavioural” 241 parameter sets; that is, parameter sets consistent with the observations according to an informal 242 likelihood function (see Smith et al. (2008a) for a review). Following some posterior sampling, 243 parameter sets whose likelihood is greater than a user defined threshold are normalised to unity to 244 derive probabilistic information. Such an approach does not make assumptions about the nature of 245 residual errors in the likelihood function, and in doing so avoids potential over-conditioning of the 246 posterior parameter distribution (Beven et al. 2008). However, in doing so GLUE does not explicitly 247 consider other forms of uncertainty, which may lead to poorly constrained parameters (Stedinger et al. 248 2008). 249 Recent GLUE developments include the limits of acceptability approach (Beven 2006), which 250 like some formal Bayesian approaches, attempts to incorporate the effects of observation error in 251 model evaluation. The approach calculates a normalised evaluation score at each time-step to evaluate 252 where during a simulation a model is behavioural (Liu et al. 2009). The approach has also been 253 extended within the GLUE framework to consider model structural uncertainties (Krueger et al. [Type text] 254 2010). The MCMC SCEM-UA algorithm has also been adapted to adequately explore posterior 255 parameter space within GLUE (McMillan and Clark 2009). 256 Much debate exists in the modelling literature regarding the most appropriate application of formal 257 and informal Bayesian approaches (Beven et al. 2008; Mantovan and Todini 2006; Stedinger et al. 258 2008; Vrugt et al. 2009b), the suitability of simpler approximations for quantifying parameter 259 uncertainty (Gallagher and Doherty 2007; Kang et al. 2009), and the application of the probabilistic 260 approach more generally (Hall 2003) Probabilistic methods are potentially limited by the requirement 261 to select a uniform distribution in the face of ignorance about the probability of an event; however, 262 during such initial assumptions should no longer be influential, provided the model has been 263 confronted with sufficient data (Freni and Mannina 2010). The problem with such an approach is that 264 it conflates indeterminanacy with equiprobability (Dubois 2010; Hall 2003); natural variability and a 265 lack of system knowledge are merged and represented probabilistically. There is no theory to manage 266 epistemic model uncertainty, as by definition, it is poorly known (Beven and Alcock 2012). Unless 267 these sources of uncertainty (e.g. input measurement data and model structural error) have been well 268 defined a priori, posterior separation may be difficult (Renard et al. 2010; Willems 2008). When 269 aleatory and epistemic uncertainties are both present, imprecise probabilities have been employed, 270 where a family or interval of probability distributions is used to represent imprecision (Merz and 271 Thieken 2005). 272 273 Possibility Theory and Evidence Theory based Calibration 274 In possibility theory, available (typically linguistic) knowledge about a quantity may be expressed as 275 an interval, or a fuzzy number – which defines the imprecise degree of membership of an element 276 (such as a parameter) to belong to a specific value (Revelli and Ridolfi 2002). To propagate such 277 uncertainty through system models, a sampling procedure is required to reconstruct the posterior 278 possibility distributions of different system states. Such methods have been applied in WDS to 279 propagate uncertainties in pipe roughness (Revelli and Ridolfi 2002) and demand (Branisavljevic et 280 al. 2009). Fu and Kapelan (2011) quantified epistemic uncertainty in the probability distribution [Type text] 281 parameters of future water demand using a fuzzy procedure. Such an approach may be applied to 282 account for uncertainty in the parameters of a probabilistic error model; however, sampling may come 283 at significant computational cost when the inner (aleatory) and outer (epistemic) distributions need to 284 be sampled (Sun 2010). 285 286 Evidence theory is the simplest method of combining probability and possibility theory into the same 287 theoretical framework (Hall et al. 2007; Hall 2003). The basic probability assignment (BPA), m() (as 288 opposed to p() in probability theory) is assigned to sets as opposed to mutually exclusive singletons. 289 As a result, two measures of likelihood for a subset, belief Bel(A) and plausibility Pl(A) are obtained 290 from m(A) (Helton et al. 2004). The interval between belief and plausibility represents the range in 291 which the true probability may lie. Fu and Kapelan (Fu and Kapelan in press) combined probabilistic 292 information on rainfall uncertainty, with fuzzy model parameters using evidence theory. The ability of 293 evidence theory to better handle incomplete or imprecise information has resulted in greater 294 application in WDS to deal with anomalous real-time conditions (Bicik et al. 2011a; Sadiq et al. 295 2006). 296 Data Assimilation 297 298 Concept 299 To derive a real-time forecast the calibrated model requires an estimate of the initial model 300 Μ π¨ ). Data Assimilation (DA) is a name provided to a class of methods that combines the states (π 301 uncertain model with new (and uncertain) data to derive an estimate of the system state (Table 2). The 302 Μ π‘−1 ) to t in response to the model propagates forwards from a set of initial conditions at time t-1 (π 303 driving forces and a set of time invariant model parameters: 304 305 Μπ‘ = π Μ (θΜ, π Μ π‘−1 , π Μ ) + ωπ‘ π 306 [Type text] (4) 307 where ωπ‘ represents the model error term, with zero mean and covariance Vπ‘ . To assimilate 308 observations, model outputs need to be related to the model states, using an observation operator, Ht: 309 310 Μ π‘ ) + επ‘ πΜπ‘ = π»π‘ (θΜ, π (5) 311 312 where επ‘ denotes the observation error with zero mean and covariance R π‘ . Model states are updated 313 considering the relative difference between model and observation errors. 314 Advances in monitoring data, sensor placement and telemetric methods have facilitated the 315 collection of data in real-time for DA (see: Hart and Murray 2010; Ruggaber et al. 2007; Storey et al. 316 2011). Automated methods have also been developed to process these data (Branisavljevic et al. 2010; 317 Schilperoort et al. 2008), which should consider measurement uncertainty when determining between 318 good and bad data (Bertrand-Krajewski et al. 2003; Winkler et al. 2008). 319 A number of DA methods have been developed and applied to reduce uncertainty in real-time 320 model application (see Table 2), particularly in related scientific disciplines, including meteorology, 321 hydrology and climatology (Evensen 2003; Matgen et al. 2010; van Leeuwen 2009). 322 323 Predictor-Corrector Schemes 324 The basic Kalman Filter (KF) is a sequential filter method that provides a solution to 325 equations (4) and (5), provided that they represent a Gaussian Linear System (Burgers et al. 1998). 326 The state forecast (superscript f) is updated (superscript a) using system observations and the Kalman 327 gain, πΎπ‘ whose value depends on the relative magnitudes of observation and model error: 328 329 Μ ππ‘ = π Μ π‘π + πΎπ‘ (πΜπ‘ − π»π‘ π Μ π‘π ) π (6) 330 331 The KF algorithm has been applied to a WDS to estimate unknown roughness in a linear estimation 332 problem (Todini 1999). The KF has worked well for demand estimation in branched networks, 333 however, has performed poorly in nonlinear looped WDS (Kang and Lansey 2009). The KF update is [Type text] 334 sensitive to the (often limited) information available to constrain model and observations errors. 335 Furthermore, state update means the KF no longer conserves mass and momentum of the physical 336 system being simulated. Further research is recommended to understand general applicability of the 337 KF to WDS. 338 The Extended Kalman Filter (EKF) was developed to work better than the KF in cases of 339 strong system non-linearity, where the model is approximated with a tangent linear operator 340 (Jacobian; Evensen 2003). The EKF has been applied for water demand estimation (Nasseri et al. 341 2010) and also by Shang et al. (2006) for real-time update of demand estimates. In other scientific 342 fields EKF can provide comparable performance with ensemble methods (see Section 5.3; Reichle et 343 al. 2002), however is unsuitable in the case of large system non-linearities (Hoteit et al. 2005). EKF is 344 limited by the computational expense of integrating the tangent linear model forwards to derive the 345 error covariance matrix, and by the potential for unbounded error growth resulting from neglecting 346 higher order terms in the closure scheme (Zhang and Pu 2010). 347 Other Predictor-Corrector schemes include that developed by Preis et al (2011) who applied a 348 GA to update Demand Multiplication Factors for a skeletonised network in real-time, with an 349 objective function that accounted for measurement noise. Davidson and Bouchart (2006) developed a 350 weighted least squares method to adjust nodal demand whilst satisfying SCADA constraints. 351 However, in the under-determined problem considered, a range of solutions would be feasible. Kang 352 and Lansey (2009) applied a Tracking State Estimator (TSE) iteratively to obtain demand estimates in 353 real-time for a WDS. The TSE approach was better suited to looped WDS areas than the KF. 354 355 Ensemble Based Approaches 356 Ensemble assimilation procedures have been developed to overcome some of the problems 357 involved in providing a correction to a deterministic forecast. In a generic ensemble approach the 358 Μ π‘−1 |πΜπ‘−1 ), becomes the prior (or forecast density): posterior density of model states at time t-1, π(π 359 Μ π‘ |π Μ π‘−1 , πΜπ‘−1 ). The density is then combined with a likelihood function of the newly available π(π 360 observations to derive the posterior density of model states represented by Bayes’ equation: [Type text] 361 362 Μ π‘ |πΜπ‘ ) ∝ π(πΜπ‘ |π Μ π‘ )π(π Μ π‘ |π Μ π‘−1 , πΜπ‘−1 ) π(π (7) 363 364 The Ensemble Kalman Filter (EnKF) was developed to overcome some of the problems 365 with EKF by propagating an ensemble (n) of model states derived from Monte Carlo perturbations of 366 the input states. Equations (4) and (6) are applied to propagate each ensemble member forwards in 367 time, and the model error covariance calculated using the ensemble mean, which avoids the 368 computational costs associated with propagating the error covariance matrix (Burgers et al. 1998). To 369 maintain ensemble variance noise is added to the observations during state update (Burgers et al. 370 1998; Evensen 2003). The EnKF has been applied most widely in the disciplines of climatology, 371 meteorology (Hargreaves et al. 2004), and more recently to hydrological systems (Xie and Zhang 372 2010), and has been shown to work well for nonlinear problems (Evensen 2003; Pu and Hacker 373 2009). 374 To avoid perturbing initial system states, which may increase sampling error, an alternative 375 set of schemes apply square root filter methods (Tippett et al. 2003), which have been shown to 376 outperform the EnKF when applied to a rainfall-runoff model (Clark et al. 2008). Further 377 methodological advances to the EnKF (see Evensen (2003) and Zhang and Pu (2010) for reviews) 378 include the Retrospective EnKF where updates are applied to t-n time steps using the current 379 observation (Pauwels and De Lannoy 2006). Komma et al. (2008) avoided calculating the Jacobian 380 between states and observations by propagating multiple realisations of each ensemble member, and 381 retaining the ensemble member closest to the observations. 382 Sequential Monte Carlo Sampling, also known as the Particle Filter (PF) applied a recursive 383 Bayesian Filter for state update (Arulampalam et al. 2002). The posterior density in Equation 7 is 384 represented by a number (np) of particles (e.g. independent models) each with an associated 385 weight (π€π‘π ). In the sequential case as described in Equation 7, the prior PDF can be set 386 Μ π‘ |π Μ π‘−1 ), and is obtained using the system model. As sampling from the posterior density is as π(π [Type text] 387 generally impossible, the transition prior is typically used as the proposal density. Models are 388 propagated forwards, and each reweighted as a function of the likelihood: 389 390 π€π‘π = Μ ππ‘ ) π(πΜπ‘−1 |π π Μ ππ‘ ) ∑ π π(πΜπ‘−1 |π (8) π=1 391 392 The chosen likelihood function, which accounts for measurement error, is often assumed Gaussian 393 due to a lack of information about errors in the observations (Moradkhani et al. 2005a), but 394 heteroscedastic likelihoods have also been used (Salamon and Feyen 2010). Work is required to 395 evaluate whether other approaches might be applicable to determine model weights, including 396 informal likelihood functions, and methods employing possibility theory and/or evidence theory. The 397 more particles are included the closer the PF approximates the true posterior. The PF has been applied 398 in climatology, meteorology and hydrological modelling (Pham 2001; Salamon and Feyen 2009; van 399 Leeuwen 2009; Vossepoel and van Leeuwen 2007; Weerts and El Serafy 2006) and can outperform 400 EnKF, but at increased computational cost (Pham 2001). Unlike EnKF the state-space model need not 401 be linear or Gaussian and no modifications to system states are made, preserving mass and 402 momentum in the model equations. However, not updating model states can lead to filter degeneracy, 403 where particles (models) evolve away from the observations over time, leading to poor posterior 404 representation (van Leeuwen 2009). 405 Methods for dealing with filter degeneracy include: re-sampling particles with higher weights 406 at the expense of poorer performing particles; the guided particle filter, which confronts particles with 407 observations prior to the measurement time; and the backtracking filter, which re-samples back in 408 time with a larger ensemble (see review by van Leeuwen (2009)). An alternative to resampling is to 409 modify the proposal density by using an EnKF to bring the particles closer to future observations (van 410 Leeuwen 2010). 411 412 Joint State and Parameter Estimation [Type text] 413 Data assimilation generally assumes an optimal parameter set is known prior to model 414 simulation (e.g. Shang et al. 2006). However, prior off-line calibration may not be representative of 415 real-time conditions (Preis et al. 2011), and not consider prior parameter uncertainty. Ensemble based 416 approaches for joint state and parameter estimation have been developed implementing both the EnKF 417 and PF (Moradkhani et al. 2005a; Moradkhani et al. 2005b; Salamon and Feyen 2010; Smith et al. 418 Μ π‘ is expanded to also represent the model parameters, which are 2008b). The vector of model states π 419 therefore considered time-varying. As changes to a parameter may not be observed until a number of 420 time-steps ahead, observations may be assimilated after t+n observations (Salamon and Feyen 2009), 421 and then the model re-propagated forwards again from t+1 to t+n+1 (Noh et al. 2011). Model states 422 and parameters may also be updated at different intervals. 423 In WDS, state (demand) and parameter (roughness) estimation is typically separate, however, 424 the availability of data means estimates of either are undermined by the uncertainty in the other. Brdys 425 and Chen (1994) applied a recursive branch and bound method to jointly infer demand and roughness, 426 accounting for measurement uncertainty. Kang and Lansey (2011) showed improved performance 427 over joint inference of demand and roughness by updating them separately using different observation 428 weights for each in a two-stage Weighted Least Squares (WLS) approach. 429 An issue for all real-time DA methods is the computational time required to propagate an 430 ensemble of models, or run an optimisation of model states in real-time. Ensemble sizes reported in 431 the literature vary depending on the specific application (Madsen and Skotner 2005; Weerts and el 432 Serafy 2005), and scale exponentially with problem size (Snyder et al. 2008; van Leeuwen 2009). 433 Thus Joint State and Parameter estimation will be more expensive computationally, but may be 434 necessary to correct the parameter estimates to the latest conditions. 435 436 Model Forecasting and Decision-Support 437 438 General [Type text] 439 Calibration and Data Assimilation should quantify and reduce uncertainties in the vector of 440 Μ) and initial system states (π Μ π¨ ) in the chosen (or optimised) model structure (π Μ ). model parameters (π 441 The forecasting time horizon in a specific WDS is determined by the time required to initiate control 442 actions to mitigate the impacts of a forecasted event. Unlike rainfall-runoff systems, the lag time 443 Μ and changes in system state is short, and therefore a demand forecast is required between demand (π) 444 to drive a real-time control model (Rao and Salomons 2007). 445 Short-term water demand fluctuations (e.g. hourly and daily timescales) are a complex 446 function of climatic, social, economic and cultural drivers (Arbues et al. 2003). To overcome the 447 difficulty (if not impossibility) of explicitly modelling human water demand, a range of data-driven 448 models have been applied (see Herrera et al. (2010) for a comparison of methods). The relative 449 performance of different models will depend on the specific WDS, however, regular model updating 450 (calibration) as new information becomes available is likely to improve the performance of all models 451 (Herrera et al. 2010). Quantifying uncertainty in demand forecasting has received relatively little 452 attention (Cutore et al. 2008; Zhang et al. 2007), in particular uncertainty in downscaling demand 453 predictions to individual nodes (Kang and Lansey 2009). The techniques considered in section 2 and 3 454 for calibration and data assimilation are also applicable for dealing with demand model uncertainty 455 (e.g. Cutore et al. 2008). 456 457 Combining sources of uncertainty in model forecasting 458 Applying a model in real-time requires estimates of both the values and uncertainties in all of 459 the terms present in the right hand side of the equation (1). A thorough method for combining and 460 propagating these terms into a system forecast has only been considered in different modelling 461 contexts (e.g. flood prediction), and is here introduced specifically for WDS: 462 1. Calibrate the WDS model(s) (e.g. different levels of skeletonisation) on a vector of past 463 Μ|π Μ,π Μ, π Μπ¨, π Μ ), considering the observations to derive the posterior parameter distribution, π(π 464 key assumptions in the chosen statistical method for quantifying uncertainty. [Type text] 465 466 2. Sample the posterior ππ times to derive an ensemble of models, each with a state vector that can include: tank levels, pump settings, nodal demands and parameters derived from stage 1. 467 3. Propagate each model forwards in time using a perturbed demand distribution for each 468 ensemble member, based on measured demand (e.g. at the DMA level), and a suitable down 469 scaling technique. 470 471 4. Assimilate system observations (e.g. nodal head or pipe flow) once available, to derive an Μ π‘ |πΜπ‘ ) using an ensemble assimilation approach. updated posterior of model states, π(π 472 5. Use updated demand alongside ancillary data to derive a demand forecast model and forecast 473 Μ ) for the time horizon/control horizon for an ensemble of size ππ of future water demands (π 474 the specific WDS. 475 Μ π‘ |πΜπ‘ ), (ππ ) becomes the prior, π(π Μ π‘ |π Μ π‘−1 , πΜπ‘−1 ), which is 6. The ensemble representation of π(π 476 propagated forward for each member of the demand forecast ensemble (ππ ) over the time 477 horizon, to give a forecasting ensemble of ππ β ππ . 478 479 7. Use the posterior spread to infer the statistical likelihood of the future event to inform a rulesbased decision, or repeat stage 6 for different potential control options. 480 8. Update system actuator settings based on the simulations conducted in Step 7. 481 9. At the next time-step for which observations are available repeat steps 2 through 6. 482 483 The stages represented above, whereby and ensemble is used to sample the specified (typically 484 probabilistic) uncertainty distributions to cascade different forms of uncertainty through to a model 485 forecast, represents a full, robust and desirable treatment of model uncertainty, provided adequate 486 methods have been applied to derive the necessary information on uncertainty. 487 The ensemble approach is however, currently impractical in many, if not all model 488 forecasting situations. The impracticality is primarily a result of computational limitations, however 489 advances in the application of multi-core CPU (Lopez-Ibanez et al. 2008) and in particular GPU 490 processing for hydraulic solvers and optimisation problems (Guidolin et al. 2011; Harding and 491 Banzhaf 2007) look to improve the potential for representing the full ensemble in real-time. The [Type text] 492 approach may also be limited in some cases by the inability to constrain and quantify all sources of 493 uncertainty, particularly within the probabilistic framework, within which most reviewed (and 494 applied) methods have been developed 495 forecasting to make robust control decisions for WDS still remains. In light of the above limitations, 496 the remainder of section 6 will consider alternative methods to quantify and reduce model uncertainty 497 in model forecasting (see Table 3). However, the need to consider uncertainty in model 498 499 Reduced and Data-driven models 500 Reduced (skeletonised) models are widely applied in WDS to overcome the 501 computational impracticality of running an all pipes model, and should be compared to full models to 502 evaluate the validity of the simulation during calibration (Preis et al. 2011). However, such validation 503 may not hold for future system conditions. Another approach to reduce computational burden is to 504 derive a linearisation of the hydraulic solver (Xu 2003).An alternative approach to model reduction is 505 to abandon a physical basis for the model, and apply data-driven models (also termed meta-models) to 506 the control process (Broad et al. 2010). The computational burden is moved off-line when such 507 models are calibrated to the physical model. Rao and Salomons (2007) and Broad et al. (2010) both 508 trained ANNs to offline EPANET simulations. GAs were used with ANN simulations to optimise 509 system performance and reduce operational cost (Martinez et al. 2007; Salomons et al. 2007). 510 Calibration methods presented in Section 4 should be used to account for uncertainty in the initial 511 system model and the meta-model approximation, which if done so correctly, may conversely 512 increase computational time. 513 514 Approximate Forecasting 515 A smaller ensemble forecast may be applied to provide an approximate representation of the 516 key uncertainties affecting model predictions (Zappa et al. 2011). Pappenberger et al (2005) applied a 517 clustering method to the posterior parameter distribution to derive a set of 6 representative parameter 518 sets in a rainfall runoff model, one for each cluster, which was then applied with 52 rainfall forecasts [Type text] 519 and used to drive 10 flood inundations models. Sensitivity analysis, such as Sobol’s method (Saltelli 520 et al. (2006)), may be applied to identify representative parameter sets for further propagation. 521 An alternative approach is to approximate the posterior distribution with uncertainty bounds 522 obtained off-line prior to model forecasting. Shrestha et al. (2009) trained an ANN to reproduce the 523 upper and lower prediction intervals (90%) derived from GLUE calibration, which was then ran 524 alongside the deterministic model to produce upper and lower forecast prediction bounds. Hostache et 525 al. (2011) applied a similar approach to a flood forecasting model, whereby a forecast was made over 526 a calibration period and used to develop a bivariate meta-gaussian model, which was used as an 527 integrated measure of all errors. Such methods can only be as good as the original specification of 528 uncertainty bounds. The conditions under which the original uncertainty/predictive bounds were 529 derived needs to be representative of the forecasting conditions, which includes the effect of 530 uncertainties in forecasted input drivers on model uncertainty. Approximate forecasting methods 531 require further evaluation in the context of WDS model forecasting. 532 533 Error correction 534 The accuracy of a system forecast – or forecast skill - is determined by the initial conditions 535 provided when the last observations were assimilated, which will eventually become ”washed out” 536 (Madsen and Skotner 2005). Data-driven models, such as ANN models (Shamseldin and O'Connor 537 2001), have been calibrated to residual error time-series (from calibration of a process-based model) 538 and used to calculate innovations, which are then added to the deterministic model forecast. Abebe 539 and Price (2003) improved 1-6 hour rainfall-runoff model forecasts by adding an ANN model 540 prediction, which was calibrated to past model residuals. Other error correction methods include 541 autoregressive time series models (Hostache et al. 2011; Lekkas et al. 2001) and genetic programming 542 (Khu et al. 2001). Though comparative studies have been conducted between different methods (e.g. 543 Goswami et al. 2005), such methods remain to be compared in the WDS context, and must be 544 compared using a forecasted as opposed to “measured” demand. 545 Madsen and Skotner (2005) extended error correction procedures to also update model states; 546 pre-determined gain functions were used to update state variables using the innovation determined [Type text] 547 off-line at measurement locations. Cañizares et al. (2001) applied a constant KF gain matrix for on- 548 line DA, derived from off-line EnKF simulations. Mancarella et al. (2008) employed a local model to 549 estimate forecast error at measured locations in the computational domain, and correlations between 550 model states to distribute these error corrections over the model domain. State error-correction 551 represents a computationally efficient method for state updating as innovations and the state 552 covariance matrix may be calculated off-line. However, the methods need to be evaluated in the 553 context of WDS where actuator settings may affect the state covariance matrix. 554 555 Incorporating uncertain information into WDS Real-time Control (RTC) 556 Regardless of the method(s) used for quantifying model and predictive (forecast) uncertainty, 557 incorporating such uncertain information into the real-time control processes faces a number of 558 challenges, including educating system controllers on the presentation and interpretation of new 559 information (Frick and Hegg 2011), However, the key challenge to address is how to collapse 560 uncertain information (e.g. an ensemble forecast) into a single control decision. Whether real-time 561 control takes the form of a rule-based decision in response to a system forecast, or an iterative 562 numerical model optimisation (Rao et al. 2007), each model run is typically deterministic. The 563 preceding discussion, however, clearly illustrates that this may not adequately represent system 564 uncertainty. 565 The simplest case for incorporating uncertain information is that an ensemble (Section 5.3) is 566 run for each control option considered, and the posterior uncertainty in the decision variable is 567 collapsed to the ensemble mean, or the location with highest probability (Hall and Solomatine 2008). 568 The ensemble mean can then be compared between control options to determine optimal system 569 control. Error-correction procedures may also be used to reduce uncertainty for each optimisation run. 570 However, collapsing uncertainty to a single metric may identify some optimal operating conditions, 571 but given uncertainty in this prediction, another area of the decision space might be more robust 572 against the model uncertainty. 573 An alternative approach has been considered in WDS design optimisation, where the 574 optimisation objective is set in terms of reducing risk and maximising system robustness (Kapelan et [Type text] 575 al. 2005; Sun et al. 2011). In such a procedure the risk and robustness objectives are set in 576 probabilistic terms, and for each fitness evaluation (when employing a Genetic Algorithm solver) an 577 ensemble is ran, reflecting the model uncertainty. Kapelan et al. (2005) accounted for demand 578 uncertainty in a WDS design problem, and showed that a small ensemble size (n=10-50) was adequate 579 to characterise the full ensemble Pareto front, with a larger ensemble size (n=1000). Such approaches 580 can be facilitated through application of high performance computing in optimisation problems 581 (Harding and Banzhaf 2007). 582 The probabilistic interpretation of an ensemble of model forecasts has been the subject of 583 debate in the climate research literature (Hall et al. 2007). As model structural uncertainties are not 584 readily measureable, probabilistic representation may only capture part of the full model uncertainty. 585 Fuzzy representations of uncertain variables (e.g. projected future water demand) and separation of 586 aleatory and epistemic uncertainties through evidence theory may provide more useful information to 587 the decision maker (Hall and Solomatine 2008). In addition, approaches for robust decision making 588 when information on system uncertainty is hard to define have been applied, including Info-Gap 589 methods and the Robust Decision Making (RDM) approach (Hall and Solomatine 2008; Hall et al. 590 2012).A number of additional challenges are present when dealing with model uncertainty in real- 591 time control. 592 ο· 593 594 uncertainty, meaning it may be difficult to choose the best control procedure. ο· 595 596 First, a number of control solutions may be produced that perform equally well given forecast Second, the time required to take appropriate actions (e.g. time horizon) may not be known a priori until modelling is undertaken to predict future conditions. ο· Third, propagating and reducing uncertainty in real-time requires an increased modelling time, 597 which will require a longer input forecast at the start of the modelling period. As forecast 598 accuracy declines with increased lead time, accounting for this input uncertainty may 599 conversely increase modelling uncertainty. 600 601 ο· Finally, the ability to have models that are valid (calibrated) for difficult to foresee future conditions, such as pipe burst, remains a challenge. [Type text] 602 603 Conclusions and Recommendations 604 Uncertainty in WDS model forecasting may originate from measurement error, parameter and 605 initial conditions uncertainty, and also from inherent model structural errors. This paper has presented 606 a framework considering each stage of model development, and reviewed a number of the most 607 promising methods available to quantify and reduce uncertainty at each of these stages. An ideal 608 treatment of uncertainty should proceed through adequate sampling from the relevant posterior 609 distributions that represent uncertainty in parameters and system states, and propagate these samples 610 through the forecasting chain. However, in light of resource limitations, different approximations for 611 dealing with uncertainty may be required at each stage of model development. Figure 2 provides 612 guidance when choosing method(s) for dealing with uncertainty in calibration, data assimilation and 613 model forecasting. 614 It is important to consider, however, that there is no universal method applicable to all cases, and 615 nor is the decision process linear As in the context of flash flood forecasting, any approach for dealing 616 with the cascade of uncertainty will include a mixture of more formal and informal approaches (Cloke 617 and Pappenberger 2009). Thus, decisions regarding which method to adopt are based initially on a 618 conditional yes/no. Iterative development is required in light of new data, and also depending on the 619 extent to which the chosen method reliably informs the decision making process regarding the 620 different forms of uncertainty present in the forecasting procedure (Beven and Alcock 2012; Hall and 621 Solomatine 2008). 622 623 A number of work areas therefore need to be addressed to implement the methods for uncertainty quantification and reduction considered in this paper, specifically for WDS: 624 1. Methods for dealing with model uncertainty need to be evaluated specifically in WDS, with 625 explicit evaluation of computational speed versus model accuracy/uncertainty quantification. 626 For example, during calibration research is required to determine the optimal trade-off 627 between model structural uncertainty (e.g. skeletonisation) and the ability to constrain 628 distributed system demand; [Type text] 629 2. Whilst synthetic studies are seen as necessary steps to ensure internal consistency of 630 statistical methods (Renard et al. 2010), many techniques developed in the research literature 631 (e.g. Kang and Lansey 2009) need further evaluation against data from real WDS systems, 632 with real (and often unknown) error structures. Increasing availability of SCADA data should 633 facilitate this development; 634 635 3. Further work is required to quantify and improve the validity of calibrated models and methods for dealing with uncertainty during anomalous system conditions; 636 4. The added benefit of accounting for model uncertainty needs to be demonstrated specifically 637 for WDS, and also in the RTC context. Such a step is important to encourage water 638 companies to adopt more robust decision making approaches when using model based system 639 control; 640 641 642 643 In order to address these requirements, a number of additional, general guidelines are considered: 5. The final application of the model should guide development, combination and application of the techniques considered; 644 6. Any approach to uncertainty quantification should be justified a posteriori, using appropriate 645 tests of assumptions embodied in the likelihood function (e.g. Beven et al. 2008; Thyer et al. 646 2009), and tests against more rigorous approaches (Kang and Lansey 2009; Kapelan et al. 647 2007). Such tests should occur prior to propagation where it may be more difficult to 648 disentangle sources of uncertainty; 649 650 651 652 653 654 7. Any method for quantifying uncertainty offline must be applicable for the conditions to which the model is to be applied; 8. Global sensitivity analysis, considering higher order interactions between parameters should be applied to help constrain ensemble sizes; 9. Computational Resources should be applied to the most uncertain part of the model prediction; [Type text] 655 10.Finally, although the flow of information considered in Figure 2 is largely one way, 656 implementation and development is an iterative process between data and models; 657 information gained during model forecasting should be fed back to guide further data 658 collection and calibration to constrain model predictions and forecasts. 659 660 Acknowledgements 661 This paper has resulted from work conducted as part of 'PREPARED, Enabling Change', a European 662 Commission Seventh Framework project (Grant agreement no.: 244232, 2010-2014). Three 663 anonymous reviewers are also thanked for thorough and constructive comments. 664 6.6 Real-Time Control System Demand Offline system measurements Sensor 4. Calibration and Parameter Uncertainty System Process Initially Calibrated Model Actuator Data Processing Control Decision Optimisation State (and Parameter) Estimation 5. Data Assimilation Rule-Based Control Forecasted System Demand 6. Model Forecasting Control Schedule Forecast Simulation Model 665 666 Figure 1. Framework of Modelling in support of real-time WDS control. Numbers refer to specific 667 sections of the review. 668 669 670 671 [Type text] 672 673 674 675 676 677 678 679 680 681 [Type text] Calibration Linear model with normally distributed parameters? Y Optimisation Algorithm FOSM Errors sources identifiable? Explicit Error Model Y Informal Bayesian N N Formal Bayesian Errors sources identifiable? Informal Likelihood N Implicit Error Model Limits of Acceptability Y N Posterior Diagnostics Y Error Model Assumptions Valid? Data Assimilation Imprecise Probability Computationally expensive model for SE? Y Data/Process-Model Driven Meta-model Ensemble Approach N Model errors Gaussian? Known model/data errors with (mildly non-) linear model? Y Predictor-Corrector Scheme Y N N (Extended) Kalman Filter Ensemble Kalman Filter Iterative Optimisation Data/Process-Model Driven Meta-model Y Deterministic Forecast Error-Correction Y Deterministic Objective Functions N Approximate Ensemble Forecast Forecasted Uncertainty bounds Optimisation? Particle Filter Computationally expensive model for RTC optimisation? Model Forecasting N Evidence Theory Ensemble Forecast Optimisation? N Risk and Robustness Objective Functions Y Rules Based Control Rules Based Control N Control Decision Assumptions regarding model uncertainty acceptable? 682 683 Figure 2. Flow diagram illustrating the key decisions and assumptions that need to be considered 684 when dealing with the propagation of model uncertainty. 685 686 [Type text] 687 [Type text] 688 Table 1. Methods applied for quantification of uncertainty during model calibration Key References Sampling Method Likelihood Function Key assumptions, strengths and limitations GA; GN; GB; RMSE; (W)LS; (W)SSE. Efficient methods for reducing parameter uncertainty, but no quantification of model uncertainty. LHS; FOSM; Assumes linearity of model response, and Gaussianity of parameter uncertainty; computationally expensive when calculating derivatives with respect to calibration parameters. A more robust means to obtain posterior parameter and predictive distributions through calibration; implicit and explicit error models may be defined, with error model parameters to jointly infer during calibration; posterior parameters are sensitive to the chosen error model; error model assumptions should be evaluated using posterior diagnostic checks; explicit error models are limited by our ability to specify structural and output error a priori; see Beven et al. (2008), Vrugt et al.(2009b), Schoups and Vrugt (2010) and (Renard et al. 2010) for more detail on strengths and limitations. Optimisation (Savic et al. 2009) review paper FOSM (Kang and Lansey 2009; Lansey et al. 2001) Formal Bayesian Approaches (Kapelan et al. 2007) SCEM-UA SLS. (Freni and Mannina 2010) (Schaefli et al. 2007) (Yang et al. 2007) MCS M-H MCMC M-H MCMC (Schoups and Vrugt 2010) DREAM-ZS (Thyer et al. 2009); (Renard et al. 2010) MCMC ND after BCT of data. (M)ND with AR1 model ND after BCT with AR model. PD checks. SEPD with AR model and H Model. PD checks. Explicit ND with hyper parameters Informal Bayesian Approaches (Beven and Freer 2001); (Smith et al. 2008a); (McMillan and Clark 2009) (Krueger et al. 2010); (Liu et al. 2009) MCS; SCEM-UA NSE; NSSE; ESE MCS; Limits of Acceptability (LA) evaluates model performance at each time-step considering uncertainty in output data. A range of informal likelihood error models have been applied (Smith et al. 2008a), with flexibility depending on the specific problem; informal error models are less sensitive to overconditioning the posterior parameter distribution; parameter and predictive uncertainty is sensitive to subjective, user defined thresholds; LA approach also reveals general problems found with Explicit formal error models: input errors can be difficult to disentangle from model structural errors; See Beven et al. (2008), Vrugt et al.(2009b), Mantovan and Todini (2006) and Stedinger et al. (2008) for mode detail on strengths and limitations. Possibilistic, and Evidence Theory Approaches 689 (Branisavljevic et al. 2009; Revelli and Ridolfi 2002); NT/GA sampling Fuzzy membership of states and parameters (Fu and Kapelan 2011) LHS;GA Fuzzy membership (Fu and Kapelan, in press) MCS - Output distribution dependent on prior parameter and/or demand uncertainty. Not used for calibration but to represent input uncertainty Propagation of fuzzy probabilities in WDS design. Approach to separate out epistemic and aleatory uncertainty. Combines fuzzy representation of parameter uncertainty with a probabilistic representation of input (rainfall) uncertainty. GA, genetic algorithm; GN, Gauss-Newton technique; NT, Newton Technique; GB, Gradient-Based optimisation; FOSM, First-Order Second-Moment; SCEM-UA, Shuffled Complex Evolution Metropolis algorithm; RMSE, Root Mea Square Error; (W)LS, (Weighted) Least Squares; (W)SSE, (Weighted) Sum Squared Errors; LHS, Latin Hypercube Sampling; (M)ND, (Mixture) Normal Distribution; MCS, Monte Carlo Simulation; M-H, Metropolis-Hastings; MCMC, Markov Chain Monte Carlo; AR, Autoregressive Model; BCT, Box-Cox Transformation; PD, Posterior Diagnostics; DREAM-ZS, DiffeRential Evolution Adaptive Metropolis Algorithm; SEPD, Skewed Exponential Power Density; H, Heteroscedastic; (E)NSE, (Extended) Nash-Sutcliffe Efficiency; NSSE, Normalised Sum Square Errors; LA, Limits of Acceptability. 690 691 692 693 694 695 [Type text] 696 Table 2. Methods applied for Data Assimilation and Joint State and Parameter Estimation 697 Key Reference Model/Algorithm Uncertainty Key assumptions, strengths and limitations Quantification Parameter Key assumptions, strengths and limitations Uncertainty Key References State Update Predictor-Corrector Schemes (Todini 1999); (Shang et al. 2006) (Preis et al. 2011) KF EKF GA APP APP APP (Davidson and Bouchart 2006) (Kang and Lansey 2009) WLS APP TSE APP ND error model for demand observation errors; Model linearity required; Assumed Gaussianity of errors; Requires Linear adjoint model. GA to solve real-time optimisation; incorporated measurement error into calibration; may be expensive computationally in real-time. Assumes demand and roughness if correct; does not provide the upper and lower bounds of solutions that satisfy SCADA constraints. Iterative filter application may be expensive computationally, and can only occur once the data are available. Model linearity required for KF, which is sensitive to specification of observation errors; Testing required for real data. Ensemble Based Approaches (Clark et al. 2008); (Pu and Hacker 2009); (Evensen 2003) EnKF; EnSRF; EnAKF; EnTKF; (Weerts and El Serafy PF-RR; 2006); (van Leeuwen 2009; van Leeuwen 2010); APP - EnKF method is the most sensitive to ensemble size; EnKF perturbs observations prior to assimilation, which may increase sampling error; assimilation performance is sensitive to (specification of) observation errors; EnKF might be limited in strongly nonlinear systems; Kalman Filter based approaches modify system states which could lead to model instability. PF performed better than EnKF with larger ensemble sizes (32-128); PF does not make assumptions about model linearity; PF may require large ensemble sizes to overcome problems of filter degeneracy and sample impoverishment; EnKF found to be less sensitive to specification of model and measurement error; EnKF modification to the proposal density to draw PF to future observations, reduced required ensemble size. See van Leeuwen (2009) for methodological details. Joint State and Parameter Estimation (Moradkhani et al. 2005b) EnKF EnKF Separate EnKF update for States and Parameters; Inputs perturbed for state update, kernel smoother for parameter perturbation; Computational expensive for joint update in real-time; (Moradkhani et al. 2005a); (Salamon and Feyen 2009) PF(RR); PF(SIR). Perturbed at each time step (Brdys and Chen 1994) RBB - Posterior predictive distribution of model parameters and states at each time-step; Computationally expensive for joint update in real-time; Error model assumptions also require posterior evaluation (see Table 1: Formal Bayesian Approaches). Joint inference of demand and roughness. (Kang and Lansey 2011) Iterative two step WLS Iterative two step WLS Demand and roughness optimised using separate weighting schemes for pipe flow and nodal pressure in each iteration. ο· KF, Kalman Fitler; APP, A priori parameters; GA, Genetic Algorithm; WLS, Weighted Least Squares; TSE, Tracking State Estimator; EnKF, Ensemble Kalman Filter; EnSRF, Ensemble Square Root Filter; EnAKF, Ensemble Adjusted Kalman Filter; EnTKF, Ensmeble Transform Kalman Filter; PF, Particle Filter; RR, Residual Resampling; SIR, Sequential Importance Resampling; RBB; Recursive Branch and Bound. 698 699 700 701 702 703 704 [Type text] 705 Reduced and Data-Driven Models (Preis et al. 2011) RM, skeletonised EPANET2 model None (Rao and Salomons 2007); (Broad et al. 2010) GA-ANN for control optimisation None Ensemble and Approximate Forecasts (Pappenberger et al. 2005); (Zappa et al. 2011) IL for offline calibration; Ensemble Forecast. Approximate ensemble of input and parameter uncertainty (Shrestha et al. 2009) ANN; GLUE Offline calibration of error bounds applied to forecast (Hostache et al. 2011) BMG model for representing forecasting error; linear correction for model error Posterior Predictive Uncertainty in Forecast for integrated model errors ANN; AR; LTF; NARXM Error correction only Skeletonised EPANET2 models can simulate overall network performance; Model performance should be compared to706 full model performance to evaluate the effect of skeletonisation; Skeletonisation affects simulation of pressure surge, demand satisfaction predictions, and contaminant consequence 707 assessment. 708 ANN trained offline to hydraulic model scenarios; Data-driven metamodels can speed up computational time, making them a suitable alternative for real-time modelling, forecasting and 709 system performance optimisation; Data-driven models should be calibrated using methods to account for uncertainty, and 710 errors in original physically based model. Sensitivity Analysis of model parameters can help chose 711 representative samples from the posterior parameter 712 distribution; computational resources should be directed to represent the factor that introduces the most uncertainty into the model forecast; method requires comparison to more713 thorough ensemble forecasts to evaluate robustness. Reliant on the strength of the error assumptions made during 714 the initial calibration; data-driven models should be calibrated using methods to account for uncertainty, and errors in 715 original physically based model. Results should be compared to ensemble error bounds. Gaussianity assumption of errors does not always hold. 716 717 718 Error-Correction (Abebe and Price 2003); (Goswami et al. 2005) Errors made during calibration are the only errors corrected, 719 which may not include errors in the forecasted input conditions, an additional and potentially significant source of uncertainty. 720 Error in all model states reduced through state updating; the strength of the relationship between model errors at 721 observation points and other states in the system may be controlled in WDS by control structures (e.g pumps and 722 valves), which may limit application. (Madsen and Error Correction Uncertainty Skotner 2005); applied to all model reduction via DA (Canizares et al. states using gain 2001); functions determined (Mancarella et al. offline. 2008) 723 RM, Reduced Model; GA, Genetic Algorithm; ANN, Artificial Neural Network; GLUE, Generalised Likelihood Uncertainty Estimation; IL, Informal Likelihood Functions; BMG, Bivariate Meta Gaussian Model; AR, Autoregressive; LTF, Linear Transfer Function; NARXM, Non-linear, Auto-Regressive eXogenous input Model. 724 725 Table 3. Methods applied to quantify and reduce error in model forecasting 726 727 728 729 730 731 732 [Type text] 733 734 References 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 Abebe, A. J., and Price, R. K. (2003). "Managing uncertainty in hydrological models using complementary models." Hydrological Sciences Journal-Journal Des Sciences Hydrologiques, 48(5), 679-692. Arbues, F., Garica-Valinas, M. A., and Martinez-Espineira, R. (2003). "Estimation of residential water demand: a state-of-the-art review." Journal of Socio-Economics, 32, 81-102. Arulampalam, M. S., Maskell, S., Gordon, N., and Clapp, T. (2002). "A tutorial on particle filters for online nonlinear/non-Gaussian Bayesian tracking." Ieee Transactions on Signal Processing, 50(2), 174-188. Bahadur, R., Johnson, J., Janke, R., and Samuels, W. B. (2006). "Impact of Model Skeletonization on Water Distribution Model Parameters as Related to Water Quality and Contaminant Consequence Assessment." 8th Annual Water Distribution Systems Analysis Symposium, CIncinati, Ohio, USA. Bakker, K. J. (2003). "From public to private to ... mutual? - Restructuring water supply governance in England and Wales." Geoforum, 34(3), 359-374. Bargiela, A., and Hainsworth, G. D. (1989). "Pressure and Flow Uncertainty in Water-Systems." Journal of Water Resources Planning and Management-Asce, 115(2), 212-229. Bates, B. C., and Campbell, E. P. (2001). "A Markov chain Monte Carlo scheme for parameter estimation and inference in conceptual rainfall-runoff modeling." Water Resources Research, 37(4), 937-947. Bertrand-Krajewski, J. L., Bardin, J. P., Mourad, M., and Beranger, Y. (2003). "Accounting for sensor calibration, data validation, measurement and sampling uncertainties in monitoring urban drainage systems." Water Science and Technology, 47(2), 95-102. Beven, K. (2006). "A manifesto for the equifinality thesis." Journal of Hydrology, 320(1-2), 18-36. Beven, K., and Binley, A. (1992). "The Future of Distributed Models - Model Calibration and Uncertainty Prediction." Hydrological Processes, 6(3), 279-298. Beven, K., and Freer, J. (2001). "Equifinality, data assimilation, and uncertainty estimation in mechanistic modelling of complex environmental systems using the GLUE methodology." Journal of Hydrology, 249(14), 11-29. Beven, K. J., and Alcock, R. E. (2012). "Modelling everything everywhere: a new approach to decision-making for water management under uncertainty." Freshwater Biology, 57, 124-132. Beven, K. J., Smith, P. J., and Freer, J. E. (2008). "So just why would a modeller choose to be incoherent?" Journal of Hydrology, 354(1-4), 15-32. Bicik, J., Kapelan, Z., Makropoulos, C., and Savic, D. (2011a). "Pipe burst diagnostics using evidence theory." Journal of Hydroinformatics, 13(4), 596-608. Bicik, J., Kapelan, Z., and SaviΔ, D. (2011b). "Challenges in the implementation of a DSS for real-time WDS management." Computing and Control in the Water Industry 2011, University of Exeter. Boulos, P. F., Lansey, K. E., and Karney, B. W. (2004). Comprehensive Water Distribution Systems Analysis Handbook For Engineers and Planners, MWH Soft, Inc., Pasadena, California. Box, G. E. P., and Cox, D. R. (1982). "An Analysis of Transformations Revisited, Rebutted." Journal of the American Statistical Association, 77(377), 209-210. Branisavljevic, N., Kapelan, Z., and Prodanovic, D. (2010). "Online time data series pre-processing for the improved performance of anomaly detection methods." Integrating water systems: Proceedings of the Tenth International Conference on Computing and Control for the Water Industry, CCWI 2009 'Integrating Water Systems', Sheffield, UK. Branisavljevic, N., Prodanovic, D., and Ivetic, M. (2009). "Uncertainty reduction in water distribution network modelling using system inflow data." Urban Water Journal, 6(1), 69-79. Brdys, M. A., and Chen, K. (1994). "Joint State and Parameter Estimation of Dynmaic Water Supply System Under bounded Uncertainty Using Geometric Programming." 10th IFAC Symposium of System Identification, Copenhagen, Denmark, 331-336. Broad, D. R., Maier, H. R., and Dandy, G. C. (2010). "Optimal Operation of Complex Water Distribution Systems Using Metamodels." Journal of Water Resources Planning and Management-Asce, 136(4), 433-443. Burgers, G., van Leeuwen, P. J., and Evensen, G. (1998). "Analysis scheme in the ensemble Kalman filter." Monthly Weather Review, 126(6), 1719-1724. [Type text] 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 Bush, C. A., and Uber, J. G. (1998). "Sampling design methods for water distribution model calibration." Journal of Water Resources Planning and Management-Asce, 124(6), 334-344. Canizares, R., Madsen, H., Jensen, H. R., and Vested, H. J. (2001). "Developments in operational shelf sea modelling in Danish waters." Estuarine Coastal and Shelf Science, 53(4), 595-605. Clark, M. P., Rupp, D. E., Woods, R. A., Zheng, X., Ibbitt, R. P., Slater, A. G., Schmidt, J., and Uddstrom, M. J. (2008). "Hydrological data assimilation with the ensemble Kalman filter: Use of streamflow observations to update states in a distributed hydrological model." Advances in Water Resources, 31(10), 1309-1324. Cloke, H. L., and Pappenberger, F. (2009). "Ensemble flood forecasting: A review." Journal of Hydrology, 375(3-4), 613-626. Cutore, P., Campisano, A., Kapelan, Z., Modica, C., and Savic, D. (2008). "Probabilistic prediction of urban water consumption using the SCEM-UA algorithm." Urban Water Journal, 5(2), 125-132. Davidson, J. W., and Bouchart, F. J. C. (2006). "Adjusting nodal demands in SCADA constrained real-time water distribution network models." Journal of Hydraulic Engineering-Asce, 132(1), 102-110. Deletic, A., Dotto, C. B. S., McCarthy, D. T., Kleidorfer, M., Freni, G., Mannina, G., Uhl, M., Henrichs, M., Fletcher, T. D., Rauch, W., Bertrand-Krajewski, J. L., and Tait, S. (2011). "Assessing uncertainties in urban drainage models." Physical and Chemistry of the Earth, doi:10.1016/j.pce.2011.04.007. Draper, D. (1995). "Assessment and Propagation of Model Uncertainty." Journal of the Royal Statistical Society Series B-Methodological, 57(1), 45-97. Dubois, D. (2010). "Representation, Propagation, and Decision Issues in Risk Analysis Under Incomplete Probabilistic Information." Risk Analysis, 30(3), 361-368. Evensen, G. (2003). "The Ensemble Kalman Filter: theoretical formulation and practical implementation." Ocean Dynamics, 53, 343-367. Freni, G., and Mannina, G. (2010). "Bayesian approach for uncertainty quantification in water quality modelling: The influence of prior distribution." Journal of Hydrology, 392(1-2), 31-39. Frick, J., and Hegg, C. (2011). "Can end-users' flood management decision making be improved by information about forecast uncertainty?" Atmospheric Research, 100(2-3), 296-303. Fu, G., and Kapelan, Z. (in press). "Flood analysis of urban drainage systems: probabilistic dependence structure of rainfall characteristics and fuzzy model parameters." Journal of Hydroinformatics. Fu, G. T., and Kapelan, Z. (2011). "Fuzzy probabilistic design of water distribution networks." Water Resources Research, 47. Gallagher, M., and Doherty, J. (2007). "Parameter estimation and uncertainty analysis for a watershed model." Environmental Modelling & Software, 22(7), 1000-1020. Giustolisi, O. (2010). "Considering Actual Pipe Connections in Water Distribution Network Analysis." Journal of Hydraulic Engineering-Asce, 136(11), 889-900. Giustolisi, O., Kapelan, Z., and Savic, D. (2008a). "Extended Period Simulation Analysis Considering Valve Shutdowns." Journal of Water Resources Planning and Management-Asce, 134(6), 527-537. Giustolisi, O., Savic, D., and Kapelan, Z. (2008b). "Pressure-driven demand and leakage simulation for water distribution networks." Journal of Hydraulic Engineering-Asce, 134(5), 626-635. Goswami, M., O'Connor, K. M., Bhattarai, K. P., and Shamseldin, A. Y. (2005). "Assessing the performance of eight real-time updating models and procedures for the Brosna River." Hydrology and Earth System Sciences, 9(4), 394-411. Guidolin, M., Savic, D., and Kapelan, Z. (2011). "Computational performance analysis and improvment of the demand-driven hydraulic solver for the CWSNET library." Computing and Control in the Water Industry (CCWI), Exeter, UK. Hall, J., Fu, G. T., and Lawry, J. (2007). "Imprecise probabilities of climate change: aggregation of fuzzy scenarios and model uncertainties." Climatic Change, 81(3), 265-281. Hall, J., and Solomatine, D. (2008). "A framework for uncertainty analysis in flood risk management decisions." International Journal of River Basin Management, 6(2), 85-98. Hall, J. W. (2003). "Handling uncertainty in the hydroinformatic process." Journal of Hydroinformatics, 5, 215-232. Hall, J. W., Lempert, R. J., Keller, K., Hackbarth, A., Mijere, C., and McInerney, D. J. (2012). "Robust Climate Policies Under Uncertainty: A Comparison of Robust Decision Making and Info-Gap Methods." Risk Analysis, doi: 10.1111/j.1539-6924.2012.01802.x. Harding, S., and Banzhaf, W. (2007). "Fast genetic programming on GPUs." Genetic Programming, Proceedings, 4445, 90-101. [Type text] 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 Hargreaves, J. C., Annan, J. D., Edwards, N. R., and Marsh, R. (2004). "An efficient climate forecasting method using an intermediate complexity Earth System Model and the ensemble Kalman filter." Climate Dynamics, 23(7-8), 745-760. Hart, W. E., and Murray, R. (2010). "Review of Sensor Placement Strategies for Contamination Warning Systems in Drinking Water Distribution Systems." Journal of Water Resources Planning and Management-Asce, 136(6), 611-619. Helton, J. C., Johnson, J. D., and Oberkampf, W. L. (2004). "An exploration of alternative approaches to the representation of uncertainty in model predictions." Reliability Engineering & System Safety, 85(1-3), 3971. Herrera, M., Torgo, L., Izquierdo, J., and Perez-Garcia, R. (2010). "Predictive models for forecasting hourly urban water demand." Journal of Hydrology, 387(1-2), 141-150. Hostache, R., Matgen, P., Montanari, A., Montanari, M., Hoffmann, L., and Pfister, L. (2011). "Propagation of uncertainties in coupled hydro-meteorological forecasting systems: A stochastic approach for the assessment of the total predictive uncertainty." Atmospheric Research, 100(2-3), 263-274. Hoteit, I., Korres, G., and Triantafyllou, G. (2005). "Comparison of extended and ensemble based Kalman filters with low and high resolution primitive equation ocean models." Nonlinear Processes in Geophysics, 12(5), 755-765. Hutton, C. J., Vamakeridou-Lyroudia, L. S., Kapelan, Z., and Savic, D. (2011). "Uncertainty Quantifcation and Reduction in Urban Water Systems Modelling: Evaluation Report." PREPARED: Enabling Change, Work Package 3.6, European Commision 7th Framework Programme. Jacobsen, L. B., and Kamojjala. (2009). "Tools and processes for calibrating large all-pipes models." Urban Water Journal, 6(1), 29-38. Jamieson, D. G., Shamir, U., Martinez, F., and Franchini, M. (2007). "Conceptual design of a generic, real-time, near-optimal control system for water-distribution networks." Journal of Hydroinformatics, 9(1), 3-14. Jonkergouw, P. M. R., Khu, S. T., Kapelan, Z. S., and Savic, D. A. (2008). "Water quality model calibration under unknown demands." Journal of Water Resources Planning and Management-Asce, 134(4), 326-336. Jung, B. S., Boulos, P. F., and Wood, D. J. (2007). "Pitfalls of water distribution model skeletonization for surge analysis." Journal American Water Works Association, 99(12), 87-98. Kang, D., and Lansey, K. (2009). "Real-Time Demand Estimation and Confidence Limit Analysis for Water Distribution Systems." Journal of Hydraulic Engineering-Asce, 135(10), 825-837. Kang, D. S., and Lansey, K. (2011). "Demand and Roughness Estimation in Water Distribution Systems." Journal of Water Resources Planning and Management-Asce, 137(1), 20-30. Kang, D. S., Pasha, M. F. K., and Lansey, K. (2009). "Approximate methods for uncertainty analysis of water distribution systems." Urban Water Journal, 6(3), 233-249. Kapelan, Z. S., Savic, D. A., and Walters, G. A. (2005). "Multiobjective design of water distribution systems under uncertainty." Water Resources Research, 41(11). Kapelan, Z. S., Savic, D. A., and Walters, G. A. (2007). "Calibration of water distribution hydraulic models using a Bayesian-Type procedure." Journal of Hydraulic Engineering-Asce, 133(8), 927-936. Kavetski, D., Kuczera, G., and Franks, S. W. (2006a). "Bayesian analysis of input uncertainty in hydrological modeling: 1. Theory." Water Resources Research, 42(3). Kavetski, D., Kuczera, G., and Franks, S. W. (2006b). "Bayesian analysis of input uncertainty in hydrological modeling: 2. Application." Water Resources Research, 42(3). Khu, S. T., Liong, S. Y., Babovic, V., Madsen, H., and Muttil, N. (2001). "Genetic programming and its application in real-time runoff forecasting." Journal of the American Water Resources Association, 37(2), 439-451. Kleiner, Y., and Rajani, B. (2001). "Comprehensive review of structural deterioration of water mains: statistical models." Urban Water 3, 131-150. Komma, J., Bloschl, G., and Reszler, C. (2008). "Soil moisture updating by Ensemble Kalman Filtering in real-time flood forecasting." Journal of Hydrology, 357(3-4), 228-242. Krueger, T., Freer, J., Quinton, J. N., Macleod, C. J. A., Bilotta, G. S., Brazier, R. E., Butler, P., and Haygarth, P. M. (2010). "Ensemble evaluation of hydrological model hypotheses." Water Resources Research, 46. Lane, S. N. (2005). "Roughness - time for a re-evaluation?" Earth Surface Processes and Landforms, 30(2), 251253. Lansey, K. E., El-Shorbagy, W., Ahmed, I., Araujo, J., and Haan, C. T. (2001). "Calibration assessment and data collection for water distribution networks." Journal of Hydraulic Engineering-Asce, 127(4), 270-279. [Type text] 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 Lekkas, D. F., Imrie, C. E., and Lees, M. J. (2001). "Improved non-linear transfer function and nerual network methods of flow routing for real-time forecasting." Journal of Hydroinformatics, 3, 153-164. Liu, Y. L., Freer, J., Beven, K., and Matgen, P. (2009). "Towards a limits of acceptability approach to the calibration of hydrological models: Extending observation error." Journal of Hydrology, 367(1-2), 93-103. Liu, Y. Q., and Gupta, H. V. (2007). "Uncertainty in hydrologic modeling: Toward an integrated data assimilation framework." Water Resources Research, 43(7), -. Lopez-Ibanez, M., Prasad, D. T., and Paechter, B. (2008). "Parallel Optimisation of pump schedules with a threadsafe variant of epanet toolkit." Proceedings of the 10th Annual Water Distribution Systems Analysis Conference WDSA2008, Kruger National Park, South Africa. Machell, J., Mounce, S. R., and Boxall, J. B. (2010). "Online modelling of water distribution systems: a UK case study." Drinking Water Engineering and Science, 3, 21-27. Madsen, H., and Skotner, C. (2005). "Adaptive state updating in real-time river flow forecasting - a combined filtering and error forecasting procedure." Journal of Hydrology, 308(1-4), 302-312. Mancarella, D., Babovic, V., Keijzer, M., and Simeone, V. (2008). "Data assimilation of forecasted errors in hydrodynamic models using inter-model correlations." International Journal For Numerical Methods in Fluids, 56, 597-605. Mantovan, P., and Todini, E. (2006). "Hydrological forecasting uncertainty assessment: Incoherence of the GLUE methodology." Journal of Hydrology, 330(1-2), 368-381. Martinez, F., Hernandez, V., Alonso, J. M., Rao, Z. F., and Alvisi, S. (2007). "Optimizing the operation of the Valencia water-distribution network." Journal of Hydroinformatics, 9(1), 65-78. Matgen, P., Montanari, M., Hostache, R., Pfister, L., Hoffmann, L., Plaza, D., Pauwels, V. R. N., De Lannoy, G. J. M., De Keyser, R., and Savenije, H. H. G. (2010). "Towards the sequential assimilation of SAR-derived water stages into hydraulic models using the Particle Filter: proof of concept." Hydrology and Earth System Sciences, 14(9), 1773-1785. McMillan, H., and Clark, M. (2009). "Rainfall-runoff model calibration using informal likelihood measures within a Markov chain Monte Carlo sampling scheme." Water Resources Research, 45. Menaia, J., Coelho, S. T., Lopes, A., Fonte, E., and Palma, J. (2003). "Dependency of bulk chlorine decay rates on flow.velocity in water distribution networks." 3rd World Water Congress: Water Services Management, Operations and Monitoring, 3(1-2), 209-214. Merz, B., and Thieken, A. H. (2005). "Separating natural and epistemic uncertainty in flood frequency analysis." Journal of Hydrology, 309(1-4), 114-132. Misiunas, D., Lambert, M., Simpson, A., and Olsson, G. (2005). "Burst detection and location in water distribution networks." Efficient Use and Management of Urban Water Supply (Efficient 2005), 5(3-4), 71-80. Moradkhani, H., Hsu, K. L., Gupta, H., and Sorooshian, S. (2005a). "Uncertainty assessment of hydrologic model states and parameters: Sequential data assimilation using the particle filter." Water Resources Research, 41(5), -. Moradkhani, H., Sorooshian, S., Gupta, H. V., and Houser, P. R. (2005b). "Dual state-parameter estimation of hydrological models using ensemble Kalman filter." Advances in Water Resources, 28(2), 135-147. Muste, M., Lee, K., and Bertrand-Krajewski, J. L. (2012). "Standardized uncertainty analysis for hydrometry: a review of relevant approaches and implementation examples." Hydrological Sciences Journal-Journal Des Sciences Hydrologiques, 57(4), 643-667. Nasseri, M., Moeini, A., and Tabesh, M. (2010). "Forecasting monthly urban water demand using Extended Kalman Filter and Genetic Programming." Expert systems with Applications, Article in Press, corrected proof. Noh, S. J., Tachikawa, Y., Shiiba, M., and Kim, S. (2011). "Applying sequential Monte Carlo methods into a distributed hydrologic model: lagged particle filtering approach with regularization." Hydrology and Earth System Sciences, 8, 3383-3420. Ofwat. (2005). "Water Framework Directive - Economic Analysis of Water Industry Costs. Final Report." Ofwat, UK. Ove Arup and Partners Ltd, In association with Oxera Consulting Ltd. Ogden, S., and Watson, R. (1999). "Corporate performance and stakeholder management: Balancing shareholder and customer interests in the UK privatized water industry." Academy of Management Journal, 42(5), 526-538. Ozger, S., and Mays, L. W. (2004). "Optimal Location of Isolation Valves: A Reliability Approach." Water Supply Systems Security, L. W. Mays, ed., McGraw-Hill, New York. [Type text] 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 Pappenberger, F., Beven, K. J., Hunter, N. M., Bates, P. D., Gouweleeuw, B. T., Thielen, J., and de Roo, A. P. J. (2005). "Cascading model uncertainty from medium range weather forecasts (10 days) through a rainfallrunoff model to flood inundation predictions within the European Flood Forecasting System (EFFS)." Hydrology and Earth System Sciences, 9(4), 381-393. Pauwels, V. R. N., and De Lannoy, G. J. M. (2006). "Improvement of modeled soil wetness conditions and turbulent fluxes through the assimilation of observed discharge." Journal of Hydrometeorology, 7(3), 458477. Pham, D. T. (2001). "Stochastic methods for sequential data assimilation in strongly nonlinear systems." Monthly Weather Review, 129(5), 1194-1207. Preis, A., Whittle, A. J., Ostfeld, A., and Perelman, L. (2010). "On-line hydraulic state estimation in urban water networks using reduced models." Integrating Water Systems, 319-324. Preis, A., Whittle, A. J., Ostfeld, A., and Perelman, L. (2011). "Efficient Hydraulic State Estimation Technique Using Reduced Models of Urban Water Networks." Journal of Water Resources Planning and ManagementAsce, 137(4), 343-351. Pu, Z. X., and Hacker, J. (2009). "Ensemble-based Kalman filters in strongly nonlinear dynamics." Advances in Atmospheric Sciences, 26(3), 373-380. Rao, Z. F., and Salomons, E. (2007). "Development of a real-time, near-optimal control process for waterdistribution networks." Journal of Hydroinformatics, 9(1), 25-37. Rao, Z. F., Wicks, J., and West, S. (2007). "Optimising water supply and distribution operations." Proceedings of the Institution of Civil Engineers-Water Management, 160(2), 95-101. Reichle, R. H., Walker, J. P., Koster, R. D., and Houser, P. R. (2002). "Extended versus ensemble Kalman filtering for land data assimilation." Journal of Hydrometeorology, 3(6), 728-740. Renard, B., Kavetski, D., Kuczera, G., Thyer, M., and Franks, S. W. (2010). "Understanding predictive uncertainty in hydrologic modeling: The challenge of identifying input and structural errors." Water Resources Research, 46, -. Revelli, R., and Ridolfi, L. (2002). "Fuzzy approach for analysis of pipe networks." Journal of Hydraulic EngineeringAsce, 128(1), 93-101. Rossman, L. A. (2000). EPANET2 Users Manual National Risk Management Research Laboratory, U.S. Environmental Protection Agency, http://www.image.unipd.it/salandin/IngAmbientale/Progetto_2/EPANET/EN2manual.pdf, Cincinnati, OH 45268, USA. Ruggaber, T. P., Talley, J. W., and Montestruque, L. A. (2007). "Using Embedded Sensor Networks to Monitoring, Control, and Reduce CSO Events: A Pilot Study." Environmental Engineering Science, 24(2), 172-182. Sadiq, R., Kleiner, Y., and Rajani, B. (2006). "Estimating risk of contaminant intrusion in water distribution networks using Dempster-Shafer theory of evidence." Civil Engineering and Environmental Systems, 23(3), 129-141. Salamon, P., and Feyen, L. (2009). "Assessing parameter, precipitation, and predictive uncertainty in a distributed hydrological model using sequential data assimilation with the particle filter." Journal of Hydrology, 376(3-4), 428-442. Salamon, P., and Feyen, L. (2010). "Disentangling uncertainties in distributed hydrological modeling using multiplicative error models and sequential data assimilation." Water Resources Research, 46. Salomons, E., Goryashko, A., Shamir, U., Rao, Z. F., and Alvisi, S. (2007). "Optimizing the operation of the Haifa-A water-distribution network." Journal of Hydroinformatics, 9(1), 51-64. Saltelli, A., Ratto, M., Tarantola, S., Campolongo, F., Commission, E., and Ispra, J. R. C. (2006). "Sensitivity analysis practices: Strategies for model-based inference." Reliability Engineering & System Safety, 91(10-11), 1109-1125. Savic, D. A., Kapelan, Z. S., and Jonkergouw, P. M. R. (2009). "Quo vadis water distribution model calibration?" Urban Water Journal, 6(1), 3-22. Schaefli, B., Talamba, D. B., and Musy, A. (2007). "Quantifying hydrological modeling errors through a mixture of normal distributions." Journal of Hydrology, 332(3-4), 303-315. Schilperoort, R. P. S., Dirksen, J., and Clemens, F. H. L. R. (2008). "Practical aspects for long-term monitoring campaigns: pifals to avoid to maximise data yield." 11th International Conference on Urban Drainage, Edinburgh, Scotland, UK. [Type text] 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 Schoups, G., and Vrugt, J. A. (2010). "A formal likelihood function for parameter and predictive inference of hydrologic models with correlated, heteroscedastic, and non-Gaussian errors." Water Resources Research, 46, -. Shamseldin, A. Y., and O'Connor, K. M. (2001). "A non-linear neural network technique for updating of river flow forecasts." Hydrology and Earth System Sciences, 5(4), 577-597. Shang, F., Uber, J. G., van Bloemen Waanders, B. G., Boccelli, D., and Janke, R. (2006). "Real Time Water Demand Estimation in Water Distribution System." 8th Annial water Distribution Systems Analysis Symposium, Cincinnati, Ohio, USA. Shrestha, D. L., Kayastha, N., and Solomatine, D. P. (2009). "A novel approach to parameter uncertainty analysis of hydrological models using neural networks." Hydrology and Earth System Sciences, 13(7), 1235-1248. Smith, P., Beven, K. J., and Tawn, J. A. (2008a). "Informal likelihood measures in model assessment: Theoretic development and investigation." Advances in Water Resources, 31(8), 1087-1100. Smith, P. J., Beven, K. J., and Tawn, J. A. (2008b). "Detection of structural inadequacy in process-based hydrological models: A particle-filtering approach." Water Resources Research, 44(1), -. Snyder, C., Bengtsson, T., Bickel, P., and Anderson, J. (2008). "Obstacles to High-Dimensional Particle Filtering." Monthly Weather Review, 136(12), 4629-4640. Sorooshian, S., and Dracup, J. A. (1980). "Stochastic Parameter-Estimation Procedures for Hydrologic RainfallRunoff Models - Correlated and Heteroscedastic Error Cases." Water Resources Research, 16(2), 430-442. Stedinger, J. R., Vogel, R. M., Lee, S. U., and Batchelder, R. (2008). "Appraisal of the generalized likelihood uncertainty estimation (GLUE) method." Water Resources Research, 44, -. Storey, M. V., van der Gaag, B., and Burns, B. P. (2011). "Advances in on-line drinking water quality monitoring and early warning systems." Water Research, 45(2), 741-747. Sun, S. (2010). "Decision-Making under Uncertainty: Optimal Storm Sewer Network Design Considering Flood Risk," Unpublished PhD, University of Exeter. Sun, S., Khu, S. T., Kapelan, Z., and Djordjevic, S. (2011). "A fast approach for multiobjective design of water distribution networks under demand uncertainty." Journal of Hydroinformatics, 13(2), 143-152. Thyer, M., Renard, B., Kavetski, D., Kuczera, G., Franks, S. W., and Srikanthan, S. (2009). "Critical evaluation of parameter consistency and predictive uncertainty in hydrological modeling: A case study using Bayesian total error analysis." Water Resources Research, 45, -. Tippett, M. K., Anderson, J. L., Bishop, C. H., Hamill, T. M., and Whitaker, J. S. (2003). "Ensemble square root filters." Monthly Weather Review, 131(7), 1485-1490. Todini, E. (1999). "Using a kalman filter approach for looped water distribution network calibration." Water industry systems: modelling and optimization applications, Volume 1., D. A. Savic and G. A. Walters, eds., Research Studies Press, Baldock, Hertforshire, England, 327-336. van Leeuwen, P. J. (2009). "Particle Filtering in Geophysical Systems." Monthly Weather Review, 137(12), 40894114. van Leeuwen, P. J. (2010). "Nonlinear data assimilation in geosciences: an extremely efficient particle filter." Quarterly Journal of the Royal Meteorological Society, 136(653), 1991-1999. Vossepoel, F. C., and van Leeuwen, P. J. (2007). "Parameter estimation using a particle method: Inferring mixing coefficients from sea level observations." Monthly Weather Review, 135(3), 1006-1020. Vrugt, J. A., Gupta, H. V., Bouten, W., and Sorooshian, S. (2003). "A Shuffled Complex Evolution Metropolis algorithm for optimization and uncertainty assessment of hydrologic model parameters." Water Resources Research, 39(8), -. Vrugt, J. A., ter Braak, C. J. F., Diks, C. G. H., Robinson, B. A., Hyman, J. M., and Higdon, D. (2009a). "Accelerating Markov Chain Monte Carlo Simulation by Differential Evolution with Self-Adaptive Randomized Subspace Sampling." International Journal of Nonlinear Sciences and Numerical Simulation, 10(3), 273-290. Vrugt, J. A., ter Braak, C. J. F., Gupta, H. V., and Robinson, B. A. (2009b). "Equifinality of formal (DREAM) and informal (GLUE) Bayesian approaches in hydrologic modeling?" Stochastic Environmental Research and Risk Assessment, 23(7), 1011-1026. Walski, T., Chase, D. V., Savic, D. A., Grayman, W. M., Beckwith, S., and Koelle, E. (2003). Advanced Water Distribution Modeling and Management, Haestead Methods Press, Waterbury, Connecticut. Weerts, A. H., and el Serafy, G. (2005). "Particle Filtering and Ensemble Kalman Filtering For Input Correction in Rainfall Runoff Modelling." International Conference on Innovation Advances And Implementation of Flood Forecasting Technology, Tromso, Norway. [Type text] 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 Weerts, A. H., and El Serafy, G. Y. H. (2006). "Particle filtering and ensemble Kalman filtering for state updating with hydrological conceptual rainfall-runoff models." Water Resources Research, 42(9). Willems, P. (2008). "Quantification and relative comparison of different types of uncertainties in sewer water quality modeling." Water Research, 42(13), 3539-3551. Winkler, S., Bertrand-Krajewski, J. L., Torres, A., and Saracevic, E. (2008). "Benefits, limitations and uncertainty of in situ spectrometry." Water Science and Technology, 57(10), 1651-1658. Xie, X. H., and Zhang, D. X. (2010). "Data assimilation for distributed hydrological catchment modeling via ensemble Kalman filter." Advances in Water Resources, 33(6), 678-690. Xu, C. (2003). "Discussion of "Fuzzy approach for Analysis of Pipe Networks" by Robero Revelli and Luca Ridolfi." American Society of Civil Engineers (ASCE), 129(7), 549-550. Yang, J., Reichert, P., Abbaspour, K. C., and Yang, H. (2007). "Hydrological modelling of the chaohe basin in china: Statistical model formulation and Bayesian inference." Journal of Hydrology, 340(3-4), 167-182. Zappa, M., Jaun, S., Germann, U., Walser, A., and Fundel, F. (2011). "Superposition of three sources of uncertainties in operational flood forecasting chains." Atmospheric Research, 100(2-3), 246-262. Zhang, H., and Pu, Z. (2010). "Beating the uncertainties: Ensemble forecasting and ensemble based data assimilation (review article)." Advances in Meteorology, 2010, Article ID 432160 10pp. Zhang, J., Song, R., Bhaskar, N. R., and French, M. N. (2007). "Ensemble Forecasting of Daily Water Demand." World Environmental and Water Resources Congress 2007: Restoring our Natural Habitat, Tampa, Florida, U.S.A. 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 [Type text]