Available online at www.sciencedirect.

com

ScienceDirect
Procedia Engineering 70 (2014) 339 – 342

12th International Conference on Computing and Control for the Water Industry, CCWI2013

Demand forecasting for water distribution systems

J. Chena , D.L. Boccellia,∗
a Environmental Engineering Program, University of Cincinnati, Cincinnati, OH (USA)

Abstract
Short-term water demand forecasts can provide valuable information to distribution system operators for controlling the produc-
tion, storage and delivery of drinking water. Our current research is focused on developing an integrated Time Series Forecasting
Framework (TSFF) to statistically predict hourly/quarter-hourly demands in real-world, real-time scenarios. The first version of
TSFF has been prototyped within Matlab. Two forecasting models, a fixed seasonal auto-regressive (FSAR) model and an adaptive
seasonal auto-regressive (ASAR) model have been included within the framework for evaluation. The ASAR model self-updates
model parameters at run-time using maximum likelihood estimates (MLE). The framework has been applied to a real-world case
study of a system-wide water demand time series. With an underlying auto-regressive model, AR(3), the ASAR provides, on
average, 5.3% absolute relative prediction error (ARE) for lead-1 forecasts, 10.2% for lead-2 forecasts, and 14.2% for lead-3 fore-
casts. Computationally, the speed of the algorithm is such as to easily accommodate real-time activities. The framework will be
applied to additional demand time series in order to evaluate the performance of the forecasting algorithm and compare the water
consumption characteristics of different distribution systems.


© The Authors.
c 2013 The Authors. Published
PublishedbybyElsevier Ltd. Open access under CC BY-NC-ND license.
ElsevierLtd.
Selection and
Selection andpeer-review
peer-reviewunder
underresponsibility
responsibility
ofof
thethe CCWI2013
CCWI2013 Committee
Committee.

Keywords: Demand; forecasts; real-time; time-series; seasonal auto-regressive model

1. Introduction

Water demands are important indicators referred to by water distribution system operators to adjust production
and control pumps and valves. Research efforts on modeling demands have two major categories: the “bottom-
up” approach that models water usage at the level of individual users/meters, usually as stochastic processes (e.g.,
Buchberger and Wu (1995), Blokker et al. (2009)); and the “top-down” approach that models water usage at the
level of the whole system or pressure regions with statistical tools (e.g., Maidment and Miaou (1986), Magini et al.
(2008)). In practice, current Supervisory Control and Data Acquisition (SCADA) systems can usually provide system
or region-wide demand data in the form of time series, typically with resolution of equal or less than 15 minutes. The
high-resolution demands fit naturally in the statistical time series modeling and forecasting framework.
Time series analysis (Box and Jenkins (1976)) is a classic statistical method for analyzing discrete, infinite series of
random variables. A category of probabilistic models named AutoRegressive Integrated Moving-Average (ARIMA)

∗ Corresponding author. Tel.: +1-513-375-6901 ; Fax: +1-513-556-4162.

E-mail address: dominic.boccelli@uc.edu

1877-7058 © 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
Selection and peer-review under responsibility of the CCWI2013 Committee
doi:10.1016/j.proeng.2014.02.038
340 J. Chen and D.L. Boccelli / Procedia Engineering 70 (2014) 339 – 342

represents a type of time series analysis methods. Periodic time series can be modeled by seasonal ARIMAs. A
number of authors have applied the statistical method of time series analysis in the field of water demand forecasting
at different time scales (Sterling and Bargiela (1985), Jowitt and Xu (1992), Shvartser et al. (1993), Zhou et al. (2000),
Jain and Ormsbee (2002), Bougadis et al. (2005), Caiado (2009), Wong et al. (2010)). The early work of Sterling and
Bargiela (1985) recommended a seasonal ARIMA(1, 0, 1) × (1, 0, 1)48 model for half-hourly water demands and an
ARIMA(1, 0, 1) × (0, 0, 1)7 model for daily water demands. Similar work done by Jowitt and Xu (1992) identified an
ARIMA(7,1,1) model for daily water demand series of Thames Water Authority, UK. Maidment and Miaou (1986)
used an ARIMA(7,0,0) model in the daily demand model based on a study of nine American cities. Caiado (2009)
identified a double-level seasonal model of ARIMA(4, 0, 9) × (2, 1, 3)7 × (0, 1, 1)365 to represent daily water demands.
However, the aforementioned research did not consider the integration of time series methodology and the real-
time data provision facilities. Our current research is focused on both a time series estimation-forecasting algorithm
for hourly/sub-hourly demand data and a design of the real-time modeling/forecasting framework.

2. Time series analysis of water demands

2.1. SARIMA models

Assuming the demands can be expressed as an array of random variables with equal intervals. The ARMA models
are linear, Gaussian models with three subtypes (Box and Jenkins (1976)):

• Autoregressive-p (AR(p)) models,

• Moving-average-q (MA(q)) models, and
• ARMA(p,q) models.

Applying time series methodology involves three steps: model identification, parameter estimation, and forecast-
ing. The structural parameters p and q are usually determined by investigating the characteristic autocorrelation/partial
autocorrelation (AC/PAC) diagrams. Seasonal autoregressive integrated moving-average (SARIMA) models add pe-
riodicity to the ARMA models. In SARIMA, the differenced series is assumed to follow ARMA equations, formally
expressed as ARIMA(p, 0, q) × (0, D, 0) s when D is the level of differencing based upon seasonality s.

2.2. Model identification

The data used in this research was a system-wide demand series of a medium-sized water supply system located
in Florida, USA. The hourly demand data were preprocessed by dividing a benchmark demand to yield non-negative
demand multipliers. The hourly demand multipliers from 39 consecutive days were collected for statistical analysis.
The autocorrelation and partial autocorrelation diagram for the studied demand series suggests the importance of the
24-hour period of water consumption. As a result, the time series model was identified as a SAR(3) with 24-hour
period.

3. Online algorithms for estimation and forecasting

As described in the previous section, the model identification step, primarily via visual inspection, determined
the ARIMA model structure. If the water consumption behavior for the system remains stable, the model structural
parameters (p, q, and s) are assumed to remain unchanged. However, in typical real-time settings, updating model
parameters “on-the-fly” may be beneficial since the forecasts have to be renewed in real-time. Therefore, two combi-
nations of algorithms are designed and investigated in this article. The fixed seasonal auto-regressive (FSAR) model
has the parameters estimated only once, using the training data provided at the beginning of the forecasting; the adap-
tive seasonal auto-regressive (ASAR) model has a moving window of training data, and the parameters are updated in
real-time. Both methods include an algorithm for multi-step-ahead forecasting with confidence limits.
 J. Chen and D.L. Boccelli / Procedia Engineering 70 (2014) 339 – 342 341

The parameters of the SAR(3) model were estimated using the approach of Minimum Sum of Square Errors,
which is approximately equal to the maximum likelihood estimates (Box and Jenkins (1976)). While time series
models enable the forecasting of any time horizon into the future, the uncertainties of long-time forecasts are usually
very high, since two points in a series are only weakly correlated. Under stationary conditions, Box and Jenkins
(1976) showed that the Least Mean Squared Error (LMSE) forecasts were equal to the conditional expectations of the
variable given previous observations with confidence limits based residual white noise terms and current estimates of
the model parameters.

4. Results

Both FSAR and ASAR models were applied to the SAR(3) model with 24-hour periodicity. With FSAR, different
training set ratios were used to estimate the parameters. With ASAR, different lengths of the moving window of the
training set were used. Generally, estimation with shorter training periods captured more of the local properties of the
time series and had larger variance in parameter estimates. The autocorrelations of all estimated SAR models were
similar with each other.
The forecasting performance of FSAR and ASAR in the study case is summarized in Table 1 . Forecasting horizons
were set from one hour up to ten hours. The Average Absolute Relative Errors (AARE) were computed for the data
points in the test set. The Absolute Relative Error is defined as Bougadis et al. (2005):
 
1  |qobs − q pre |
AARE = (1)
N t qobs

Table 1: Results of hourly water demand forecasting with time series models

Time ASAR FSAR

horizon 100 200 500 50% 90%
Lead-1 0.059 0.055 0.054 0.053 0.060
Lead-2 0.113 0.104 0.104 0.102 0.112
Lead-3 0.156 0.140 0.144 0.142 0.155
Lead-4 0.183 0.167 0.171 0.169 0.187
Lead-5 0.197 0.183 0.188 0.186 0.205
Lead-6 0.207 0.192 0.197 0.194 0.215
Lead-7 0.211 0.195 0.201 0.198 0.220
Lead-8 0.210 0.197 0.202 0.200 0.222
Lead-9 0.209 0.197 0.201 0.199 0.222
Lead-10 0.205 0.197 0.199 0.198 0.220

According to Table 1, for the ASAR models, slight improvements in prediction power were observed if we in-
creased the size of the moving window used for parameter estimation, and using more than 200 data points for
training the model did not yield significantly better forecasts. For FSAR models, the model taking 50% of the data for
training generated marginally better results than the model taking 90% of the data, possibly due to the shift of weekly
demand pattern in the very last part of the data set. This weekly pattern can potentially be characterized in future
studies.

5. Summary

The discretized, aggregated hourly system-wide water demand series has significant daily periods. Applying a 24-
hour seasonal differencing operator filtered out these periodicities and effectively improved the model performance.
A general-purpose time series forecasting framework TSFF was developed in the study to generate both forecasts
342 J. Chen and D.L. Boccelli / Procedia Engineering 70 (2014) 339 – 342

and forecasting confidence limits for multiple time steps ahead of current time. The TSFF approach is applicable
for interfacing with water utilities’ SCADA infrastructures. The speed of the estimation/forecasting algorithm is
such as to easily accommodate real-time activities. The TSFF will be evaluated with additional time series model
structure (e.g., 24 hour and 7 day periodicities) and applied to additional demand time series in order to evaluate the
performance of the forecasting algorithm and compare the water consumption characteristics of different distribution
systems.

Acknowledgements

The authors would like to gratefully acknowledge the financial support provided by the CMMI Directorate, Civil
Infrastructure Systems (NSF) through Grant No. 0900713.

References

Blokker, E., Vreeburg, J., van Dijk, J., 2009. Simulating residential water demand with a stochastic end-use model. Journal of Water Resources
Planning and Management 136, 19–26.
Bougadis, J., Adamowski, K., Diduch, R., 2005. Short-term municipal water demand forecasting. Hydrologic Processes 19, 137–148.
Box, G.E.P., Jenkins, G.M., 1976. Time series analysis: forecasting and control. Holden-Day series in time series analysis and digital processing,
Holden-Day, Inc., Oakland, Calif.
Buchberger, S., Wu, L., 1995. Model for instantaneous residential water demands. Journal of Hydraulic Engineering 121, 232–246.
Caiado, J., 2009. Performance of combined double seasonal univariate time series models for forecasting water demand. Journal of Hydrologic
Engineering 15, 215–222.
Jain, A., Ormsbee, L.E., 2002. Short-term water demand forecast modeling techniques - conventional methods versus AI. Journal American Water
Works Association 94, 64–72.
Jowitt, P.W., Xu, C., 1992. Demand forecasting for water distribution systems. Civil Engineering Systems 9, 105–121.
Magini, R., Pallavicini, I., Guercio, R., 2008. Spatial and temporal scaling properties of water demand. Journal of Water Resources Planning and
Management 134, 276–284.
Maidment, D.R., Miaou, S.P., 1986. Daily water use in nine cities. Water Resources Research 22, 845–851.
Shvartser, L., Shamir, U., Feldman, M., 1993. Forecasting hourly water demands by pattern-recognition approach. Journal of Water Resources
Planning and Management - ASCE 119, 611–627.
Sterling, M.J.H., Bargiela, A., 1985. Adaptive forecasting of daily water demand, in: Comparative models for electrical load forecasting. John
Wiley & Sons Ltd, pp. 213–225.
Wong, J.S., Zhang, Q., Chen, Y.D., 2010. Statistical modeling of daily urban water consumption in Hong Kong: Trend, changing patterns, and
forecast. Water Resources Research 46, W03506.
Zhou, S.L., McMahon, T.A., Walton, A., Lewis, J., 2000. Forecasting daily urban water demand: a case study of Melbourne. Journal of Hydrology
236, 153–164.