A simple forecast technique for estimating the statistical characteristics of water consumption József Bene, Dávid Bóka, Csaba Ho ˝ s and Varga Roxána ABSTRACT The forecast water consumption values are the most critical input data in the pump schedule optimization of water distribution systems. The aim of this paper is to present a simple technique which is able to estimate the mean consumption and its distribution for a given demand zone with an hourly resolution. Simplicity (low computational cost) is advantageous since the forecast model needs to be run for any optimization computation. The proposed technique uses a recorded hourly consumption database and consists of two steps. First, the database content is automatically grouped based on the similarity of the elements (more precisely, their normality). This step is time- consuming but is performed only once for a given database independently of optimization. The second step which is quick but has to be performed before the actual optimization makes use of this grouping for forecasting mean value and standard deviation. The proposed technique provides hourly water consumption predictions independently; that is, the neighbouring hours do not effect each other, which prevents the accumulation of prediction errors. The daily overall consumption is computed a posteriori. The test results presented in this paper prove the applicability of the technique for real-life problems. Moreover, it is demonstrated that the condence interval provided by the technique includes the actual measured data. József Bene (corresponding author) Dávid Bóka Csaba Ho ˝ s Varga Roxána Department of Hydrodynamic Systems, Budapest University of Technology and Economics, H-1521 Budapest, P.O. Box 91, Hungary E-mail: bene@hds.bme.hu József Bene Systems Engineering Laboratory, Department of Process and Environmental Technology, University of Oulu, FIN-90014 Oulu, P.O. Box 4300, Finland Key words | forecast technique, normality test, pump schedule optimization, water consumption INTRODUCTION Pump schedule optimization problems of water distribution systems have inspired many scientic studies resulting in a rich literature, see for example, Mays (), Nicklow et al. () and Ormsbee & Lansey (). These optimization methods are based either on heuristic (e.g. genetic algor- ithms) or deterministic (e.g. dynamic programming) methods; they can take into account full hydraulics or use simple ow-only models (Cembrano et al. ). Neverthe- less, even though the modelling and the solution technique of these methods are different, the aim is the same: to deter- mine the pump schedule for a given time horizon (typically for the next 24 h) which satises the complex constraint system of the optimization problem and offers the lowest possible energy cost or energy consumption. All of these methods require forecastwater consumption values as input data. The authorsexperience showed that the accuracy of the consumption is the bottleneck of the pro- blem: weak and unreliable forecast consumption results in such a signicant difference between the forecast and the measured values, that the computed optimalschedule is far from being optimal, or, even worse, it might not even satisfy the constraint system. Consumption can be involved in the computations as deterministic or stochastic data. In the rst case, when the water demands are assumed to be deterministic (Certainty Equivalent Control; Bertsekas ), they are usually deter- mined a priori from statistics or by some forecasting approach (Alvisi et al. ; Adamowski ; Bárdossy et al. ). When the stochastic behaviour is taken expli- citly into account, the consumptions are described by a priori determined distributions (Cervellera et al. ; Ikonen & Bene , ; Selek et al. ). 593 © IWA Publishing 2014 Water Science & Technology: Water Supply | 14.4 | 2014 doi: 10.2166/ws.2014.013