Optimised control and pipe burst detection by water demand forecasting

Abstract

Water demand forecasting The total water demand in an area is the sum of the water demands of all individual domestic and industrial consumers in that area. These consumers behave in repetitive daily, weekly and annual patterns, and the same repetitive patterns can be observed in the drinking water demand. The observations of the water demand were used to develop a fully adaptive forecasting model for short-term drinking water demand. The forecasting model automatically stores and updates water demand patterns and demand factors, and uses these when forecasting the water demand for the next 48 hours with 15-min. time steps (192 values). The model uses as single input the measured water demand and calendar information that appoints deviant days. The model is easy to implement, fully adaptive and accurate, which makes it suitable for application in real time control and pipe burst detection. The model was tested on datasets containing six years of water demand data in six different areas in the central and southern part of Netherlands. The areas vary in size from very large (950,000 inhabitants) to small (2,400 inhabitants). The mean absolute percentage error (MAPE) for the 24-hours forecasts varied between 1.44-5.12%, and for the 15-min. time step forecasts between 3.35-10.44%. When using temperature information as extra input, the average forecasting errors could be reduced by 6.3%, and the largest forecasting errors by 9.4%. ? Optimised control A first application of a short-term water demand forecasting model is using it for optimised control of water supply systems. The conventional automatic control of the production flow or the clear water pumps is often quite simple, resulting is highly varying production or transportation flows. This basic control results in sub-optimal operation of the system, and the operation can be improved when forecasts of the water demands in the system are used. To assess the differences between conventional basic control and optimised predictive control, five existing water supply systems in the Netherlands and one system in Poland were examined. The operational results in a period with conventional control were compared to the results in a period with optimised control. The results showed that the variation in the production flow was 75% lower with optimised control, which resulted in 17% lower turbidity rates in the clear water. The optimised control also resulted in a reduction of the energy costs of 5.2% at the Dutch systems and 11.5% at the Polish system. Pipe burst detection A second application of a short-term water demand forecasting model is using it for pipe burst detection. The now-cast of the water demand a good estimator for the actual water demand under normal circumstances. By comparing the measured water demand to this forecasted water demand, anomalies like pipe bursts can be detected. A pipe burst detection method based on this principle was developed. In the method, all measured and forecasted signals are transformed to moving averaged values over time frames of 2 up to 240 minutes. The transformation to longer moving average time frames resulted in lower threshold values which enabled the detection of smaller pipe bursts. The threshold values that distinguish between normal forecasting inaccuracies and pipe bursts, are derived by evaluating the forecasting deviations in the year prior to the monitoring year. The method was tested on different historic datasets with hydraulic data and pipe burst information in three areas in the western part of the Netherlands, and six areas in the northern part of the Netherlands. The method proved to be effective for detecting the relatively larger bursts: 80-90% of the bursts could be detected within 20 minutes, while generating false alarms on 3% of the days without a burst. The size of the pipe burst that can be detected showed a strong relation with the size of the area. Based on an analysis of problematic bursts, it was found that the burst detection method can effectively be applied to areas with an average demand of 150 m3/h or less.