Linear regressions and neural approaches to water demand forecasting in irrigation districts with telemetry systems

Information regarding water demand is key to managing consumption in irrigation districts. Forecasting water demand is one of the main problems for designers and managers of water delivery systems. This paper evaluates the performance of linear multiple regressions and feed forward computational neural networks (CNNs) trained with the Levenberg–Marquardt algorithm for the purpose of irrigation demand modelling. The models are established using data recorded from an irrigation water distribution system located in Andalusia, southern Spain, during two irrigation seasons (2001/2002, 2002/2003). A commercial telemetry system was installed on 28 farms of the irrigation network to record water volumes in real time. The input or independent variables used in various CNN and multiple regression models are: (a) water demands from previous days; (b) water demands and climatic data (rainfall, maximum, minimum and average temperatures, relative humidity and wind speed) from previous days. Good predictions were obtained when water demand original data were modified in the calibration period by a smoothing process to reduce the noise in the data acquisition during the start-up of the research project. The best predictions were obtained when water demand recorded during the two previous days was used as input data.