Short-Term Water Demand Forecasting Using Nonlinear Autoregressive Artificial Neural Networks

Abstract

Short-term water demand forecasting models address the case of a real-time optimal water pumping schedule. This study focuses on developing artificial neural network (ANN) models to forecast water demand 24 h and 1 week ahead. A number of studies have shown that the relationship between water demand and the driving variables is nonlinear. Two ANN time-series models were developed, a nonlinear autoregressive with exogenous (NARX) model with historical demand and weather data as an exogenous input, and a nonlinear autoregressive (NAR) model with only historical demand as an input. This investigation examines how model structure, length of historical data span, and improvement of an exogenous input can influence model performance. The results show that on average, using a nonlinear ANN model can improve water demand prediction by 18% and 25% when forecasting 24 h and 1 week ahead, respectively. The results also show that training the model (i.e., NARX) with correlated exogenous parameters dropped the error by 30% on average compared with a single-input model. In addition, using historical data for only 4 months compared with 5 years and 1 year decreased the error by 76% and 68% for NARX models and 35% and 33% for NAR models, forecasting 24 h and 1 week ahead, respectively.