Dynamic Forecast of Daily Urban Water Consumption Using a Variable-Structure Support Vector Regression Model

Abstract

A reliable forecasting model for daily water consumption would provide the data basis for scheduling urban water supply facilities. In this paper, a variable-structure support vector regression (VS-SVR) model is developed for dynamic forecast of the water consumption. Considering the nonlinear mapping capability of the SVR, the next-day water consumption is associated with the past water consumption series using the SVR model. To better accommodate the dynamic characteristics, the model structure of the SVR is variable in response to the receding horizon of the water consumption series. The variable model structural parameters are obtained using an extended Kalman filter (EKF) as the feedback correction tool. Combining the robustness of the model predictive control framework and the nonlinearity of the SVR, the proposed VS-SVR model is a dynamic approach to forecasting daily urban water consumption, evaluated using real data collected from a water company from January 2010 to December 2011. Compared with the SVR model, the dynamic forecast of daily urban water consumption using the proposed VS-SVR method improves the one-day-ahead forecast mean absolute error by