A new hybrid approach for short-term water demand time series forecasting

This study proposed a new hybrid approach for short-term water demand prediction. Raw water demand data were decomposed into a set of intrinsic mode functions (IMFs) component and a residue by using ensemble empirical mode decomposition (EEMD) method. IMF₁ is the main random component of the raw water demand data, the continuous deep belief neural network (CDBNN) model is used to predict IMF₁. Other IMFs and residue can be recombined employing autoregressive integrated moving average (ARIMA) model for forecasting, this method reduces forecasting steps without increasing the forecasting error. The final forecasting result was obtained by summing the forecasting results of the two models. A historical hourly water demand data from an urban waterworks of Zhuzhou, China were investigated by the hybrid method. Simulation results indicate that the proposed approach has higher forecasting accuracy than back propagation neural network (BPNN), support vector regression (SVR), EEMD and their combinations.