Short-term water demand forecast based on automatic feature extraction by one-dimensional convolution

Abstract

Short-term water demand forecast is one of the most important technology for urban water supply management. The accuracy and timeliness of the forecast have an important impact. Most of the reported water demand forecast models based on deep learning methods apply a manual features extraction strategy, resulting in incomplete mining of the data and weak model self-adaptability capability. To address these issues, a new framework of short-term water demand forecast is proposed, in which a data preprocessing approach, S-H-ESD (Seasonal Hybrid Extreme Student Deviate), and a forecasting model, Conv1D-GRU (one-dimensional convolution-gated recurrent unit) are mainly developed. Based on the historical monitoring data, different hyper-parameter settings and training strategies were carried out with the proposed models. The results show that the data preprocessing model S-H-ESD can effectively deal with a variety of abnormal values, therefore significantly improving the accuracy of forecast (When the training dataset length is 7 days, the average accuracy of the three models is improved by1.23% when using S-H-ESD method compared with Z-Score method) and the Conv1D-GRU model shows better capability in forecast accuracy and self-adaptability of data features extraction compared with other models in literature (GRUN, ANN). With the achieving optimal parameter setting and training strategy, the developed methodology shows the best forecasted value of MAPE and NSE indicator are 1.677%, 0.983, respectively.