A Novel Model Based on Deep Learning Approach Combining Data Decomposition Technique and Grouping Distribution Strategy for Water Demand Forecasting of Urban Users

Abstract

Urban user water demand prediction (WDP) is of significant importance for smart water supply system, which can provide a strong decision-making basis for the dispatching and management of smart water supply system. However, owing to the fluctuation, intermittence and nonstationarity of the user’s water consumption in urban buildings, it is extremely difficult to predict accurately. Therefore, a novel short-term WDP model (Singular Spectrum Analysis Convolutional Neural Network Bidirectional Gate Recurrent Unit, SSA-CNN-BiGRU) is proposed to promote the stability and accuracy of WDP, which successfully introduces organic combinations including deep learning, decomposition technique, and data partitioning policies into the domain of WDP. First, raw data are decomposed into components that carry distinct frequency signals for weakening its nonstationarity and complexity. Then, all the components are automatically divided into several groups using clustering algorithm based on their entropy, after which deep learning method is adopted to predict by groups. Finally, the predicted result of each group is summed up to be fused as the final value. To validate the predictive performance of SSA-CNN-BiGRU, real data have been selected for this study. In experiments, SSA-CNN-BiGRU achieved a fitting of 94.73%. Comparison by relevant evaluation metrics demonstrates that the proposed model exhibits superior performance, thus providing a more accurate basis for WDP.