Deep learning with long short-term memory neural networks combining wavelet transform and principal component analysis for daily urban water demand forecasting

Abstract A reliable and accurate urban water demand forecasting plays a significant role in building intelligent water supplying system and smart city. Due to the high frequency noise and complicated relationships in water demand series, forecasting the urban water demand is not an easy task. In order to improve the model’s abilities in handling the complex patterns and catching the peaks in time series, we propose a hybrid long short-term memory model combining with discrete wavelet transform (DWT) and principal component analysis (PCA) pre-processing techniques for water demand forecasting, i.e., DWT-PCA-LSTM. First, the outliers of water demand series are identified and smoothed by 3σ criterion and weighted average method, respectively. Then, the noise component of water demand series is eliminated by DWT method and the principal components (PCs) among influencing factors of water demand are selected by PCA method. In addition, two LSTM networks are built to yield the daily urban water demand predictions using the results of DWT and PCA techniques. At last, the superiorities of the proposed model are demonstrated by comparing with the other benchmark predictive models. The water demand from 2016 to 2020 of a waterworks located in Suzhou, China is used for the experiment. The predictive performance of the experiments are evaluated by the mean absolute percentage error (MAPE), mean absolute percentage errors of peaks (pMAPE), explain variance score (EVS) and correlation coefficient (R). The results show that the DWT-PCA-LSTM model outperforms the other models and has satisfactory performance both in catching the peaks and the average prediction accuracy.