Wavelet Transform based Bidirectional Long Short-Term Memory for Predicting Water Demand

The prediction of urban water demand is essential in rational allocation of water and developing smart city. Influenced through multiple factors, the water demand huge frequency noise and difficult patterns, that is due to nostationary of daily data has huge influence on generalization capability of methods. In this research, Wavelet Transform (WT) and Hodrick-Prescott (HP) techniques are utilized to decompose of daily data for resolving an issue of non-stationary. Bidirectional Long Short-Term Memory (BiLSTM), Seasonal Auto Regressive Integrated Moving Average (SARIMA) and Gaussian Radial Basis Function Neural Network (GRBFNN) are introduced to predicting various subseries. Ensemble learning is implemented for enhancing a generalization capability of methods and interval prediction is produced depended on t-test of students is handled along difference of laws in water supply. The coupling and WT techniques correctly predicted the water demand and give optimum 0.19% of Mean Square Error (MSE), 3.37% of Mean Absolute Error (MAE), 0.2 of Mean Relative Error (MRE), 97.25% of Nash-Sutcliffe Efficiency (NSE) and 0.99 of Correlation coefficient.