Deep Learning Model-Based Demand Forecasting for Secondary Water Supply in Residential Communities: A Case Study of Shanghai City, China

To promote intelligent water services and accelerate the water industry's modernization process, accurately predicting regional residents' water demand and reducing energy consumption for secondary water supply is a major challenge for scientific scheduling and efficient management of urban water supply. This paper proposes a deep learning-based approach for demand forecasting in residential communities. The approach first identifies and corrects outliers in raw water supply data, and incorporates additional features such as epidemics and meteorological information. A long and short time Transformer model (LTMFormer) is then proposed, combining the recursive mechanism of the LSTM model and the parallel mechanism of the Transformer to achieve parallel output in the long-time series modeling task, improving both the prediction length and accuracy of the model. We evaluate our model on a metering dataset of 20 cells in Shanghai and compare it to traditional deep-learning models. Our experimental results demonstrate that the proposed model outperforms other deep learning models, achieving MSE, RMSE, and MAE scores of 3.337, 4.536, and 1.848 respectively on the test set. These results provide a theoretical and technical basis for further safeguarding public water health and meeting the growing demand for better urban water management.