TimesInformer: An Approach to Predicting Urban Water Supply with Multi-Periodic Features

Urban water supply forecasting is a critical component of urban planning and management. By predicting the city’s water supply, water companies or relevant management units can better schedule and manage the water supply system. This allows for the rational arrangement of water supply plans, adjustment of water pressure and flow, improvement of water supply efficiency and service quality, and avoidance of water shortages or waste. According to the requirements of the water utility group, the water supply data is aggregated into hourly units. The time series data of water supply by hour exhibits distinct multiperiod characteristics. Therefore, it is necessary to consider both the intra-period (24 hours) regularities of the water supply data and the inter-period (weekly, monthly, and quarterly) regularities. In response to the intra-period and inter-period characteristic relationships in water supply data, we propose the TimesInformer model. This model transforms 1D time series data into a 2D tensor and utilizes the Inception module within it to extract the multiperiod features of water supply data, thereby capturing its periodic trends. Through adaptive fusion, the 2D tensor is converted back into one-dimensional features to achieve information aggregation. This model was compared with 15 cutting-edge time series forecasting models. The results show that the TimesInformer model proposed in this paper achieves the highest prediction accuracy and best performance in preprocessing water supply data.