Water Level Prediction at Cascade Pump Stations Based on Multi-Scale Augmented Temporal Decomposition Network

Xumeng Shen, Yongchuan Yu, Jianzhuo Yan, Chunyi Hou, Siyao Zhang

Abstract: Maintaining the pool water level within a safe range is crucial for the operational stability and safety of cascade pump stations, making accurate water level forecasting essential for optimal performance. In response, this study proposes a multi-scale augmented temporal decomposition network model, consisting of four main parts: a multi-scale feature augmented module, a feature data decomposition module, a feature data extrapolation module, and a fusion prediction module. The multi-scale feature enhancement module uses pooling layers and linear variations to enhance the comprehension of the original time series data, and the feature data decomposition module progressively decomposes the enhanced time series data into information containing different patterns such as trends and cycles, which are further analyzed by the corresponding extrapolation module. In this case, the regular information is extrapolated according to its evolution, the irregular information is extrapolated through the Basic Probability Assignment (BPA) module, and then the fusion prediction module integrates the extrapolation results of each part of the information to generate the final prediction results. In this paper, a study is conducted based on the actual operation data of the three-stage cascade pump stations of the East- West Water Transfer. The results show that our multi-scale augmented temporal decomposition network has better error metrics such as MAE and MSE than other models, and can capture and predict the evolution of the temporal data more accurately. The significance of this study is that by using the multi-scale augmented time-series decomposition network-based model, we can accurately predict the water level in the pools of the cascade pump stations, which provides a reliable reference basis for water dispatching.