A channel water temperature prediction method based on transfer learning and spatial-temporal graph neural networks

Accurate water temperature forecasting during winter operation periods is critical for the North Extension of the Eastern Route of the South-to-North Water Diversion Project, yet remains challenging due to limited historical data and sparse monitoring records. This study proposes a novel water temperature prediction model called Transfer-Learning Graph Temporal Convolutional Network(TF-GTCN) that integrates transfer learning with spatial-temporal graph neural networks. The model initially employs transfer learning techniques to capture the periodic variations in water temperature and its correlation with air temperature, leveraging ice-period scheduling data from the Central Route of the South-to-North Water Diversion Project. Subsequently, a spatial-temporal graph neural network is utilized to extract temporal features of water temperature, air temperature, and flow, alongside spatial dependencies across different cross-sections of the North Extension Project. By integrating the experiential knowledge obtained through transfer learning with the extracted spatial-temporal features, the proposed model effectively forecasts future water temperatures. The TF-GTCN model demonstrates significant improvements compared to traditional deep learning methods such as LSTM and GRU, achieving a mean absolute error (MAE) of [Formula: see text]C and reducing MAE by 0.72-3.29 at key monitoring stations. These advancements provide valuable insights for water transfer scheduling during ice-period operations.