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

Area of Science:

• Environmental Engineering
• Hydrology
• Artificial Intelligence

Background:

• Accurate water temperature forecasting is crucial for the South-to-North Water Diversion Project's winter operations.
• Challenges include limited historical data and sparse monitoring in the North Extension area.

Purpose of the Study:

• To develop a novel water temperature prediction model for winter operations.
• To improve forecasting accuracy in the North Extension of the Eastern Route.

Main Methods:

• Proposed a Transfer-Learning Graph Temporal Convolutional Network (TF-GTCN) model.
• Integrated transfer learning with spatial-temporal graph neural networks.
• Leveraged historical ice-period data from the Central Route for transfer learning.

Main Results:

• The TF-GTCN model achieved a mean absolute error (MAE) of [Formula: see text]°C.
• Demonstrated significant improvements over traditional methods like LSTM and GRU.
• Reduced MAE by 0.72-3.29°C at key monitoring stations.

Conclusions:

• The TF-GTCN model effectively forecasts water temperatures by integrating transfer learning and spatial-temporal features.
• Provides valuable insights for optimizing water transfer scheduling during winter ice periods.