Physics-Informed Directed Graph Network-Based Temperature Forecasting Model.

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Accurate temperature forecasting is crucial, but faces challenges in modeling complex spatial and temporal data.
• Purely data-driven models struggle with the intricate relationships in temperature datasets.

Purpose of the Study:

• To develop a physics-informed directed-graph-based model for enhanced temperature prediction.
• To address limitations in existing data-driven temperature forecasting methods.

Main Methods:

• A directed graph design module constructed an asymmetric adjacency matrix based on station locations.
• Graph attention and graph-gating modules extracted spatial-temporal features using the directed adjacency matrix.
• A fusion module integrated features and the adjacency matrix for improved prediction.

Main Results:

• The proposed model demonstrated superior prediction performance in numerical simulations.
• Real-world data from southern China showed a mean absolute error of less than 0.75 °C.
• The physics-informed approach effectively captured asymmetric relations in temperature data.

Conclusions:

• The physics-informed directed-graph model offers a robust solution for accurate temperature forecasting.
• This method successfully integrates physical insights with graph-based deep learning.
• The model shows significant potential for improving environmental monitoring and prediction systems.