A spatiotemporal wind power forecasting method based on dual-view graph fusion and dual-granularity residual learning

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence
• Graph Neural Networks

Background:

• Accurate wind power forecasting is essential for stable power grid operation.
• Existing spatiotemporal graph neural network (ST-GNN) methods struggle with static spatial representations, limiting their ability to model complex turbine interactions.

Purpose of the Study:

• To develop an advanced model for wind power prediction that overcomes limitations of existing methods.
• To improve the accuracy and robustness of wind power forecasting for turbine clusters.

Main Methods:

• Proposed a Spatiotemporal Dual-View Graph Network (ST-DVGN) utilizing two complementary views: physical proximity and historical power dependency.
• Integrated dual-view graph representations over a fixed topology using a unified fusion mechanism.
• Incorporated a dual-granularity temporal module to capture short-term fluctuations and long-term trends.

Main Results:

• ST-DVGN consistently outperformed baseline models in prediction accuracy.
• Demonstrated high prediction stability and robustness, with a maximum RMSE decrease of only 6% under incomplete observation scenarios.
• Validated the effectiveness of dual-view spatial modeling in realistic operating environments.

Conclusions:

• The proposed ST-DVGN offers a significant advancement in wind power forecasting accuracy and reliability.
• Dual-view spatial modeling provides a robust solution for wind power prediction, particularly in challenging operational conditions.
• The model's ability to handle incomplete data highlights its practical applicability for power systems.