A short-term power forecasting method for wind farm clusters considering feature enhancement and real-time optimization of dynamic graphs

This paper proposes a short-term power forecasting method for wind farm clusters that integrates feature enhancement and real-time dynamic graph optimization. The proposed method improves the accuracy of wind farm cluster power forecasting. First, we propose a historical wind power output scenario reuse mechanism. This mechanism provides a new decision-making basis for power forecasting, thereby achieving feature enhancement. Then, the deep deterministic policy gradient algorithm is adopted to construct a data pool for long-term correlation representation. It perceives real-time correlation changes to dynamically adjust the graph structure. The objective function is optimized based on three indicators: accuracy, complexity, and stability. Finally, a residual graph network structure is proposed to improve the stability of network training. A dynamic-static dual-channel graph convolution is adopted to enhance the effectiveness of spatial correlation mining. The multi-task learning mechanism is introduced to optimize the output layer. The proposed method is validated on a wind farm cluster in Jilin Province, China. The average short-term power forecasting accuracy for 20 wind farms reaches 92.96%, demonstrating the effectiveness of the proposed method.