Enhancing wind power prediction accuracy: A novel method integrating seasonal temporal factors and advanced spatio-temporal feature extraction

Accurate wind power forecasting is important for day-ahead power dispatch and the widespread adoption of renewable energy sources. However, the complex spatio-temporal dependencies and seasonal variations of meteorological factors pose substantial challenges to forecasting accuracy. In this paper, we propose a novel wind power forecasting method based on enhanced spatio-temporal feature extraction, which improves the accuracy and robustness of predictions through innovative model design. Firstly, to address the differences in multiple meteorological factors across different seasons, we construct a seasonal time factor expression that enables the model to dynamically adapt to seasonal meteorological changes. Secondly, we propose a Local-Global Spatio-Temporal Graph Convolutional Network (LG-STGCN) for correlation analysis, incorporating a Random EdgeDrop module proposed in this study to capture the spatio-temporal dependencies of meteorological data. Subsequently, the Pearson correlation coefficient is employed for feature dimensionality reduction and weighting. Finally, by integrating the seasonal time factor weights with the correlation weights, we input the combined features into a Bidirectional Long Short-Term Memory (BiLSTM) network for power forecasting, thereby enhancing the model's adaptability and dynamic interpretability. Experimental results demonstrate that the proposed method significantly outperforms existing benchmark models in terms of prediction performance across various wind farm datasets, particularly in scenarios with pronounced seasonal variations. This study provides a new perspective and an effective solution for wind power forecasting.