通过特征空间匹配分析解释基于时空相关性的LASSO回归模型用于风电功率预测

Abstract Interpreting well-performing wind power forecasting (WPF) models is essential to advance more trustworthy and accurate forecasting methodologies. Current research primarily focuses on interpreting black-box deep learning models, overlooking self-interpreting models that can directly indicate feature importance but fail to explain the underlying reasons. Self-interpreting regression models based on the least absolute shrinkage and selection operator (LASSO) excel in WPF. Therefore, it is crucial to explore their underlying decision logic and the practical implications of their coefficients to extract beneficial domain knowledge. An interpreting framework is proposed to elucidate the decision logic of the LASSO regression in WPF considering spatio-temporal correlations. The framework includes four main components. Firstly, a spatio-temporal correlation quantification system is established for feature selection for target wind farms , utilizing metrics that reflect spatial correlations , temporal fluctuations, geostatistics, and causalities of wind farms’ power output . Secondly, feature matching analysis is performed by comparing the features selected and ranked by the quantification system with those selected by the LASSO model. Thirdly, based on the spatio-temporal patterns and key features identified from the preliminary feature matching analysis, a feature perturbation analysis is conducted by modifying the feature space to assess how changes in spatio-temporal features impact forecasting accuracy. Finally, a sensitivity analysis is conducted by setting different LASSO parameters to verify the consistency of the extracted domain knowledge. The proposed framework is applied to two datasets, yielding substantial qualitative and quantitative results. Critical factors affecting WPF accuracy, such as feature collinearity, the number and spatial dispersion of reference wind farms , and how these factors influence forecasting accuracy are effectively identified. The framework and findings are effective, consistent and demonstrates generalizability across different datasets.