Explainable deep learning with Shapley additive explanations for multivariate wind power forecasting

Wind is a popular renewable energy source, but it is highly dependent on environmental conditions and internal dynamics. Given these complexities, the need for accurate forecasting models that account for multiple variables is critical. However, the challenge extends beyond accuracy to the interpretability of these complex models. This study focuses on the interpretability and explainability of complex deep learning models, specifically for wind power forecasting. A stacked bidirectional long short-term memory (LSTM) model is utilized, which can capture intricate patterns. After ten independent training rounds with varying initializations, the model achieved an average coefficient of determination of 0.923. To interpret the influence of various input variables, including wind speed, wind direction, air density, turbulence intensity, and below hub height wind shear, on the model’s power predictions, Shapley additive explanations (SHAP) were applied. This analysis revealed that wind speed is the most influential variable, with a mean absolute SHAP value of 30.73. It is followed by below-hub-height wind shear, which had a mean absolute SHAP value of 5.12. Besides the SHAP analysis, a comparative analysis with traditional correlation coefficients (Pearson, Spearman, and Kendall) was conducted. The results indicate that correlation coefficients cannot reliably assess feature influence on model predictions unless a feature strongly dominates the model’s output.