Interpretable Wind Power Prediction: A Machine Learning Perspective Using Lightgbm and SHAP

As the global demand for renewable energy continues to grow, wind power, as a vital component of clean energy, is gaining increasing attention. Accurately predicting the power output of wind turbines is a challenging task, which is crucial for the power sector in scheduling and reducing grid impact. This paper addresses the issue by studying the power generation dataset from a wind farm using various Prediction models. With a training-to-testing ratio of 8:2, we optimized the model's hyperparameters through grid search to predict power output. The results show that the LightGBM model performed best, with an R-squared value of 0.9937 and a Root Mean Square Error (RMSE) of 0.023. Compared to traditional numerical calculations, artificial intelligence predictions not only offer high accuracy but also save considerable computational resources. Furthermore, this paper investigates the interrelationships among various indicators based on SHAP theory and Pearson correlation coefficients. We found that maximum power (MAXP) and minimum power (MINP) had the greatest impact on average power (AP), followed by density-adjusted wind speed (DA WS) and power standard deviation (SDP). We validate the model's performance across different geographic locations, including additional turbines in England and data from China, achieving consistent accuracy. The analysis highlights the potential for simplifying wind power predictions by focusing on maximum and minimum wind speeds, which could reduce the need for high-frequency measurements and improve operational efficiency. This approach aids in optimizing grid integration and minimizing the impact on power stability.