Machine Learning-Based Multi-model Ensemble Approach of Wind Speed Forecasting for Wind Farm

Based on the forecasts of four numerical models: CMA-WSP, CMA-MESO, CMA-GD and WRF-SOLAR, as well as the observation data collected from the wind farm in 2022, the ensemble forecasting experiments were conducted using four machine learning ensemble models. Results show that multi-model ensemble significantly outperforms individual numerical models. The optimal ensemble model, LightGBM, reduces the mean absolute error (MAE) and root mean square error (RMSE) by 25.29% and 21.82%, respectively, compared with the best-performing numerical model, and improves the correlation coefficient by 10.81%. The best-performing ensemble model varies across months and seasons: LightGBM performs best from January to June and in October, ET achieves the best results from July to September and in December, while XGBoost is optimal in November. When wind speed is below 8 m·s−1, the multi-model ensemble reduces RMSE by more than 30% compared to the optimal numerical model. However, when wind speed exceeds 8 m·s−1, the ensemble model deteriorates, likely due to the limited sample size and the complexity of weather processes. The multi-model ensemble effectively mitigates the systematic daytime overestimation of the numerical model, achieving reductions in MAE and RMSE of 44.74% and 40.85%, respectively. Moreover, as forecast lead time increases, the multi-model ensemble exhibits a slower degradation in performance. Case studies indicate that the multi-model ensemble captures wind speed peaks more accurately during cold surge events, but its performance during typhoon events remains inferior and requires further improvement.