Wind power forecasting is important for renewable energy integration, grid stability, and operational planning. However, model performance may be overestimated when temporal data leakage occurs or when historical power values are used without careful validation. This study proposes a leakage-aware AutoML framework for wind power forecasting using only instantaneous wind speed and wind direction information. The publicly available Wind Turbine SCADA Dataset from Kaggle was used. The dataset was sorted chronologically and divided into training and test sets using an 80%–20% split. Outliers were replaced with median values using thresholds derived from the training set. Wind direction was transformed into sine and cosine components to represent its circular nature. AutoGluon was used to compare multiple regression models under the same experimental setting. The selected Weighted Ensemble model achieved an RMSE of 505.90 kW, MAE of 250.63 kW, and R² of 0.8579 on the chronological test set. For interpretability, SHAP analysis was conducted using the LightGBM model, which achieved a similar R² of 0.8560. SHAP results showed that wind speed was the dominant predictor, while wind direction components had secondary effects. The cubic fit of wind speed SHAP contributions achieved an R² of 0.9231, supporting the physical consistency of the learned nonlinear relationship. The results show that leakage-aware AutoML combined with SHAP can provide accurate and interpretable wind power forecasts.

Keywords

Wind Power Forecasting, AutoML, SHAP, Physical Consistency, Data Leakage