DOI: 10.67047/tepes.1911904 ISSN: 2791-6049

Interpretable Machine Learning for Wind Power Forecasting: A Decision Support Framework with Hybrid ARIMAX Residuals

Halit Alper Tayalı, Emrah Akbulut, Birgül Küçük Çırpın, Lale Erdem Atılgan
Show PDF Cite
The integration of wind energy into power grids requires reliable and efficient decision support tools. While complex machine learning models are prevalent for wind power forecasting, their black-box nature and susceptibility to temporal overfitting can limit their operational utility. This study introduces a decision support framework for 10-minute-ahead wind power forecasting. Unlike many prior studies, the framework deliberately retains negative power values to represent turbine idle and auxiliary consumption states. It also applies strict chronological data partitioning to prevent temporal leakage. The framework further employs a novel hybrid feature engineering strategy, incorporating the residuals from the autoregressive integrated moving average with exogenous variables (ARIMAX) model to capture latent temporal structure in the data. The proposed framework evaluates machine learning algorithms—multiple linear regression (MLR), ARIMAX, decision tree, random forest, and extreme gradient boosting. Empirical results from the supervisory control and data acquisition (SCADA) benchmark dataset indicate that the interpretable MLR model consistently outperforms complex tree-based ensemble models on unseen data. Integrating a lagged dependent variable into the MLR model further captures previously unmodeled temporal dependence, reducing the test root mean squared error by 45% (from 361.0 to 200.0) and increasing the coefficient of determination by 5%, outperforming benchmark models reported in the literature for this dataset. The findings suggest that prioritizing statistical transparency, temporal validation, and physically consistent data preprocessing leads to an operationally reliable decision support tool. The proposed framework supports grid operators in improving turbine operational planning.

More from our Archive