Mingu Kang  Taewook Heo

Vol. 31,  No. 3, pp. 182-191, Sep.  2025
10.7464/ksct.2025.31.3.182

Wind power forecasting Computational Efficiency Time-Series Forecasting Turbine clustering XGBoost Wind farm heterogeneity

PDF

Accurate wind power forecasting is crucial for grid stability and renewable energy management. While large-scale, data-driven AI models demonstrate high forecasting capabilities, their real-world application is often constrained by significant computational resource limitations. This paper presents a spatially-aware clustering framework for improving the accuracy and computational efficiency of wind power forecasting. The approach leverages turbine-level SCADA data and applies KMeans clustering based on geographical and operational features. For each resulting cluster, individual forecasting models are constructed using the XGBoost algorithm. Experimental results demonstrate that the proposed framework achieves a forecasting accuracy within a 0.9% RMSE difference compared to the global baseline model (RMSE of 467.06 vs. 464.60) for the challenging 48-hour forecast, and reduces the inference time by approximately 86% (from 0.27 s to 0.04 s on average). Additional evaluations using hierarchical agglomerative clustering (HAC) and DBSCAN confirm the robustness of the method. To enhance the interpretability, cluster-level analyses are performed, including comparisons of the average wind speed, output variability, and feature importance. The results indicate that clusters exhibiting more stable environmental conditions tend to yield higher forecasting accuracy. Overall, the proposed approach both enables localized modeling for faster inference and maintains reliable forecasting performance, making it suitable for practical deployment in real-time energy management systems.