Decoding wind power uncertainty: A Bayesian-optimized machine learning approach for multi-timescale dynamic transitions

Abstract

Due to the inherent instability and dynamic multi-timescale transitions of wind power generation, effectively decoding its uncertainty is critical for operational planning and grid stability. This study developed a machine learning framework using real-world data from a wind farm in Xinjiang. The framework integrates tree-based ensemble models to capture non-linear interactions and handle anomalies, Bayesian optimization for hyperparameter tuning, and Shapley Additive Explanations (SHAP) to quantify feature contributions. CatBoost, identified as the best-performing model, achieved an R² improvement from 0.9372 to 0.9631 when combined with Bayesian optimization. The results reveal that 50m wind speed positively impacts power generation when exceeding 5 m/s, whereas 10m turbulence intensity exhibits nonlinear degradation characteristics, specifically manifesting detrimental influences on energy conversion efficiency when exceeding the threshold of 0.2. Wind speed emerges as the most critical factor across all seasons. Beyond wind speed, the importance of features varies seasonally. Key feature impacts vary throughout the day: 50m wind speed in spring, wind direction in summer, and temperature in winter generally enhance power generation, though their effects weaken during the day. In autumn, 10m wind direction aids generation during the day but hinders it at night. Conversely, higher summer temperatures and lower winter wind speeds variably negatively affect power output. Wind direction fluctuations at 50m reduce daytime generation in spring and autumn. This study provides critical insights for optimizing wind farm operations and enhancing grid stability, offering actionable recommendations for policymakers and renewable energy stakeholders.