Photovoltaic power prediction based on BO-LightGBM collaborative prediction and K-means partition optimization

Abstract

As the global energy mix accelerates toward a clean and sustainable system, accurate and reliable photovoltaic power generation forecasting is crucial for ensuring grid stability and optimizing resource allocation. This study aims to improve the day-ahead forecasting performance of photovoltaic power generation under high-volatility scenarios. A collaborative forecasting model based on Bayesian Optimization (BO) and LightGBM is proposed, supplemented by a K-means clustering and partitioning optimization strategy. The model first applies K-means clustering to analyze total irradiance, a core driver, and successfully identifies high-risk periods of high power accompanied by high volatility, providing a divide-and-conquer foundation for the forecasting strategy. LightGBM is selected as the core forecasting engine to accurately capture the complex nonlinear relationship between meteorological characteristics and power output. Furthermore, the BO mechanism is innovatively introduced for global hyperparameter optimization, significantly improving model learning efficiency and forecast robustness. Feature importance analysis reveals the dominant feature sets consisting of light intensity, air pressure, and temperature. Experimental results demonstrate that this integrated model not only achieves superior forecast accuracy compared to traditional methods but also effectively controls forecast errors caused by peak fluctuations through a partitioning optimization strategy, providing a data-driven technical solution for optimizing resource allocation in power grid systems.