Abstract: To address the strong fluctuations, pronounced intermittency, and limited predictive accuracy of single models in photovoltaic power generation under complex meteorological conditions, this paper proposes a photovoltaic power forecasting method that integrates K-Means++ clustering with multi-time-window weighted Stacking ensemble learning. First, the measured data from a photovoltaic power station are cleaned and feature-engineered, and weather types are classified using K-Means++. Then, a heterogeneous forecasting framework composed of XGBoost, ExtraTrees, RF, KNN, and Ridge is constructed, and robust fusion is achieved through multi-time-window weighted Stacking and a safety-anchoring mechanism. Under sunny conditions, the RMSE and MAE of the Stacking model are reduced by 1.14% and 1.52%, respectively, compared with the best single model. Under cloudy and rainy conditions, the RMSE is reduced by 0.19% and 1.18%, respectively, while the MAE is reduced by 0.64% and 1.32%, respectively. The proposed method can provide technical support for the optimal operation of photovoltaic power stations and the secure dispatch of power grids under complex meteorological conditions.