This paper introduces a comprehensive forecasting model that integrates kernel principal component analysis (KPCA), grey wolf optimizer (GWO) and light gradient boosting machine (LightGBM), which aims to address the challenges posed to grid scheduling by large-scale photovoltaic (PV) grid connection. Firstly, KPCA is used to identify key factors influencing PV power, selecting correlated meteorological data as relevant features. These features and historical power data are used to train the LightGBM model. GWO optimization is applied to prevent the model from getting stuck in local optima, resulting in highly accurate ultra-short-term PV power predictions. The simulation results demonstrate that the GWO-optimized LightGBM significantly enhances prediction accuracy, reducing the mean absolute percentage error (MAPE) by 4.482% and the root mean square error (RMSE) by 0.289 kW. Furthermore, this proposed model surpasses XGBoost, achieving a reduction in MAPE and RMSE by 7.324% and 0.394 kW, respectively. These findings suggest that the proposed model effectively captures the variability in photovoltaic power and delivers more consistent prediction results. This research represents a promising avenue for further exploration in the field of ultra-short-term PV power prediction.