The photovoltaic microgrids for power systems has witnessed a steady increase over the years. Its principal objective is to schedule the photovoltaic power generation and stored electricity, thereby optimizing the utilization of solar energy. Accurately predicting photovoltaic power generation at each juncture stands as a prerequisite for proficiently managing the electricity scheduling. Most of the existing popular forecasting methods forecasting techniques often prioritize overall prediction performance, neglecting significant prediction errors occurring at peak and valley points. In order to mitigate the challenge posed by significant prediction errors at peak and valley points, this paper proposes a novel switched data-driven model based precise forecasting method. In this method, we first establish a global model, then identify regions with relatively large errors in the predictions of the global model. Subsequently, we train several local models using data from these regions and re-train the global model with the remaining data. Finally, we combine the updated global model with the local models to establish the switched data-driven model. Experimental findings indicate that in comparison to traditional forecasting approaches, the proposed switched data-driven model yields a notable enhancement in prediction accuracy within localized regions, thereby refining the overall model precision, with a minimum reduction of 36% in RMSE and 31% in MAE.