Abstract: To address the issues of low accuracy in traditional neural network-based forecasting models under specific weather conditions and the lack of consideration for environmental changes, a short-term photovoltaic（PV） power forecasting method based on multi-mode incremental update is proposed. By analyzing weather features, generalized weather types are forecasted based on historical data. Then, corresponding training methods and data enhancement techniques are developed according to the forecasting weather types for the following day. Finally, by using parameter freezing technology, the model is incrementally updated so that its ability to depict special weather and adapt to subsequent environments is enhanced. Experiments on a real-world PV dataset demonstrate that the proposed method effectively improves forecasting accuracy.