Ultra-Short-term PV Power Forecasting Based on Multiple Weather Scenaios and Incremental Learning

Photovoltaic (PV) power generation is subject to fluctuations caused by complex weather conditions, resulting in challenges for ultra-short-term forecasting models. The accuracy of traditional forecasting methods is often reduced in such scenarios due to limited available data. This paper presents an approach that improves ultra-short-term PV power forecasting by leveraging multiple weather scenario identification and incremental learning techniques. The method identifies complex weather conditions using a improved K-means clustering method and Local Outlier Factor (LOF) algorithm. The approach then updates the forecasting model incrementally using parameter freezing, enhancing its ability to adapt to newly identified weather patterns. The proposed method is tested on real-world data, demonstrating improvements in forecasting accuracy under complex weather conditions while maintaining performance for more typical weather scenarios. Experimental results show that incremental learning significantly enhances the model’s predictive capabilities for ultra-short-term PV power forecasting.