A double-layer forecasting model for PV power forecasting based on GRU-Informer-SVR and Blending ensemble learning framework

With the advancement of power market reform, the proportion of new energy sources participating in power market transactions has been increasing. Photovoltaic (PV) power generation is characterized by intermittency and volatility, which brings risks to power system operation. To improve the prediction accuracy of PV power, we propose a double-layer prediction model of GRU-Informer and SVR based on the Blending ensemble learning framework that considers feature screening and weather clustering. First, XGBoost is used to calculate the importance of weather features and filter the features. Second, K-means clustering is used to classify the weather data into three types: sunny, cloudy and mutation. Third, a GRU-Informer and SVR double-layer prediction model that was constructed based on the Blending ensemble model is used. The first-layer base-learner uses the training set to train the GRU with Informer and outputs the first-layer prediction results. The second-layer metalearner uses the first-layer prediction results to train the SVR and generate the final prediction results. The empirical analysis results show that the proposed model can achieve higher fitting accuracy R2 of more than 98 % under all three weather conditions and is a promising approach in terms of PV power generation forecasting.