A short-term PV power forecasting method based on weather type credibility prediction and multi-model dynamic combination

Accurate short-term photovoltaic (PV) power forecasting results can provide solid supports for power gird management. The PV power generation exhibits different daily output patterns in short-term time scale, which are closely related to daily weather types. The existing research proves classifying weather types appropriately and classification modeling for each weather type is an effective approach to improve the accuracy, comparing to unified modeling. However, this classification modeling framework is a double-edged sword, with the prerequisite for its effective works is the accurate prediction of weather types in day-ahead. Once the weather type is misjudged, the wrong application of power forecasting models will lead to a decrease in forecasting accuracy, but existing research has mostly ignored this issue. To this end, this paper proposes a short-term PV power forecasting method based on weather type credibility prediction and multi-model dynamic combination. Firstly, a weather type identification model is established by extracting the measured irradiance features, to identify historical days with unknown weather types. Secondly, one-hot encoding is employed to represent the probability information within the original weather type, then an attention mechanism based deep learning model is constructed by extracting features from day-ahead acquired numerical weather prediction (NWP) to achieve the prediction of weather type credibility. Finally, by establishing a credibility level (CL) optimization mechanism, the unified forecasting model based on Transformer and classification forecasting model based on double Q learning are established for unreliable and credible scenes respectively. Simulation results show that the accuracy of the proposed method is improved by 4.46% and 2.79% respectively compared with the models that do not introduce classification modeling and do not consider the credibility of weather type prediction results.