Research on stacking ensemble method for day-ahead ultra-short-term prediction of photovoltaic power

The intermittency and uncertainty of photovoltaic power output make it difficult to match real-time electricity demand, and accurate prediction of photovoltaic power is the basis for improving photovoltaic energy consumption. This article proposes a new stacking ensemble forecast model based on the time series generative adversarial network (TimeGAN-SEFM), to improve the accuracy of ultra-short-term forecasting of photovoltaic power. Firstly, a dual layer feature weighted clustering weighted fuzzy C-means clustering algorithm is proposed to accurately partition the original dataset according to weather types. Secondly, in the proposed TimeGAN-SEFM, a Particle Swarm Optimization based BP neural network is used as a meta learner to integrate three types of base learners, thereby improving the structural diversity of the model. The dataset for each type of weather is proportionally divided into multiple sub-datasets for training base learners. In response to the limitation that the amount of data used for training a single base learner is relatively small after multiple splitting of the original data, a TimeGAN augmentation network is introduced to perform multi-dimensional data augmentation on the sub-training sets, which improves the size and diversity of the dataset used for training base learners, thereby improving the predictive performance of the model. Finally, the predictive performance of the proposed TimeGAN-SEFM was evaluated using measurement data from a 20 MW photovoltaic power station in North China. The results show that TimeGAN-SEFM effectively improves the accuracy of photovoltaic prediction and helps promote real-time matching of photovoltaic energy and power demand, thereby improving the utilization of photovoltaic energy.