Ultra-Short-Term Photovoltaic Power Forecasting Based on Transformer Model

With the rapid depletion of traditional fossil energy and the increasing global environmental crisis, sustainable development has become the theme of today’s era. In order to ensure energy supply and control the degree of pollution, governments of various countries are making every effort to develop renewable energy. Among the new energy sources that are clean and renewable, the largest amount is solar energy. But its randomness and intermittent nature greatly limit the process of large-scale development and utilization. This requires accurate and effective prediction of photovoltaic power generation to reduce power system security issues caused by large-scale grid-connected photovoltaic power plants. In this paper, a novel transformer-based deep learning model is established for ultra-short-term photovoltaic power generation forecasting. In this study, the data set has a 15min interval for sampling, and the Transformer model is evaluated against the convolutional neural network(CNN) model and the long short-term memory(LSTM) model based on MSE, MAE, RMSE and R
²
. The experiment was analyzed from three perspectives: overall, sunny and cloudy. In the overall evaluation, the CNN model achieves an accuracy of 97.8%, the LSTM model achieves an accuracy of 97.9%, and the Transformer model achieves an accuracy of 98.1%. In the sunny evaluation, the CNN model reaches an accuracy of 98.5%, the LSTM model reaches an accuracy of 98.6%, and the Transformer model reaches an accuracy of 98.9%. In the cloudy evaluation, the CNN model attains an accuracy of 91.5%, the LSTM model attains an accuracy of 91.7%, and the Transformer model attains an accuracy of 92.1%. The prediction results indicate that the Transformer model surpasses other models.