Solar Irradiance Forecasting Utilizing Distributed and Federated Learning

Photovoltaic energy generation and prediction are crucial for the integration between solar plants and smart grid. Accurate solar radiation prediction is one of the most important research topics for solar energy generation. Existing deep learning methods require sufficient data for training for achieving the desired results. In traditional deep learning modeling, the data sets required for model training usually come from various data in different regions and organizations, and it is difficult to ensure data security and privacy protection. Federated learning can effectively utilize the data optimization model while ensuring data security and privacy, but there is still the risk of privacy leakage by analyzing the parameters uploaded by the client. To effectively utilize data optimization models while ensuring data security and privacy, we propose a method based on a federated learning framework and N-BEATS to achieve interpretable time series prediction of solar irradiance. The model builds the N-BEATS model on the solar irradiance datasets of multiple institutions in a distributed method, does not share local data to protect institutional data privacy, the concept of differential privacy (DP) is introduced, Gaussian noise is added to the parameters of the client before the server-side aggregation processing, the model parameters of each distributed node are collected through TCP communication for federated average, and the parameters are returned to each distributed node. The experimental results show that compared with other deep learning models, the model has better prediction accuracy under the premise of ensuring data security and privacy.