Regional photovoltaic power prediction based on multi-agent federated learning and deep learning

Accurately predicting regional photovoltaic power is a crucial foundation for integrating photovoltaic power generation into the grid. This paper introduces a regional photovoltaic power prediction method based on a Convolutional Long Short-Term Memory Network (CNN-LSTM) model by using satellite cloud imagery within the framework of federated learning (FL). Firstly, the CNN-LSTM model is established to predict power for individual photovoltaic stations within the region. Secondly, within a multi-agent federated learning framework, the prediction model is trained locally, and encrypted training gradient parameters are uploaded to a central server, where aggregation computations occur before returning the parameters to local terminals iteratively, leading to the final regional photovoltaic power prediction model. Lastly, based on the semantic segmentation technique of DeepLabv3+deep learning, the satellite cloud image is processed to obtain correction factors of the cloud cluster on the regional photovoltaic power station, and the prediction results are corrected. Experimental results demonstrate that the proposed model achieves highly accurate ultra-short-term predictions of regional photovoltaic power while considering data privacy and cloud cover obstructions.