Novel PV Power Hybrid Prediction Model Based on FL Co-Training Method

Existing photovoltaic (PV) power prediction methods suffer from insufficient data samples, poor model generalization ability, and the inability to share data. In this paper, a hybrid based on federated learning (FL) is proposed. To improve communication efficiency FL introduced combine multiple locations without sharing collaboratively train model. Furthermore, LSTM-BPNN designed accuracy of predictions. LSTM used extract important features time-series data, BPNN maps extracted high-dimensional low-dimensional space outputs predicted values. Experiments show that minimum MAPE constructed in paper can reach 1.2%, effect improved by 30% compared with traditional Under mode, trained not only improves more than 20% but also has excellent ability scenarios.