摘要
Federated learning(FL)is a promising technique to build a power generation model in photovoltaic(PV)scenarios.However,due to the heterogeneity of power generation data,there is a problem of slow convergence of the global model or even deviation from the optimal solution during model training.Therefore,to improve the prediction accuracy and accelerate the model convergence speed,this paper proposes a model functional blocking and differentiated scheduling mechanism under personalized FL framework for intermittent PV power generation.Firstly,cluster analysis is conducted according to longitude,latitude,and altitude to form a model collaborative training region(MCTR).Then,based on the constructed MCTRs,a personalized FL model training method is proposed.This method is based on a combination of global shared convolutional neural network(CNN)model and local personalized long short term memory(LSTM)model,where CNN model block is responsible for feature extraction and LSTM model block is responsible for prediction.It adopts synchronous aggregation for global shared CNN and asynchronous aggregation for personalized LSTM.Furthermore,the FL server performs block scheduling of the CNN-LSTM models and aggregates them based on the regional membership which can provide differentiated and accurate prediction models with different power generation patterns.The simulation results show that the proposed algorithm has the highest accuracy of 85.1%and the best performance on mean absolute error(MAE),root mean square error(RMSE)and mean absolute percentage error(MAPE),with 0.1105,0.1224 and 0.4383 respectively.
基金
supported by the Science and Technology Project of State Grid Corporation of China(5108-202218280A-2-394-XG)。
