Distributed Photovoltaic Power Prediction With Limited Data: A Combined Ensemble Learning and Transfer Learning Approach

The intermittent and stochastic nature of photovoltaic (PV) power generation poses significant challenges to the operation of modern power systems, necessitating precise predictions for efficient integration of renewable energy. This paper addresses the difficulty in training deep learning models for PV power stations in regions with limited sample data by proposing a novel few-sample PV power generation prediction method. The approach combines ensemble learning and migration learning, which adopts the MLP-LightGBM-LSTM adaptive fusion integration model training for the power stations with complete power generation data in the source domain to improve the accuracy of the power station power generation model in the source domain. subsequently transferring the trained model to less-sampled PV power stations in the target domain based on optimized weights. This migration improves prediction accuracy for power stations with limited data in the target domain. Experimental validation demonstrates the effectiveness of the proposed method, showcasing a 13.56 %-14.39 % reduction in prediction errors for individual power stations and a 1.95 % 4.09 % decrease in prediction errors for overall regional power generation compared to alternative methods.