Short-Term PV Output Prediction Method Based on Feature Analysis and Multi-model Fusion

Abstract

In the future, photovoltaic power generation will usher in a larger market. The rapid development of artificial intelligence provides a new solution for photovoltaic power generation forecasting. In this paper, combined with the current cutting-edge theoretical research in the field of artificial intelligence, a short-term photovoltaic power generation power prediction method based on multi-model fusion Stacking ensemble learning is proposed. Random Forest and correlation coefficient feature importance analysis were used to determine the important climate factors as the input characteristics of the prediction model. On this basis, multiple machine learning models with good single prediction performance and certain differences are integrated into the Stacking ensemble learning photovoltaic output prediction model. The base learner of the model includes XGBoost tree ensemble algorithm and GRU neural network algorithm. To prevent overfitting, the meta-learner consists of the LSSVM algorithm with relatively simple complexity and high accuracy. The calculation example uses the photovoltaic power and climate data provided by the Australian Solar Energy Research and Development Center to verify the effectiveness of the algorithm. The prediction results show that the Stacking model has higher prediction accuracy than the traditional single model which can better track the fluctuation of output power.