Digital Twin Model of Photovoltaic Power Generation Prediction Based on LSTM and Transfer Learning

This paper proposes a digital twin model based on the long-term & short-term memory network (LSTM) for the photovoltaic power generation prediction. This model is applied in the power generation prediction of other photovoltaic systems that have a short operating time and insufficient data through transfer learning. Due to the influences of solar irradiance, temperature and other random factors, the photovoltaic power generation has strong intermittent and volatility, which is difficult to make accurate photovoltaic power prediction. The proposed digital twin model realizes the synchronization and real-time update with the physical entity of the photovoltaic system, thus obtaining more accurate prediction results than the traditional prediction methods. At the same time, the knowledge learned from the photovoltaic systems with sufficient historical data is used to assist the photovoltaic systems with limited historical data to establish a power generation power prediction digital twin model, which can not only obtain the accurate prediction results but also save the model training time. This paper verifies the effectiveness of the proposed method through the historical photovoltaic data from three different sites on the Queensland University open source website and the Jinneng Clean Energy Company, Shanxi province. © 2022, Power System Technology Press. All right reserved.