Transfer Learning Based Power Generation Prediction and Intelligent Diagnosis for Photovoltaic Power Plant

Solar energy is an important renewable energy for our daily life. Recent years we have witnessed the development of photovoltaic (PV) power stations. Accurate forecasting of PV power generation is crucial for scheduling and maintaining the PV stations. The limited PV data presents a significant challenge in developing effective predictive models for newly-constructed PV stations. A transfer learning method is integrated with BP neural network (BPNN) to predict the power generation with a small amount of data. A BPNN pre-train model is trained by using dataset with about 7.4 million samples. The pre-train model is transferred to the newly-constructed PV stations by proposing a transfer learning techniques, constructing a transfer learning BP (TL-BP) network for PV power generation prediction. Based on the residual error between the real generation and predicted generation, a diagnosis method is presented to evaluate the status of the PV power plant. Experimental comparison studies are given to show that the proposed TL-BP network can improve the prediction accuracy with less training data.