A Short-term Photovoltaic Output Power Forecasting Method by Integrating Physical and Data-driven Models

Photovoltaic power generation efficiency is heavily influenced by factors like weather and environment, leading to poor stability and low prediction accuracy of its power generation through traditional mechanism models. Thus, in this paper, we proposed a method to integrate traditional photovoltaic output power prediction models with machine learning models for short-term photovoltaic output power forecasting. During the compensation stage, we first applied physical models to calculate the generation power and compensate the power value with gated recurrent unit (GRU) neural networks. Simultaneously, a purely data-driven model is trained to map actual generation power through condition data, environment data, calibration data. Then, the final forecasted generation power can be obtained through the theory and mapping values according to the weight coefficients which can be optimized with the recorded actual power. We also applied other data-driven models to compare the accuracy and effectiveness. Our results demonstrate that the proposed method can effectively raise the forecasting accuracy, and decrease the instability due to the aging of equipment over time.