RESIDUAL DEEP LEARNING MODEL FOR PHOTOVOLTAIC POWER PREDICTION

To ensure that the photovoltaic power prediction model has high accuracy when the meteorological abrupt changes, quantifying the weather abruptly by residuals is proposed and constructed as a new feature. After applying the maximum information coefficient (MIC) to eliminate the irrelevant meteorological features, the XGBoost model was introduced to obtain the residual series. Using the autocorrelation of the residuals, the residuals of the previous moment were used as the new features of the current moment to construct a deep learning model of residuals for photovoltaic power prediction. The experimental results show that the proposed model can achieve higher accuracy under sudden meteorological changes.