Research On Photovoltaic Power Forecasting Method Based On Ground-based Cloud Image And Deep Learning

In the context of the rapid development of global renewable energy,the installed capacity and development scale of photovoltaic power generation have been expanding.However,PV power generation is susceptible to climatic and environmental factors,showing intermittency,randomness and fluctuation.Accurate and rapid prediction of PV power is very important for safe and stable operation of power grid and PV power development and consumption.Therefore,this paper uses image processing technology and deep learning technology based on ground-based cloud image data and PV power data to achieve accurate short-term PV power prediction.First,a cloud identification method based on solar occlusion is proposed for ground-based cloud image feature extraction.Based on the shading situation between the sun and clouds and the color feature distribution of the ground-based cloud image,the ground-based cloud image is divided into four weather conditions:sunny,partly cloudy,cloudy,and overcast.Meanwhile,the cloud amount and the average pixel intensity of the sun region of the ground-based cloud image are extracted using the cloud identification methods corresponding to different weather conditions.The results show that the cloud identification method based on solar shading outperforms the clear sky library method and the fixed threshold method,and can accurately identify cloud regions and extract cloud amount features.Secondly,for the significant influence of clouds on PV power generation,a ground-based cloud image and LSTM-based PV power prediction model is proposed.Firstly,the cloud identification method based on solar shading is used to extract the ground-based cloud image features,and the PV power history values with significant correlation are selected by autocorrelation analysis,and they are used as the input of the LSTM prediction model,and finally the prediction results of the LSTM model are compared with those of the persistence model and the support vector regression model.The results show that the input of ground-based cloud features is beneficial to improve the prediction accuracy,and the LSTM model has the best prediction performance under different weather conditions,and the prediction error is reduced by 3%compared with the persistence model.Finally,the LSTNet-based PV power prediction model is proposed for different time scales of PV power prediction.Through linear correlation analysis and autocorrelation analysis,meteorological features and historical values of PV power with high correlation with the prediction target values are selected as the input of the LSTNet prediction model,and the local dependence between PV power and multivariate meteorological features and long-term dependence of PV power time series are extracted using the LSTNet network,and compared with the persistence model,support vector regression model,and LSTM model and compare the results with those of the persistence model,support vector regression model,and LSTM model.The results show that the LSTNet model has a goodness-of-fit greater than 0.98 at different time scales,which proves the reliability and validity of the LSTNet model.