Research On Extrapolation Of Ground-Based Cloud Image And Photovoltaic Power Prediction Method

Ground-based cloud image plays an important role in many fields,such as cloud cover monitoring,air pollution research,and photovoltaic power prediction systems.The prediction of sky cloud conditions in the future can help researchers make decisions in advance at the critical moment,so as to obtain higher equipment operation efficiency and avoid unnecessary losses.For example,in the ultra-shortterm photovoltaic power prediction system,the sky cloud condition is one of the most important factors affecting the output power of solar photovoltaic panels.Better forecasting of the photovoltaic output power in the future can improve the stability of the grid,help photovoltaic power plants make energy scheduling plans,and improve their operation and management efficiency.Traditional cloud image extrapolation methods based on digital image processing do not have a good solution for continuous predictions that include cloud contour changes.Therefore,from the perspective of image sequence prediction,this paper proposed a cloud image extrapolation method based on deep learning,and built an ultra-short-term photovoltaic power prediction model,which verifies the effectiveness of the cloud image extrapolation model in this paper.The specific research contents of this paper are as follows:(1)This paper draws on the Pred RNN++ model that performs well in image sequence prediction,and proposed the CCLSTM model for the more complex,changeable,and unpredictable characteristics of cloud motion.The proposed scheme strengthened the unit depth and gradient transmission of the model,and improved its fitting ability,and realized the colorful and continuous cloud motion prediction that includes cloud contour changes.(2)Since the training of the image sequence prediction model requires high computer resources,this paper proposed down-sampling operations on the cloud images before inputting them into the CCLSTM model.Such prediction results are difficult to meet the requirements of resolution and clarity in subsequent research.Therefore,this paper designed the Ground-Based cloud image Super-resolution model,added the perceptual loss on the pre-trained Res Net50,reconstructed the output of CCLSTM,effectively restored the cloud image distortion caused by down sampling and extrapolation,and provided a strong support for the follow-up research of extrapolation cloud images.(3)This paper proposed an ultra-short-term photovoltaic power prediction model based on cloud image extrapolation and 3D-CNN.The input information of the model includes historical cloud image sequence,extrapolated cloud image sequence,current cloud image frame,photovoltaic historical data and manually extracted features.Among them,historical and extrapolated cloud image sequences are coded as feature vectors by 3D-CNN,merged with historical photovoltaic data and manual features,and then supervised and regressed through multi-layer MLP.a model without cloud image extrapolation sequence is compared with the above model in the same data set.The results show that the cloud image extrapolation method proposed in this paper can effectively improve the accuracy of photovoltaic power prediction.