Prediction Of Photovoltaic Power Generation Based On Density Peak Clustering And Cloud Analysis

This subject is part of the sub-topic of the Shanxi Electric Power Company's science and technology project “Research on Key Technologies for Planning and Control of Wide Area Distributed Power Based on Big Data Analysis”.With the continuous expansion of China's photovoltaic industry,the impact of photovoltaic system power fluctuations on the power grid should not be ignored in the future,what's more,the precise PV power prediction has become the current research hotspot.This thesis proposes a short-term and ultra-short-term power forecasting method based on data mining techniques,taps potential information from low-value-density data and provides decision support for short-term and ultra-short-term power forecasting of distributed photovoltaic systems.At the same time,the method takes photos of clouds over a distributed photovoltaic system and uses image processing techniques to identify and predict cloud features,further improved the accuracy of PV power prediction.First of all,this thesis analyzed the common methods of photovoltaic power generation forecasting at the present stage,and clarified the idea of power forecasting using similar historical data clustering methods.By analyzing the traditional clustering algorithm and pointing out its existing defects,an improved algorithm based on density peak clustering is proposed.The algorithm achieves global diffusion and local convergence of data sets by means of optimization between classes and optimization within the class,and enhances the separability of photovoltaic historical data.The principle of the two optimizations is based on the histogram equalization idea and the Sigmoid function respectively.The degree of optimization is derived from the characteristics of the data set itself,and the optimization can be kept within a relatively reasonable range.Therefore,it has certain self-healing optimization capabilities.By comparing the two-dimensional MDS graphs and distance histograms of the data in different stages of optimization,the results show that the improved clustering algorithm can effectively improve the separability of photovoltaic historical data,which has laid a good foundation for the cluster analysis of similar historical data.Secondly,based on the above improved algorithm,a short-term power forecasting method based on CFSFDP clustering is proposed.The performance of the density peak algorithm before and after improvement was tested by three cluster validity indicators,and the stability test under parameter perturbation was performed to ensure that the improved algorithm can display good clustering performance and good robustness in photovoltaic historical data.Then use the grey correlation analysis method to establish a matching relationship between the date to be predicted and the clustering result,and find the training data that is most suitable for the day to be predicted.The short-term prediction model was established by using Elman neural network and the short-term prediction results were obtained after neural network training.Practical application analysis shows that the prediction results obtained by optimizing clustering can effectively reduce the prediction error and achieve a better short-term prediction effect.Finally,on the basis of the above CFSFDP clustering results,an ultra-short-term power forecasting method based on real-time cloud layer analysis is proposed.The focus of data mining is on cloud image analysis and terrestrial radiometric prediction.By collecting the cloud image above the photovoltaic site,image processing technology is used to mine the image features that affect ground radiation.Taking full consideration of the combined effects of cloud image characteristics,radiation outside the atmosphere,and air quality,an ultra-short-term radiation prediction model was constructed using Elman neural networks.At the same time,the method makes full use of the clustering results of CFSFDP to fit the non-linear conversion relationship between radiometric power and power,and then obtains the ultra-short-term prediction results of photovoltaic power generation.Practical application analysis shows that the forecast results obtained through real-time cloud layer analysis can effectively predict the power fluctuation phenomenon that is obvious on the day and achieve a better ultra-short-term prediction effect.This subject has passed the acceptance of science and technology projects of Shanxi Electric Power Company,which can provide a good application foundation for practical projects.