Research On Photovoltaic Data Anomaly Repair And Output Prediction Method Based On Deep Learning

With the rapid depletion of traditional fossil energy,the global environmental crisis is getting worse,and sustainable development has become the theme of the times.In order to ensure the energy supply and control the degree of pollution,the governments of various countries try their best to develop renewable energy.Among the new energy sources with clean and renewable characteristics,the largest quantity is solar energy.However,its randomness and intermittence greatly limit the process of large-scale development and utilization.Therefore,it is necessary to predict the photovoltaic power accurately and effectively,so as to reduce the power system security problems caused by the large-scale grid connection of photovoltaic power plants.At present,the short-term and ultra-short-term photovoltaic output prediction models need to be improved mainly in the prediction accuracy and the fit with engineering practice.In view of the above problems,this paper studies and analyzes from two aspects: bad data identification and repair,and prediction model optimization.Based on the physical modeling of power generation principle and the mathematical statistics of historical data,the correlation degree between photovoltaic power generation and weather and the distribution characteristics of daily power generation curve are analyzed.The original power data is divided into missing anomaly and singular anomaly,and the bad data is identified based on the analysis results.According to the missing abnormal data,nnz function is used to screen;According to other complete data,referring to the power output values of neighboring power stations with regional correlation and the irradiance obtained from the detection of their own power stations,the calculation of curve similarity function considering dynamic distribution characteristics is completed,the dates containing abnormal data are distinguished according to the distribution of function values,and the abnormal periods are screened out by comparing the clustering of reference curves.Aiming at these two kinds of abnormal data,this paper constructs a bad data repair model based on PSO-RBF algorithm.Compared with other algorithms,the results prove the superiority of this way.In the ultra-short period,due to the short time scale,the error is less influenced by external factors,and only historical time series is used as input.The photovoltaic prediction model based on ARIMA-RNN is constructed with the abnormal repair results as input.This part first analyzes the advantages and disadvantages of ARIMA algorithm in the field of time series prediction.Considering the disadvantage that the prediction accuracy will decrease when there is large fluctuation,the basic cyclic neural network is introduced to improve it,and the optimization principle of RNN is deduced mathematically.Combined with practical examples,it can be seen that the prediction effect is better.Short-term forecasting often faces two kinds of problems.First,it is difficult to obtain the required meteorological data in practical engineering,such as radiation angle,PM2.5 concentration near the station and other parameters,which reduces the practical applicability.Or it is necessary to predict real-time cloud cover,temperature and other change curves based on the existing meteorological data,so that error superposition occurs.In order to solve the above two dilemmas,five short-term photovoltaic output forecasting models based on deep learning are constructed in this paper.The advantages and disadvantages are analyzed through experiments,and the software platform is developed with the best method GRU-attention,which is put into use and fits the engineering practice,and the forecasting effect is good.Finally,the research results of the full text are summarized,and the development prospect of the follow-up photovoltaic output prediction research is put forward in view of the existing shortcomings.