| During the production and operation process of photovoltaic power plants,it is necessary to constantly collect power generation data to monitor the operation status of the power plant.However,in the actual production process,there are often two types of abnormal situations:data missing and data mutation,which bring great difficulties to the prediction of photovoltaic power and the formulation of power generation plans for power stations.Moreover,a large and accurate real-time output data is of crucial importance for power grid scheduling and other work,so it is necessary to study the abnormal power data of photovoltaic power plants.This article focuses on the problem of extracting similar days.By analyzing the regional differences and changes in the actual power generation data of photovoltaic power plants,it is determined that photovoltaic output data has time-varying characteristics and climate correlation,which are combined with the vertical and horizontal characteristics of photovoltaic output data.A similar day screening method based on Pearson correlation analysis combined with seasonal changes is proposed.Through practical cases,it has been proven that this method can effectively and quickly screen out several similar days with high similarity between the photovoltaic output curve and the day to be repaired.To address the issue of identifying and cleaning abnormal data in photovoltaic power plants,a density based clustering method with noise is used to identify and clean data mutation anomalies.After data cleaning,the two types of data anomalies in the power plant output data are reduced to data gaps,reducing the complexity of repair work.The feasibility of this method was demonstrated by comparing the cleaning results with the K-means partitioning clustering algorithm.Aiming at the problem of abnormal data repair in photovoltaic power stations,a generative countermeasure network data completion algorithm is proposed in the field of photovoltaic data repair to complete and repair data,and the operation effect is compared with the short-term memory model combined with similar days.The results show that the normalized generative adversarial network data completion repair method has a root mean square difference of less than 0.18 between the repaired data and the actual data,which is in line with experimental expectations.The repair results under the whole year time span reached the experimental expectations,and the comprehensive repair effect of the generative warfare network data completion algorithm was longer than that of the short-term memory model algorithm,which increased by 2.3%.The experimental operation results proved the effectiveness and superiority of this method.In summary,through the above research,similar day screening methods combining seasonal changes,abnormal data recognition and cleaning methods based on density clustering algorithm,and abnormal data repair methods for photovoltaic power plants based on generative adversarial networks have been determined.All achieved the expected results,providing a sound data foundation for production needs such as photovoltaic power prediction and power station production arrangement. |
