Research On 10kV Feeder Fault Prediction Method Based On Deep Learning

With the continuous improvement of the level of industrialization,informatization and intelligence in modern society,the electric energy,as an important part of secondary energy,is not only irreplaceable in people's daily lives,but also plays a key role in social development.At present,the daily undifferentiated patrol and passive maintenance management of distribution network can't further meet the people's growing demand for power supply reliability.In order to achieve the goal of active maintenance and reduce the occurrence of fault,the approach how to discover the potential risk early and predict fault in the distribution network,has become a hot and difficult research topic.In recent years,with the rapid development of smart grid technology,the research of power grid cyber-physical systems?CPS?is also deepening.Many distribution network information management systems have been put into use.While these systems bring convenience to the management of distribution network,they also accumulate a large amount of data,forming large data of distribution network with"5V"characteristics.How to apply big data technology to the big data of distribution network and realize effective prediction of fault has also become a major research trend of fault prediction technology in smart distribution network.Experts and scholars at home and abroad have done a lot of research on these problems and achieved a lot of results,but the research is still immature,and some algorithm theories are also inadequate,there is still a gap between the actual engineering applications.Therefore,this paper studies the fault prediction method of distribution network based on deep learning,and predicts the probability of 10kV feeder fault(meteorological factors fault^?1?,equipment factors fault^?2?and operation factors fault^?3?)according to the actual demand of the power supply company,and proposes a feeder inspection sorting algorithm based on fault prediction results.Firstly,this paper makes a thorough analysis of the causes,types and data sources of the distribution network faults,and conducts a field investigation on the distribution network information management system of a power supply company,and obtains the data needed for the fault prediction of the 10kV feeder.Then,according to the defects and characteristics of the original data of distribution network,the effective methods such as Jieba participle,negative sample generation and sample equalization are used to process the data,and the processed data are cleaned to improve the data quality.Secondly,this paper studies the inherent attribute characteristics,statistical analysis features and depth time series characteristics of distribution network faults,analyzes the relationship between each feature and the feeder fault rate,and determines the strong correlation fault characteristic variables.In order to quantify the cumulative effect of time series variables on distribution network feeder,this paper constructs“fault pictures”based on time series data according to different time steps,and uses CNN algorithm to extract time series variable features deeply.Based on stacking idea,the output of the CNN network is input to the fault prediction model as part of the feeder fault feature vector.Thus,the fault characteristics can be further enriched and the effect of fault prediction model can be improved.Thirdly,combined with the big data background and practical application requirements of distribution network,this paper proposes a fault prediction method for10kV feeder in distribution network.The method uses the LightGBM algorithm with big data processing capability to construct the distribution network fault prediction model,and uses the grid search algorithm to optimize the parameters of multiple LightGBM sub-models in the prediction model.The experiments prove that the model has high accuracy and can effectively predict the probability of faults with three typical causes in the distribution network.At the same time,the prediction results of this algorithm and random forest algorithm are compared and analyzed,and the comparative experiments of adding depth time series features or not are also carried out.The results show that the accuracy of LightGBM prediction model with depth time series features is higher,which further proves the validity of the depth time series features and the rationality of using this method to construct the feeder fault prediction model.Finally,based on the fault prediction results of 10kV feeder,considering the constraints of the power supply area of the feeder,the number of important users of the feeder,the average number of hour-households per day when power failure and the distance between the feeder and the power supply bureau,so as to minimize the impact degree of faults and maximize the operation and maintenance value,this paper studies the sorting algorithm of distribution network feeder inspection.The method uses the entropy weight method to calculate the influence weight of other constraints on the feeder inspection priority.And combined with the weight,designs the quantitative formula of feeder inspection priority based on the feeder fault probability.The experimental results show that the proposed sorting algorithm is more reasonable and practical than the simple fault probability sorting algorithm,which can provide assistant suggestions for operation and maintenance workers.