Intelligent Fault Diagnosis Based On Massive Data

With the development of new energy technology and demand,photovoltaic power generation is playing an increasingly important role in energy supply.Normal,stable and economic operation is the basic requirement of photovoltaic power stations.Therefore,the fault diagnosis of photovoltaic system is of great significance to maintain the safety,reliability and efficiency of photovoltaic power station.Nonlinear output characteristics and easy to be affected by operating environment are the characteristics of photovoltaic power generation.Machine learning fault diagnosis methods to solve these problems are relatively rich.However,there are still some problems:the number of short-term samples is too small,the samples are difficult to obtain in the field and the interference information is too much,the diagnosis accuracy rate has room to improve,there is no effective prediction for the fault data in the future period,and the accuracy,prediction efficiency and time cost of the fault prediction model have room to optimize.This paper studies the validity verification,fault diagnosis,prediction trend of characteristic parameters and fault identification of photovoltaic arrays.The main contents of this paper are as follows:First,the paper first through the Simulink simulation system for PV array modeling,comparing the physical PV cells and array parameters specifications,modify the parameters,demonstrate the usability of the simulation PV array,verify the P-V,I-V and temperature,irradiance and other characteristics of each condition.Twice,put forward based on Fuzzy C-Means clustering(FCM)and D-S evidence theory fusion method,segmentation simulation data samples,through the number of different clustering functions and fuzzy weight value modification,find the best value,fault classification of mixed samples,fault diagnosis.Third,Proposes a fault prediction model combining GRU(Gate Recurrent Unit)and Temporal Convolutional Network(TCN).Two models were used to calculate the input,connect different output vectors,and reduce dimension to form a new output,so as to predict the characteristic parameters.The performance of the short and long term characteristic parameter prediction was evaluated respectively.The fault types were determined according to the changing trend of the characteristic parameters.The performance of the evaluation indexes was compared comprehensively with that of the other four prediction models.Fourth,the attention mechanism is introduced to improve the GRU-TCN model,connect the residual block of TCN,focus the key feature information again,and send it to the GRU-TCN network for future short-term working condition data prediction.Simulation results show that this method can achieve high fault recognition rate and prediction accuracy in massive data.Compared with the improved diagnostic method,the proposed model can significantly improve the accuracy of classification and prediction of training and test samples,and reduce the time cost.It has significant advantages in diagnostic accuracy and generalization ability.