Research On Fault Location For Active Distribution Network Based On Tensorflow Deep Learning

With the gradual maturity of concepts such as environmental protection and low carbon,society’s requirements for the use of renewable energy have gradually increased.Distributed generation technology has developed rapidly due to its advantages of generating electricity mainly by using renewable energy such as wind energy and photovoltaic energy.A single large grid system connected to distributed generation can effectively save costs,improve energy efficiency,and ensure the safety and reliability of the system.However,a large number of access to distributed power sources will cause the original topology of the distribution network to change,which will change the direction of the system’s power flow.This will bring difficulties to the traditional distribution network fault location method.With the development of computer technology,some data-driven fault location methods have been proposed gradually,and exploring fault location method of the active data distribution network using big data method has become a new direction.The dissertation has carried out a study of applying deep learning ideas to fault location of active distribution networks.The main research contents include:(1)Considering the advantages of deep learning in identifying high-dimensional data features,it is applicable to the research of active distribution network fault location methods.Firstly,the reasons why the access of distributed generation restricts the application of the traditional fault location method are pointed out,and the limitations of the traditional method in some scenarios are pointed out.Secondly,an active distribution network model was built in Simulink,and the rationality of selecting the second harmonic electrical feature as the object of deep learning training was analyzed through simulation,Characteristic simulations of short-circuit faults at different locations were conducted to verify that when asymmetric short-circuit faults occur at different locations in the system,the amplitude,phase,and waveform characteristics of the sequence components produced by them are significantly different from each other.Finally,the adaptability of deep learning algorithms in learning differentiated features is analyzed.(2)Aiming at the fault location in the single-fault scenario,a deep neural network-based active distribution network fault location method is studied.Firstly,it introduces the Tensor Flow framework and the deep learning model structure under this framework in detail and by combining complex network theory,a reasonable measurement point allocation strategy is proposed.Secondly a deep neural network model based on fully connected network(FCN)and using Dropout is constructed to realize fault location.Finally,experiments show that the method can accurately locate faults with a small number of measuring points,it can be used as an effective supplement to the traditional active distribution network fault location method in the scenario of fewer measurement points.(3)Aiming at the fault location in multiple fault scenarios,by studying the structure and principle of long-short term memory(LSTM)networks,a fault-tolerant fault location method for active distribution networks based on LSTM is proposed.By training the time series of the current sequence components and the voltage sequence components before and after the fault which is collected by the FTU at both ends of the line,a dual LSTM network is formed,and logic gate rules are formulated for the neuron output of the dual network as a basis for judging whether a short-circuit fault occurs in this section.Utilizing the structural characteristics of the LSTM network,fast and accurate mining of the dual-end timing characteristics of the segment is achieved,it can locate multiple faults,which can be used as an effective supplement to the traditional active distribution network fault location method in the full measurement point scenario.(4)For the fault phase selection problem,based on mature long-short term memory network model,LSTM feature extraction network parameters trained as a historical task to solve the fault location problem are taken as knowledge transfer,thereby realizing the fault type selection task.By rationally using the model parameters of the hisrorical task,the learning efficiency of the fault type selection task is improved,and the simulation does not need to be traversed repeatedly,which reduces time of the data preprocessing and the model convergence.The results of calculation examples verify that the method can quickly identify the fault type and realize fault type selection.