Method of Fault Self-Healing in Distribution Network and Deep Learning Under Cloud Edge Architecture

.


INTRoDUCTIoN
As a key component in smart grid construction, the distribution network is directly connected to power users, and its safe and stable operation is closely related to daily life, production, and the lives of the power users.Even a very brief power supply interruption will adversely affect the power supply service and commitment of the power supply department (Wang & Wang, 2015).At the same time, as the global energy supply is shifting toward cleanness, low carbon, high efficiency, and electrification, great advances have been made in distributed generation that involve more eco-friendly technologies, and the wide applications of these technologies have made the distribution network, originally radiated by single power sources, increasingly huge and complex (Cavalcante et al., 2015).Adapting to the development of the future network on the basis of the existing network and making the feeder of the distribution network locate quickly after tripping are pressing issues for power supply enterprises and power practitioners (Srivastava et al., 2012;Zidan & El-Saadany, 2012;Cavalcante et al., 2015;Leite & Mantovani, 2016).Therefore, power supply enterprises continuously enhance the reliability of grid structure construction and power supply, and they also strive to improve the automation and intelligence level of distribution networks.In recent years, with the gradual popularization of the SCADA system and with power grid companies at all levels increasing the automation transformation and "three remote" upgrading of distribution network equipment, the distribution network has basically realized the monitoring of operating parameters, the collection and uploading of fault information, and the remote operation of switches (Wang & Wang, 2015).After comprehensively analyzing the fault information, protection action information, and switch status information fed back from the main station, the dispatcher orderly guided the onsite maintenance personnel to carry out targeted fault patrol, which greatly improved the power recovery efficiency and shortened the power supply recovery time (Li et al., 2020;Refaat et al., 2018;Elmitwally et al., 2014;Coster et al., 2013).The fault location and recovery process, however, still relies excessively on the dispatcher's short-time response, which requires the dispatcher to have rich dispatching experience and professional knowledge.When the distribution network structure is complex and multiple lines fail in a short time, the dispatcher cannot find an optimal solution swiftly (Zadsar et al., 2017;Nouri & Alamuti, 2011;Arefifar et al., 2023).
The hierarchical structure of traditional machine learning is relatively simple.If the number of samples and calculation units is not big enough, overfitting is highly likely to occur, and it is also difficult to express the correlation between the complex bidirectional power flow and faults of the active distribution network in a smart grid (Majidi, Arabali, & Etezadi-Amoli, 2014;Dashtdar et al., 2018;Liang et al., 2020;Majidi, Etezadi-Amoli, & Fadali, 2014;Orozco-Henao et al., 2018).However, a deep learning algorithm has strong advantages: On the one hand, there is an enormous amount of data for distribution network fault analysis which results from the development of wide-area measurement technology and the construction of a big data platform.The number of data-driven deep learning algorithms is big enough, and deep learning algorithm has strong ability in data mining and feature extraction when processing big data.On the other hand, the access of distributed generation and electric vehicles to the distribution network in a smart grid generates a large number of harmonics, which aggravates the nonlinearity and randomness of the distribution network (Dashti et al., 2018;Orozco-Henao et al., 2017;Deng et al., 2020).

RELATED RESEARCH
Deep learning conducts multiple feature extraction of hidden layer feature information by building a multi-hidden layer and multi-depth neural network model and obtains abstract features of the signals from deeper neurons.Deep learning has an advantage over traditional methods in locating faults in distribution networks since the former can effectively identify and classify mixed power quality interference signals.There is a good prospect for an intelligent distribution network system that can continuously update model parameters from experience and data and improve the accuracy of the whole deep learning model.Chen et al. (2019) used the quantity measurement on different buses of the distribution network in combination with the graphical convolution network to judge the fault section and obtained a high accuracy in fault section judgment.Li et al. (2019) proposed using CNN to analyze the bus voltage to achieve fault location and discussed the ability of a CNN model to identify fault sections in the case of incomplete measurement of different bus voltages.Guo et al. (2020) proposed a method of fault identification for a distribution system based on the wavelet transform and a CNN.Ahmad et al. (2020) conducted a partial discharge pattern analysis of medium voltage overhead covered conductors (OCC) based on long-term and short-term deep learning methods for fault power line detection, testing methods on the real data set provided by VSB of eNet center and trained and validated by the k-fold hierarchical cross-validation method.Livani and Evrenosoglu (2013) proposed a hybrid transmission line fault location method by combining overhead lines and underground cables.Jiao and Wu (2018) used a fuzzy inference system as the initial fusion step to identify fault scenarios and establish a fault location model for transmission lines based on multi-sensor data fusion.Liang et al. (2020) proposed a method for fault section location of a distribution network based on a double-ended zero sequence current and adaptive CNN.Bukhari et al. ( 2020) trained a CNN using the three-phase current waveform information to identify the fault type and fault section.Du et al. (2018) used CNN to distinguish the fault phases of distribution lines in the case of single-phase ground fault.
The above methods, however, are low in accuracy and do not bring satisfactory ability in feature extraction when used for distribution network fault location in distributed generation connected to a smart grid.In our study, a thermal coding and normalization method was used to preprocess the data collected by the robot fault acquisition system.We propose an improved BiLSTM fault location method, and train I-BiLSTM network and U-BiLSTM network, respectively, and can accurately locate the fault segment according to the data of each node in the robot fault collection system topology.A fault self-healing method for a distribution network based on robot inspection and deep learning under cloud edge architecture is proposed.The innovation points are as follows: 1.By combining BiLSTM and attention mechanism, we propose an improved BiLSTM fault location method to improve the feature extraction ability of the system.2. We collect and form time series of each sequence component of current and voltage, form data sets of current and voltage, train the I-BiLSTM network and U-BiLSTM network, respectively, and rely on the data of measuring points configured at each node of the robot fault collection system topology to improve the accuracy of fault location effectively.

CLoUD EDGE CoLLABoRATIoN ARCHITECTURE
The cloud edge collaboration architecture is shown in Figure 1.The model is divided into three layers: the data collection layer, edge node layer, and cloud center layer.First, in the data collection layer, the operation data concerning temperature, humidity, number of lightning strikes and lines, etc., were obtained through the robot patrol fault acquisition system, and these data go through the preliminary analysis.Second, in the edge computing layer, hot coding and normalization methods were used to preprocess the data collected by the robot fault acquisition system to prevent data flooding.Finally, the I-BiLSTM network and U-BiLSTM network were trained in the cloud, and the improved BiLSTM was used to train and learn the fault characteristics in the cloud, to accurately locate the fault segment according to the data of each node of the robot fault collection system topology.

Fault Collection
This study aims to establish a fault collection system for robots and make predictions for faults by selecting the appropriate data.Operating data concerning temperature, humidity, and lightning strikes are obtained through the robot's fault collection system, and the data regarding equipment location, operation time, and line length are derived from the geographic information system, both of which are used to divide the initial fault feature set of the active distribution network in smart grid.When a line failure occurs, the robot fault collection system can record and upload relevant pre-and post-failure line data in real time.The cloud center, by analyzing and calculating the data from the robot's fault collection system, can quickly locate the fault section and restore the power supply in the non-fault area.

Data Preprocessing
In case of a feeder fault, each terminal device arranged at the line section switch can monitor the fault current, transmit the overcurrent alarm information with a time scale to the control master station, learn the characteristic law, and realize fault location by collecting the current and voltage values at the monitoring points.The unified fault data vector is as follows: The topology is configured with n measuring points in total.FTU collects the second harmonic component i n , u n at measuring point n .Variable P DGm is the active output component of the m th distributed power supply.Because the constructed deep learning structure is supervised learning, the fault branch is used as a label.The tag needs to be encoded.Because the label data of the fault branch is discrete, it has classification characteristics.So, the hot coding rule is used to preprocess the label of the fault branch.The specific coding rules are shown in Table 1.
In practice, due to various reasons, data distortion or missing will be encountered when considering the uploading of data from the measuring point.Data cleaning includes invalid data cleaning, missing value processing, etc.When an abnormality is detected in the data preprocessing stage, the data is covered by the average value of the data at this point in the historical fault data set.At the same time, in order to prevent the difference in the order of magnitude in the data, the minimax method is used for normalization processing:

Framework for Fault Location in a Distribution Network
The framework for fault location in a distribution network is shown in Figure 2. The existing historical fault vector data is preprocessed, and the data is divided into two parts: the training set and the test set.
One-hot coding and normalization methods are used to preprocess the data collected by the robot fault collection system to prevent data flooding.Using BiLSTM and the attention mechanism, we propose an improved BiLSTM fault location method to adjust attention weight.The I-BiLSTM network and the U-BiLSTM network were trained, respectively, and the fault section can be accurately located by relying on the data of measurement points configured at each node of the robot's fault collection system topology.If a fault occurs, the collected data on the fault vector is loaded into the model, and the fault location branch label is output to achieve fault location.

BiLSTM
In this paper, at the core of the model is BiLSTM, which generates and trains the representation vector of data.In LSTM, C t and C t-1 are memory units of the current time and the historical time, respectively, h t and h t-1 are implied units of the current time and the historical time, respectively.
The input gate at the current time is i t , f t is the forgetting gate, and o t is the output gate.Variable X mt is the value of n characteristic quantities in the m th time period on day t .The LSTM memory cell stores all the state information of two adjacent times.The input gate determines the proportion of the information at the current time entering the memory cell.The forgetting gate discards part of the information of the memory cell at the historical time and updates the rest to the current time.The output gate determines the amount of the current information to be passed to the hidden cell and participate in the next state update.BiLSTM can perform high-level abstraction and nonlinear transformation on the collected fault data to provide more fine-grained calculation:

Attention Mechanism
An attention mechanism is introduced to address the problem that the existing methods cannot cope with the redundant data properly.The calculation process of the attention mechanism algorithm is described as follows: Firstly, the input is converted into three matrices: the key value K , the query value Q , and the value V The number of parallel attention headers h is used to process different parts of the matrix.Secondly, the attention score of the zoom point multiplication attention mechanism is calculated.The above process is described as follows: where Q is the query matrix, K is the key matrix, V is the numerical matrix, d is the number of hidden layer nodes, and n is the length of the input data.

Fault Location Method
The fault location method mainly comprises two steps: local training and testing, and fault location.A mature BiLSTM network grows out of local training and testing.When a fault occurs, the real-time data of the robot's fault collection system is collected and entered into the existing network, and combined with logic gate processing and judgment, real-time fault location is realized.Under the framework of tensor flow, the current sequence of the BiLSTM network is established for voltage sequence and current sequence, respectively.The voltage sequence BiLSTM (U-BiLSTM) and the current sequence BiLSTM (I-BiLSTM) networks are trained identically in the following steps: 1. Divide the active distribution network of the smart grid into multiple double-ended, non-branching sections, define the head and end, and collect the time series of current and voltage at both ends of each section before and after the fault.2. Reasonably expand the existing sampling data set, perform data preprocessing, encode labels, and divide the data set into a training set and a test set.

Based on the dimension of the training set in
Step 2, establish the BiLSTM network for the voltage sequence and the current sequence, respectively, in the tensor flow framework.4. Calculate and optimize the loss value gradient in a single step, and give the learning direction of feedback adjustment network parameters, which represents the model for learning and training.5.When the set number of steps is reached, discontinue the training; the BiLSTM network is finalized.Otherwise, adjust the network parameters and repeat Step 4 to train the network.
For the active distribution network, the voltage sequence U k and the current sequence I k of the defined section i are: and where subscripts F k and L k denote head end nodes and end nodes of section i .The subscript abc indicates that the sequence contains abc three-phase data, and u and u ' denote the voltage sequence before the fault and the voltage sequence after the fault, respectively.After a fault occurs, the collected data of section I is loaded into two networks, and a fully connected output layer is added at the end.After being classified by the softmax classifier, two neuron values are output, of which the largest neuron value is determined to be 1 and the smaller output value is determined to be 0.That is, the fault section can be represented by [1,0] and the normal section by [0,1].A logic gate is set at the outputs of the two BiLSTM networks, and the setting value of the first neuron in the output layer of each network is taken as the input, and finally, the section state determination is obtained.The process of fault location is shown in Figure 4.The basic corresponding rules of logic gates are shown in Table 2.
In this study, the section status is set to three types, namely the normal section, suspicious section, and fault section, whose main purpose is to improve the method's fault tolerance rate and provide directional suggestions.The results of the example suggest that better fault characteristics of the current sequence generate a higher final recognition rate and a more reliable model.In order to prevent the neural network from misjudging the voltage sequence, when the input is [0,1], it still indicates a section fault; that is, the U-BiLSTM network plays an auxiliary role in judging.It can be seen from the logic gate rules that if the input current sequence is partially missing or abnormal-if the I-BiLSTM determines that it is invalid and the U-BiLSTM determines that it is successful, the suspicious section can be determined, and the fault tolerance of the model can be ensured by reading the data uploaded by FTU and checking the circuit.

Network Model and Data Set
The simulation diagram of distribution network fault location with multiple DGs is established based on MATLAB, as shown in Figure 5.In the simulation diagram, S is the system power supply, and DG1-DG3 are the distributed power supplies; K1-K3 are the switches for DG access to the distribution network.If DG is connected, the corresponding switch value is 1; otherwise, it is 0; CB1-CB3 are incoming circuit breakers, S1-S28 are section switches, and L1-L29 are feeder section numbers.Since the double-ended sequence of each section to be set with traversal section fault in the traversal table is included in the samples in each simulation, a total of 260,000 samples of current section sequence and voltage section sequence are formed through automatic simulation traversal, of which 140,000 are selected as the training data set, and 20,000 are selected as the test data set.

Impact of Input Time Steps on the Results
In order to analyze the impact of the input time steps on the results, we derive the loss value of the current model from the fixed training step with the accuracy rate as the index, and the results are shown in Figure 6.For the I-BiLSTM network, the input is made in the form of 12 time steps, which will not increase after about 15 steps.The accuracy rate is low, and the convergence is slow, with the final accuracy rate being 0.795.When 288 time steps are input, the loss value converges to about 17 steps, with the discrimination accuracy finally reaching 0.917 in the test set.For the U-BiLSTM network, 288 time steps ultimately performed better, and the final accuracy reached 0.887.The final show that the proposed method achieves the best indexes.The accuracy rate of the four fault types is 0.928, 0.933, 0.948, and 0.942, respectively, all of which are higher than those in the comparative literature.This is because the proposed algorithm combines the advantages of the BiLSTM and attention mechanism, adjusts attention weight, and filters or weakens miscellaneous information.
Relying on the data of measuring points configured at each node of the robot's fault collection system topology, the fault section can be accurately located.However, the comparison methods are not deep enough to extract data features.Therefore, compared with the comparison method, the accuracy index of the proposed method is significantly improved.

Application Cases
On the basis of verifying the effectiveness of the distribution network fault self-healing method based on robot and deep learning proposed in this paper, this section introduces the application effect of the proposed method in practical projects.The verification is conducted on the reactive power and voltage coordinated regulation project, part of a distribution network of the state grid.The topology of Beizhuang line of distribution network is shown in Figure 8.The distribution network generally handles heavy loads in July and August, the peak period for power consumption that often sees short circuit faults.Before the implementation of the project, most of the lines in this distribution grid have no fault self-healing function.

CoNCLUSIoN
This paper proposes a method of fault self-healing in a distribution network based on robot inspection and deep learning in a cloud edge architecture.By using one-hot coding and normalization methods to preprocess the data collected by the robot's fault collection system, we propose an improved fault location method based on BiLSTM and integrate the attention mechanism to filter or weaken the miscellaneous information.The experimental results show that the proposed method performs well in locating the faults in a distribution network.As far as the fault line location of the active distribution network is concerned, this paper only verifies it at the simulation level.In the next step, this algorithm can be extended to the actual active distribution network system, mining more physical features, reducing the localization range, and determining the accurate fault location.In subsequent research, we will explore the way to further adjust and train the hyperparameters of the deep learning algorithm in a more scientific and refined manner, generate a simpler operation scheme in engineering practice, and verify the engineering feasibility of the proposed method.

*
Figure 1.Cloud edge collaboration architecture where W c , W f , W i , and W o are the weights, and b c , b f , b i , and b o are the corresponding offset coefficients.

Figure 2 .
Figure 2. The framework for fault location in a distribution network Figure 3.The structure of BiLSTM Figure 4.The process of fault location

Figure 5 .
Figure 5. Simulation diagram of distribution network fault location with multiple DGs Figure 6.Comparison of accuracy in different time steps (A: I-BiLSTM; B: U-BiLSTM)