Annular DC micro-grid fault positioning method based on local data driving

Abstract

The invention relates to an annular DC micro-grid fault positioning method based on local data driving, and provides an accurate positioning method based on local fault data driving aiming at the fault positioning problem under the characteristics of few fault samples and weak positive current so as to provide a basis for maintenance work after fault isolation. In order to solve the problem of insufficient fault samples, a fault data enhancement method based on an attention mechanism and a gated circulation unit neural network is provided, and training data required by intelligent fault positioning is expanded by predicting fault voltage and current dynamics after fault isolation. In order to solve the problems of high resistance and negative grounding fault positioning caused by weak positive current characteristics, a fault positioning method combining discrete wavelet transform, an extreme learning machine and an adaptive enhancement algorithm is provided, implicit fault position information is extracted and learned from a fault sample with weak characteristics, and the line fault positioning precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of DC microgrid fault diagnosis, and in particular to a ring DC microgrid fault locating method based on local data drive. Background Art

[0002] The new power system is an important hub of the new energy system and a key carrier for achieving the "dual carbon" goal. As an important part of the new power system, the DC microgrid can efficiently integrate various types of renewable power sources, energy storage units, electric vehicles and other DC loads, and has broad application prospects. DC microgrid line failures are inevitable. The low inertia and weak overcurrent capacity of the system require that the fault be diagnosed within 1-4ms after the fault occurs. However, since the failure mechanism and characteristics of the DC microgrid are very different from those of the traditional AC microgrid and DC transmission system, there is still a lack of relevant protection specifications and industry standards, which poses a challenge to the diagnosis, classification and positioning of DC microgrid line faults with fast transients, weak characteristics and little information.

[0003] DC microgrids can have radial, ring, and mesh topologies. The ring topology can be simplified to radial or expanded to mesh. The power supply in a ring microgrid can supply power in any direction to ensure stable power supply in the event of a fault. These features make the ring microgrid attractive for application. However, there are two paths between the fault node and any node in the ring structure, which increases the difficulty of fault analysis and diagnosis.

[0004] Existing DC microgrid line fault diagnosis and location mostly rely on the measurement and transmission of fault information such as line current and voltage, which increases investment costs and reduces protection reliability. Reducing the number of sensors can reduce protection costs, but it will also further reduce the available fault information, bringing new challenges to protection.

[0005] Compared with the existing technology, the technical differences are as follows:

[0006] Comparison with patent CN117872038A "A method and device for locating the source of instability faults in DC microgrids based on graph theory"

[0007] In patent CN117872038A, the key electrical characteristics are collected through the unstable fault feature collection and extraction link to accurately capture the unstable oscillation phenomenon, and the fault feature vector is constructed using wavelet packet decomposition. The graph convolutional neural network is used for deep learning training to further improve the accuracy and efficiency of diagnosis. The implementation of this method requires the collection and processing of a large amount of electrical data. If the data collection is not comprehensive or there is noise, it may affect the establishment of the model. The graph convolutional neural network used may not be able to quickly detect faults when processing large-scale power grids due to the need for a large amount of computing resources.

[0008] In the "ring DC microgrid fault location based on local data drive" proposed in the present invention, a line fault offline location method based on the bus side voltage and positive current data driven by a single-ended converter is proposed. The problem of weak fault characteristics is solved based on intelligent fault prediction and fault feature extraction, and the accuracy and reliability of DC microgrid line fault location are improved in the context of few sensors and low communication dependence.

[0009] 2. Comparison with Patent CN117872038A “Short-circuit fault location method for ring DC microgrid based on line model and Euclidean distance”

[0010] In patent CN116718872A, the sampling current of the line is obtained through intelligent electronic devices or current sensors, and after randomly generating fault schemes, the transient fault current is calculated through the established system mathematical model, and the Euclidean distance between the sampling current is calculated to determine whether it meets the relevant threshold, and the genetic algorithm is used to optimize the scheme to find the optimal solution. This method requires setting up more sensors to obtain current information on the line, which increases the construction cost, and the algorithm parameters are more and is not suitable for low-resistance fault conditions.

[0011] The present invention is driven by local data and collects the voltage and current at the output end of the converter. In the case of few sensors, the bus-side voltage and positive current required for converter control can be used to realize fault diagnosis and positioning, thereby ensuring the safe and stable operation of the DC microgrid and facilitating the promotion and application of DC microgrid technology.

[0012] 3. Comparison with Patent CN116826674A “A method for protecting inter-pole faults of a DC microgrid with photovoltaic storage and charging group adapted to control mode switching”

[0013] In patent CN116826674A, a method for positioning the inter-pole fault protection of a photovoltaic storage charging DC microgrid protection system is established to obtain protection positioning criteria, thereby performing an action export and achieving isolation of the fault area. The power electronic transformer, the circuit breaker directly connected to the power electronic transformer, and each circuit breaker in the photovoltaic storage charging station are used to obtain data information. The speed of information collection and integration will directly affect the time for accurate fault positioning, and this method is only for inter-pole short-circuit faults of DC microgrids, and its application range is not wide.

[0014] The present invention constructs positioning feature quantities using discrete wavelet transform, and then designs a positioning model by combining extreme learning machine and Adaboost algorithm, mining implicit fault location information from samples with weaker fault features and learning to accurately locate different line faults.

[0015] IV. Comparison with Patent CN118449099A “A DC Microgrid Fault Protection Method Adaptable to Various Network Topologies”

[0016] In patent CN118449099A, the DC microgrid is judged whether a fault occurs based on the current and voltage change rates collected at the circuit breaker in the DC microgrid, and then the fault mark is calculated based on the voltage and current information collected at each circuit breaker in the DC microgrid, and finally the circuit breaker is judged whether it is connected to a two-terminal line or a multi-terminal line, and whether it is a fault point. This method needs to collect current and voltage information of a large number of circuit breakers, and the configured global protection unit needs to make a global calculation judgment through all circuit breaker information, and any circuit breaker information collection will affect the final result.

[0017] The present invention proposes an Att-GRU data enhancement method, which expands training samples by predicting the voltage and current dynamics after fault isolation, reduces the use of sensors, reduces the dependence on fault samples, reduces construction costs, and improves the accuracy of fault location. Summary of the invention

[0018] In order to solve the above technical problems, the present invention proposes a ring DC microgrid fault location method based on local data drive, which solves the problems of few fault samples and weak fault characteristics based on intelligent fault prediction and fault feature extraction, improves the accuracy and reliability of DC microgrid line fault location in the context of few sensors and low communication dependence, and provides a reference for improving the DC fault protection system.

[0019] To achieve the above object, the technical solution adopted by the present invention is:

[0020] The fault location method of the ring DC microgrid driven by local data has the following specific steps:

[0021] Step 1) constructing a fault sample set;

[0022] Step 2) Fault data preprocessing based on sliding window;

[0023] Step 3) Fault prediction and location model training;

[0024] Step 4) Data enhancement based on Att-GRU prediction model;

[0025] Step 5) Fault feature extraction;

[0026] Step 6) Fault localization model based on ELM-Adaboost.

[0027] Furthermore, the step 1) of constructing the fault sample set is specifically as follows:

[0028] A ring DC microgrid simulation model was built based on MATLAB / Simulink and OPAL-RT4510 real-time simulator to simulate line faults under different fault locations, fault resistances, fault types, and load sizes. The positive current and voltage on the converter bus side before and after the fault were collected and sorted into N*S original data sets, where N is the sampling data length, which depends on the sampling frequency and time window size; S is the number of simulation cases considered.

[0029] Furthermore, the fault data preprocessing based on the sliding window in step 2) is specifically as follows:

[0030] The training data sets required for the Att-GRU prediction model and the ELM-Adaboost positioning model are constructed based on different data processing methods: For the prediction model, the variable-length sliding window method is used to divide the data into multiple subsequences in chronological order, and the data at the front of the chronological order is used as the training input, and the data at the back of the chronological order is used as the training label; For the positioning model, the fixed-length sliding window method is used to construct multiple sub-datasets containing voltage and current in different periods before and after the fault, and then the feature vector is constructed based on discrete wavelet decomposition as the training input of the positioning model, and the relative position of the fault point in the line is used as the training label;

[0031] Use w sliding windows to construct w new data sets containing different sampling time periods, and adjust the sliding step size n s and window length n w , you can control the number of sub-datasets generated and the degree of overlap. The number of windows is calculated as follows:

[0032] W=(Nn w ) / n s +1 (1)

[0033] Where N represents the length of time contained in the sampled data, which depends on the sampling frequency. Taking the data processing process of the fault voltage as an example, the generated sub-data set is expressed as:

[0034]

[0035] In the above formula, i=0,1,2,…Nn w .

[0036] Further, the step 3) fault prediction and location model training is specifically as follows:

[0037] Taking the above fault feature dataset as input and minimizing the prediction error as the objective function, the Att-GRU model is trained based on the gradient descent algorithm until the prediction error is no longer significantly improved. For the ELM-Adaboost model, the internal weights of each extreme learning machine weak locator are calculated according to the positioning error on the training dataset, and the training samples and weak locator weights are continuously adjusted according to the Adaboost algorithm until all weak locators are trained.

[0038] Furthermore, the data enhancement of step 4) based on the Att-GRU prediction model is as follows:

[0039] The fault prediction model based on Att-GRU is used to capture the changing trend of fault signals and predict the fault current and voltage data for a period of time after fault isolation;

[0040] The establishment process of the double hidden layer Att-GRU model is as follows, where N1 and N2 are the number of nodes in the two layers respectively;

[0041] Firstly, based on the simulation experiment, the bus voltage data of the converter before and after the line fault occurs are collected. Different fault locations and fault resistances are considered in the simulation process to obtain more available information.

[0042] Secondly, the input samples and corresponding prediction labels of the neural network are constructed based on the variable-length sliding window method to obtain training data sets of different lengths to meet different prediction time requirements;

[0043] Finally, the Att-GRU model is trained with the minimum prediction error as the objective function;

[0044] After the model training is completed, the voltage after a period of time after fault isolation is predicted based on the real-time sampled voltage before fault isolation; the enhanced data set can be used for subsequent feature extraction and positioning. In order to further improve the robustness and generalization ability of the prediction model, a dropout mechanism is added to the attention layer. During the calculation stage of the forward propagation of the neural network, the output of the neuron is set to zero with a certain probability.

[0045] Further, the step 5) of fault feature extraction is specifically as follows:

[0046] The features of the enhanced data are extracted using discrete wavelets, and then the total wavelet energy, relative wavelet energy and relative wavelet energy entropy are calculated to construct the localization feature vector.

[0047] According to the discrete wavelet transform decomposition result, the feature vector is constructed based on the relative wavelet energy. If the signal is transformed by discrete wavelet and K detail coefficients are obtained at the decomposition level j, then the wavelet energy E of the signal at the decomposition level j is j It can be defined as:

[0048]

[0049] In Equation (3), j = 1, 2, …, M, k = 1, 2, …, K represent different detail coefficients d at decomposition level j j,k At the corresponding positions, the total wavelet energy at all decomposition levels can be defined as:

[0050]

[0051] The relative wavelet energy represents the probability distribution of the wavelet energy at different decomposition levels, and can reduce the data size while retaining the characteristics of the original signal. Its definition is as follows:

[0052]

[0053] The relative wavelet energy is used to quantify the difference in the energy distribution of the signal at different scales, so as to screen out the important features of the signal. To further improve the positioning accuracy, the relative wavelet energy entropy is calculated to describe the uniformity information of the energy distribution of the signal at different frequencies:

[0054]

[0055] Combining the above feature quantities based on discrete wavelet transform, the input feature vector of the positioning model can be constructed as follows:

[0056]

[0057] In Equation (7), N is the length of the original signal, and M < N. After processing, the original signal with a length of 2N will be converted into a feature vector with a length of 2M + 4;

[0058] In the feature extraction based on discrete wavelet transform, the mother wavelet function ψ(t) and the maximum decomposition level M are two important parameters affecting the extraction effect;

[0059] For the maximum decomposition level M, it is usually jointly determined by the length N of the original signal and the length F of the mother wavelet filter according to the following empirical formula:

[0060]

[0061] Furthermore, the fault location model based on ELM - Adaboost in step 6) is specifically as follows

[0062] (1) Data selection and sample weight initialization: Randomly select m groups of training data from the input sample data, and initialize the sample weight of the t - th weak locator ELM t as follows:

[0063] D(t) = (ω t1,ω t1 ,L,ω tm ); ω ti =1 / m,i=1,2,Lm (9)

[0064] (2) This round of weak locator training and positioning error calculation: The weak locator is trained using the samples with initialized weights. For the t-th weak locator, the positioning error is calculated as follows:

[0065] e Li =p esti -p acti (10)

[0066] In formula (10), p esti For weak locator ELM t Estimated fault location, p acti is the actual fault location;

[0067] (3) Calculate the weight of the weak locator in this round: Calculate the weight coefficient α of the weak classifier t based on the positioning error t , giving greater weight to weak locators with better positioning effect:

[0068]

[0069] (4) Update the weight of the next round of training samples: adjust the weight D(t+1) of the t+1th weak locator training sample so that the samples that are not correctly located in this round are given more attention in the next round of training. The sample weight calculation is as follows:

[0070]

[0071] In formula (12), is the normalization factor, since a t > 0, so if the positioning error of the i-th sample is large in this round, the weight of the sample in the next round of training will be increased;

[0072] (5) Iterative training and strong classifier construction: Repeat steps (2) to (4) until T rounds of training for all weak locators are completed. Based on the weighted decision method, the final strong locator is constructed as follows:

[0073]

[0074] Beneficial effects:

[0075] This chapter of the invention proposes a line fault offline positioning method based on the bus side voltage and positive current data of a single-ended converter. It solves the problems of few fault samples and weak fault characteristics based on intelligent fault prediction and fault feature extraction, improves the accuracy and reliability of DC microgrid line fault positioning in the context of few sensors and low communication dependence, and provides a reference for improving the DC fault protection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is the overall structure diagram of the present invention;

[0077] Figure 2 This is a schematic diagram of fault voltage sliding window preprocessing;

[0078] Figure 3 Relative wavelet energy distribution of positive grounding faults at different locations: (a) p = 0.5, current RWE; (b) p = 0.9, current RWE; (c) p = 0.5, voltage RWE; (a) p = 0.9, voltage RWE;

[0079] Figure 4 Relative wavelet energy distribution of negative grounding faults at different locations: (a) p = 0.5, current RWE; (b) p = 0.9, current RWE; (c) p = 0.5, voltage RWE; (a) p = 0.9, voltage RWE

[0080] Figure 5 The Att-GRU prediction model training flow chart of the present invention;

[0081] Figure 6 This is a topological diagram of the isolated island ring DC microgrid studied in the present invention;

[0082] Figure 7 For the present invention, p=0.65, R f =4,T pre = Data prediction results after fault isolation at time T3: (a) PP fault current; (b) PP fault voltage; (c) NPG fault current; (d) NPG fault voltage;

[0083] Figure 8 Schematic diagram of positioning error of BES3 test data under different faults: (a) PP; (b) PPG; (c) NPG. DETAILED DESCRIPTION

[0084] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0085] The technical solution adopted by the present invention is:

[0086] A method for fault location of a ring DC microgrid based on local data drive mainly includes two modules: fault modeling and fault location. Each module includes three links. The steps are as shown in the attached figure. Figure 1 As shown:

[0087] Step 1) Construction of fault sample set

[0088] A ring DC microgrid simulation model is built based on MATLAB / Simulink and OPAL-RT4510 real-time simulator to simulate line faults under various conditions such as fault location, fault resistance, fault type, load size, etc. The positive current and voltage on the bus side of the converter before and after the fault are collected and sorted into N*S original data sets, where N is the sampling data length, which depends on the sampling frequency and time window size; S is the number of simulation cases considered.

[0089] Step 2) Fault data preprocessing based on sliding window

[0090] The training data sets required for the Att-GRU prediction model and the ELM-Adaboost positioning model are constructed based on different data processing methods: for the prediction model, the variable-length sliding window method is used to divide the data into multiple subsequences in chronological order. The data at the front of the chronological order is used as the training input, and the data at the back of the chronological order is used as the training label. For the positioning model, the fixed-length sliding window method is used to construct multiple sub-datasets containing voltage and current in different periods before and after the fault. Subsequently, the feature vector is constructed based on discrete wavelet decomposition as the training input of the positioning model, and the relative position of the fault point in the line is used as the training label.

[0091] Since the voltage and current data before the fault occurs contain limited information about the fault location and change trend, in order to reduce the impact of useless pre-fault data on the performance of the prediction and positioning model, the present invention adopts a sliding window method in the construction of the model training data set. Figure 2 As shown in the figure, w new data sets containing different sampling time periods are constructed using w sliding windows. By adjusting the sliding step size n s and window length n w , you can control the number of sub-datasets generated and the degree of overlap. The number of windows can be calculated as follows:

[0092] W=(Nn w ) / n s +1 (1)

[0093] Where N represents the length of time contained in the sampled data, which depends on the sampling frequency. Taking the data processing process of the fault voltage as an example, the generated sub-data set can be expressed as:

[0094]

[0095] In the above formula, i=0,1,2,…Nn w .exist Figure 2 In the example, it is assumed that the fault occurs at t = 0.5s and is isolated at 0.5015s. The original sampled voltage sequence contains data of [0.499, 0.5015]s. If w = 3, n w =2ms,n s = 0.25ms, after sliding window processing, three data windows [0.499, 0.501], [0.49925, 0.50125] and [0.4995, 0.5015] can be obtained, and three new sub-data sets are generated. In practical applications, according to specific needs, the window size n w It can be a fixed value or a variable value, thus generating multiple sub-datasets of the same or different sizes. The sliding window method can increase the amount of available data during training, reduce the impact of sampling uncertainty, and improve the generalization ability of the model.

[0096] Step 3). Fault prediction and location model training

[0097] Taking the above fault feature dataset as input and minimizing the prediction error as the objective function, the specific process of training the Att-GRU model based on the gradient descent algorithm is as follows: Figure 5 As shown in the figure, the prediction error is no longer significantly improved. For the ELM-Adaboost model, the internal weight of each extreme learning machine (ELM) weak locator is calculated according to the positioning error on the training data set, and the training samples and weak locator weights are continuously adjusted according to the Adaboost algorithm until all weak locators are trained.

[0098] Step 4). Data enhancement based on Att-GRU prediction model

[0099] In order to meet the diagnostic speed requirement, the sampled data before fault isolation may not be sufficient to achieve fault location. In order to expand the input information required for the location model under the condition of few samples, the present invention proposes a fault prediction model based on Att-GRU to capture the fault signal change trend and predict the fault current and voltage data for a period of time after fault isolation.

[0100] The dual hidden layer Att-GRU model establishment process used in the present invention is as follows: Figure 3As shown, N1 and N2 are the number of nodes in the two layers respectively. This section only gives the fault voltage prediction modeling method, and the establishment of the current prediction model is similar. First, based on the simulation experiment, the bus-side voltage data of the converter before and after the line fault occurs is collected. Different conditions such as fault location and fault resistance are considered in the simulation process to obtain more available information. Secondly, based on the variable-length sliding window method, the input samples and corresponding prediction labels of the neural network are constructed to obtain training data sets of different lengths to meet different prediction time requirements. Finally, the Att-GRU model is trained with the minimum prediction error as the objective function.

[0101] After the model training is completed, the voltage after the fault isolation can be predicted based on the real-time sampled voltage before the fault isolation. The enhanced data set can be used for subsequent feature extraction and positioning. In order to further improve the robustness and generalization ability of the prediction model, a dropout mechanism is added to the attention layer. By setting the output of the neuron to zero with a certain probability during the calculation stage of the forward propagation of the neural network, this can avoid the influence of local features on the prediction results and prevent overfitting during the training process.

[0102] Step 5) Fault feature extraction

[0103] Discrete wavelet is used to extract features from the enhanced data, and then the total wavelet energy, relative wavelet energy and relative wavelet energy entropy are calculated to construct the positioning feature vector to solve the problem of accurate positioning under weak fault characteristics.

[0104] The voltage and current signals collected after the fault are subjected to discrete wavelet transform, and a set of wavelet coefficients are output. If these coefficients are directly used as the input of the positioning model, a large amount of storage space will be occupied and the calculation cost will be increased. To this end, the present invention constructs a feature vector based on the relative wavelet energy (RWE) according to the discrete wavelet transform (DWT) decomposition result. If the signal is subjected to discrete wavelet transform and K detail coefficients are obtained at the decomposition level j, then the wavelet energy E of the signal at the decomposition level j is j It can be defined as:

[0105]

[0106] In formula (3), j = 1, 2, ... M, k = 1, 2, ... K represent different detail coefficients d at decomposition level j j,k The total wavelet energy (Energy total, ET) at all decomposition levels can be defined as:

[0107]

[0108] The relative wavelet energy represents the probability distribution of wavelet energy at different decomposition levels, and can reduce the data size while preserving the characteristics of the original signal. Its definition is as follows:

[0109]

[0110] The relative wavelet energy can be used to quantify the difference in energy distribution of the signal at different scales, so as to screen out the important features of the signal. To further improve the positioning accuracy, the relative wavelet energy entropy (RWEE) is calculated to describe the uniformity information of the energy distribution of the signal at different frequencies:

[0111]

[0112] Combining the above feature quantities based on discrete wavelet transform, the input feature vector of the positioning model can be constructed as follows:

[0113]

[0114] In Equation (7), N is the length of the original signal, and M < N. Decomposing the voltage and current respectively and then calculating the corresponding feature quantities helps to reduce the dimension of the original data while preserving the characteristics of the fault voltage and current. After processing, the original signal with a length of 2N will be converted into a feature vector with a length of 2M + 4.

[0115] In the feature extraction based on discrete wavelet transform, the mother wavelet function ψ(t) and the maximum decomposition level M are two important parameters affecting the extraction effect. At present, many mother wavelet functions have been widely used in research, such as Haar wavelet, Daubechies wavelet, Symlet wavelet, etc. Different mother wavelet functions have their own unique time-frequency domain characteristics. For different data, the mother wavelet function with the best performance may be different, and it is usually selected empirically according to the characteristics of the sampled data. When the number of samples is sufficient, the Daubechies (db) wavelet usually has good robustness and feature extraction ability, while the feature extraction ability of other wavelets is affected by the filter length. For the DC microgrid containing a large number of transient voltage and current signals, the db wavelet is selected for analysis in this chapter. The maximum decomposition level is another important parameter affecting the feature extraction effect. A larger decomposition level can挖掘 more detailed information contained in the signal, but it will increase the calculation cost. For the maximum decomposition level M, it is usually determined jointly by the length N of the original signal and the length F of the mother wavelet filter according to the following empirical formula:

[0116]

[0117] For the commonly used mother wavelets of the db family, the filter lengths are shown in Table 1. This chapter selects the db8 wavelet for feature extraction. When the sampling frequency is 20kHz, if there are 1064 groups of different fault cases considered, and the sampling data of each group of cases is 1ms, 1.5ms, and 2ms after the fault, then the size of N is 21280, 31920, and 42560 respectively. According to formula (8), the corresponding maximum decomposition levels after rounding are 10, 11, and 11 respectively. When M=11, the frequency bands corresponding to different decomposition levels are shown in Table 2.

[0118] Table 1 Filter lengths of commonly used mother wavelets in the DB family

[0119]

[0120] Table 2 Corresponding frequency bands at different wavelet decomposition levels

[0121]

[0122] To verify the effect of the feature extraction method based on discrete wavelet transform, it is assumed that at t = 0.5s, a fault with a resistance of R occurs at different positions (p = 0.5, 0.9) of line 56. f =5Ω positive-pole-to-ground fault and negative-pole-to-ground fault. The voltage and current of BES3 are sampled within 1.5ms after the fault at a sampling frequency of 20kHz, and the db8 wavelet and decomposition level 11 are selected for feature extraction. The relative wavelet energy of positive-pole-to-ground fault (PPG) and negative-pole-to-ground fault (NPG) at different fault locations is as follows: Figure 3 and Figure 4 As shown in the figure. When a PPG fault occurs, the relative wavelet energy of the voltage at different fault locations is relatively similar, while the relative wavelet energy of the current is significantly different, which is related to the drastic change of the current under the positive grounding fault. Therefore, the relative wavelet energy of the fusion current and voltage is more conducive to fault location.

[0123] When an NPG fault occurs, the relative wavelet energy of the current at different fault locations is relatively similar. Although there are differences in the relative wavelet energy of the voltage, they are not as obvious as the relative wavelet energy (RWE) of the current under the PPG fault. Therefore, combining the total wavelet energy (Energy total, ET) and the relative wavelet energy entropy (RWEE) to construct the eigenvector shown in formula (6) helps to amplify the voltage difference at different fault locations and improve the positioning accuracy.

[0124] Step 6) Fault location model based on ELM-Adaboost

[0125] ELM has many advantages, but when only a single ELM model is used for positioning, it may cause overfitting or underfitting problems, affecting the reliability of the positioning results. Therefore, this chapter combines the Adaboost algorithm to improve ELM. The Adaboost algorithm is an integrated learning method that continuously trains T weak locators and dynamically adjusts the sample weights and weak locator weights according to the error of each training. If the error exceeds the threshold θ e , more weights are given to samples with larger errors and weak classifiers with better positioning effects. After all weak locators have completed T rounds of training, a strong locator can be obtained through the Adaboost algorithm. The detailed steps are as follows:

[0126] (1) Data selection and sample weight initialization: Randomly select m groups of training data from the input sample data and assign the tth weak locator ELM t The sample weights of are initialized as follows:

[0127] D(t)=(ω t1 ,ω t1 ,L,ω tm ); ω ti =1 / m,i=1,2,Lm (9)

[0128] (2) This round of weak locator training and positioning error calculation: The weak locator is trained using the samples with initialized weights. For the t-th weak locator, the positioning error is calculated as follows:

[0129] e Li =p esti -p acti (10)

[0130] In formula (10), p esti For weak locator ELM t Estimated fault location, p acti is the actual fault location.

[0131] (3) Calculate the weight of the weak locator in this round: Calculate the weight coefficient α of the weak classifier t based on the positioning error t , giving greater weight to weak locators with better positioning effect:

[0132]

[0133] (4) Update the weight of the next round of training samples: adjust the weight D(t+1) of the t+1th weak locator training sample so that the samples that were not correctly located in this round are given more attention in the next round of training. The sample weight calculation is as follows:

[0134]

[0135] In formula (12), is the normalization factor. t >0, so if the positioning error of the i-th sample in this round is large, the weight of the sample in the next round of training will be increased.

[0136] (5) Iterative training and strong classifier construction: Repeat steps (2) to (4) until T rounds of training for all weak locators are completed. Based on the weighted decision method, the final strong locator is constructed as follows:

[0137]

[0138] In order to verify the effectiveness of the proposed positioning method, the Figure 6 The fault simulation model of the ring DC microgrid shown in Figure 3 is established, and the same simulation parameters shown in Table 3 are used. For different parts of the proposed positioning method, there are corresponding parameters that need to be determined in advance. The parameters corresponding to different parts of the proposed method are shown in Table 3. The optimal number of hidden nodes in the GRU hidden layer and ELM is N1, N2, N ELM , and the positioning error threshold θ e Determined by trial and error. The sliding window step size is 5 samples, and at a sampling frequency of 20kHz, the sliding step size is 0.25ms.

[0139] Table 3 Parameters of different parts of the proposed positioning method

[0140]

[0141] Table 4 Configuration of each line fault simulation case

[0142]

[0143]

[0144] For each fault type (PP, PPG, NPG) of each line (lines 1-6), simulations with different fault resistances and fault locations were performed under different load conditions to obtain data. The cases considered in the simulation analysis are shown in Table 4. 1064 sets of cases were generated for one line fault on each line, and a total of 19152 sets of cases were generated for all three fault types and six fault lines. The fault location is expressed as the relative distance between the fault point and the end of the line in the counterclockwise direction. For example, for fault F1, p = 0.2 means that the fault is close to the photovoltaic end. The bus side voltage and current of 3ms, 1ms before the fault and 2ms after the fault, were sampled as the original data set and divided. 70% of the training data was used to train the prediction model and positioning model after sliding window partitioning and feature extraction, and 30% of the test data was used to test the data enhancement effect and positioning accuracy.

[0145] Embodiment 1:

[0146] Table 5 Different time intervals used for data augmentation

[0147]

[0148] This example illustrates the data enhancement effect. Considering different faults occurring on line 34 with a length of 1.4 km, according to the diagnosis methods proposed in the previous two chapters, PP / PPG and NPG faults can be diagnosed and isolated within 1-2 ms. Therefore, consider the three different prediction time intervals shown in Table 5 for analysis, where T pre Indicates the time interval, “-” indicates the time before the fault, and “+” indicates the time after the fault.

[0149] Under different fault resistances, the voltage and current ranges fluctuate greatly. In order to reduce the impact of numerical values ​​on the evaluation index, the normalized root-mean-square error (NRMSE) of the test data i is defined to evaluate the prediction accuracy:

[0150]

[0151] In the above formula, n pre =10 is the length of the test signal sequence in this case (0.05ms at 20kHz sampling frequency), and It is the normalized predicted value and actual value after Z-score standardization.

[0152] Table 6 shows the voltage and current prediction errors under different scenarios, where p acti is the actual fault location. In order to more intuitively display the prediction accuracy, Figure 7 The prediction results for one of the scenarios are given in . Figure 7 As can be seen from Table 6, the prediction error of NPG fault current is relatively large in time interval T1, and high prediction accuracy can be achieved in other scenarios. The above results show that the proposed Att-GRU prediction model can accurately predict the voltage and current change trends after the fault.

[0153] To further evaluate the prediction performance of the proposed Att-GRU model, the total prediction error on the test dataset is defined as follows:

[0154]

[0155] In the above formula, N test =320 is the size of the test data set. The total prediction error of the converters at both ends of the line at different time intervals is shown in Table 7. Based on the above prediction results, the following conclusions can be drawn:

[0156] (1) Since the voltage and current change very quickly during PP and PPG faults, as shown in the prediction results under time interval T1, data containing 0.5 ms within a short period of time after the fault occurs is sufficient to train the prediction model.

[0157] (2) Since the positive current does not change significantly after the NPG fault occurs, the input data length has a greater impact on the current prediction error. When using fewer sensors, at least 1ms of post-fault data is required to train the model to ensure the accuracy of the prediction.

[0158] (3) At the same time interval, the prediction errors at both ends of the line are not much different, which shows that the proposed prediction method is robust to changes in fault resistance and location, and also indicates the availability of local data at a single end of the line.

[0159] (4) The prediction error decreases with the increase of the length of post-fault data, which indicates that the post-fault data contains the temporal information about the voltage and current changes.

[0160] (5) Under different conditions, the total prediction error on the test data set does not exceed 4%, indicating that the proposed prediction model can predict the voltage and current changes within a period of time after fault isolation with high accuracy.

[0161] Based on the above analysis, when the sampling data is insufficient before fault isolation, the trained Att-GRU prediction model can be used to increase the total length of sampling data for PP / PPG faults and NPG faults to 1.5ms and 2ms, respectively, to meet the requirements of prediction accuracy and diagnosis speed.

[0162] Table 6 Prediction error NRMSE under different scenarios i

[0163]

[0164]

[0165] Table 7 Total prediction error NRMSE of test data sets under different scenarios tol

[0166]

[0167] Embodiment 2:

[0168] This embodiment illustrates the analysis of fault location accuracy. When a fault occurs, the Att-GRU model is used to enhance the data, and then a feature vector based on DWT is constructed and input into the trained ELM-Adaboost location model to determine the relative position of the fault. In order to analyze the positioning accuracy of the proposed method, this section considers different faults occurring on line 56 with a length of 1.6 km. The fault location error calculation is shown in Equation 8. Using the proposed method, the positioning results of both ends of the fault line (Load2 and BES3) are shown in Table 8. It can be seen that the positioning error is mainly affected by the fault resistance. As the fault resistance increases, the positioning error becomes larger. For different line faults, the positioning error of all test data at the BES3 end is as follows Figure 8 The points with high individual errors are mainly concentrated at the high-resistance faults at the end of the line. The positioning errors of most data are within 1%, and the positioning accuracy is high.

[0169] Table 8 Positioning error when line 56 fails

[0170]

[0171]

[0172] In order to further evaluate the performance of the proposed positioning method, the maximum location error (Maximum location error, MLE) and the mean absolute percentage location error (Mean absolute percentage location error, MAPLE) are defined as follows:

[0173]

[0174] The MLE and MAPLE under different faults are shown in Table 9. Based on the above results, the following conclusions can be drawn:

[0175] (1) The location accuracy of PP and PPG faults is higher than that of NPG faults. This is due to the difference in current characteristics. Under NPG faults, the non-fault positive electrode sampling current does not change dramatically, which reduces the available fault characteristics.

[0176] (2) For the same fault type, the positioning results and errors at both ends of the line are not much different, indicating the feasibility of positioning based on local data.

[0177] (3) Even if the positive current feature is weak under NPG fault, the MLE is only 2.31%. The MAPLE of all test data is less than 2%, indicating that the proposed location method has high accuracy for different line faults.

[0178] Table 9 MLE and MAPLE of test data under different faults

[0179]

[0180] Embodiment 3:

[0181] This embodiment verifies the effectiveness of the proposed method in a noisy environment, considering different faults occurring in line 12 with a length of 2 km for analysis. Gaussian white noise with different signal-to-noise ratios is added to the measured bus-side current and voltage signals to simulate the noise in the actual working environment. The MAPLE and MLE on the test data under different noise levels are shown in Table 10. As the noise becomes stronger, the positioning accuracy decreases. Under PP and PPG faults, the injection of noise has little effect on the current; in addition, the noise under the three faults will distort the voltage, so the NPG fault positioning error is more susceptible to noise than the PP and PPG faults. However, even in a strong noise environment of 40 dB, the average positioning error on the test data is still less than 4%, and the maximum positioning error does not exceed 4.3%. Since the noise mainly exists in the high-frequency domain, and the signal energy is mainly concentrated in the low-frequency domain, the discrete wavelet transform is used for feature extraction, and the positioning feature vector based on the wavelet energy is constructed, which can suppress the noise to a certain extent. The above results show that the proposed method can accurately locate different line faults even in a noisy environment.

[0182] Table 10 MLE and MAPLE of test data in noisy environment

[0183]

[0184] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A ring DC microgrid fault location method based on local data drive, characterized in that: The specific steps are as follows: Step 1) constructing a fault sample set; Step 2) Fault data preprocessing based on sliding window; Step 3) Fault prediction and location model training; Step 4) Data enhancement based on Att-GRU prediction model; Step 5) Fault feature extraction; Step 6) Fault localization model based on ELM-Adaboost.

2. The method for locating a fault in a ring DC microgrid based on local data drive according to claim 1, characterized in that: The step 1) of constructing the fault sample set is specifically as follows: A ring DC microgrid simulation model was built based on MATLAB / Simulink and OPAL-RT4510 real-time simulator to simulate line faults under different fault locations, fault resistances, fault types, and load sizes. The positive current and voltage on the converter bus side before and after the fault were collected and sorted into N*S original data sets, where N is the sampling data length, which depends on the sampling frequency and time window size; S is the number of simulation cases considered.

3. The method for locating a fault in a ring DC microgrid based on local data drive according to claim 1, characterized in that: The step 2) of preprocessing fault data based on sliding window is specifically as follows: The training data sets required for the Att-GRU prediction model and the ELM-Adaboost positioning model are constructed based on different data processing methods: For the prediction model, the variable-length sliding window method is used to divide the data into multiple subsequences in chronological order, the data at the front of the chronological order is used as the training input, and the data at the back of the chronological order is used as the training label; For the location model, a fixed-length sliding window method is used to construct multiple sub-datasets containing voltage and current in different periods before and after the fault. Then, a feature vector is constructed based on discrete wavelet decomposition as the training input of the location model, and the relative position of the fault point in the line is used as the training label; Use w sliding windows to construct w new data sets containing different sampling time periods, and adjust the sliding step size n s and window length n w , you can control the number of sub-datasets generated and the degree of overlap. The number of windows is calculated as follows: W=(N-n w ) / n s +1 (1) Where N represents the length of time contained in the sampled data, which depends on the sampling frequency. Taking the data processing process of the fault voltage as an example, the generated sub-data set is expressed as: In the above formula, i=0,1,2,…Nn w .

4. The method for locating a fault in a ring DC microgrid based on local data drive according to claim 1, characterized in that: The step 3) of fault prediction and location model training is specifically as follows: Taking the above fault feature dataset as input and minimizing the prediction error as the objective function, the Att-GRU model is trained based on the gradient descent algorithm until the prediction error is no longer significantly improved. For the ELM-Adaboost model, the internal weights of each extreme learning machine weak locator are calculated according to the positioning error on the training dataset, and the training samples and weak locator weights are continuously adjusted according to the Adaboost algorithm until all weak locators are trained.

5. The method for locating a fault in a ring DC microgrid based on local data drive according to claim 1, characterized in that: The data enhancement of step 4) based on the Att-GRU prediction model is as follows: The fault prediction model based on Att-GRU is used to capture the changing trend of fault signals and predict the fault current and voltage data for a period of time after fault isolation; The dual hidden layer Att-GRU model establishment process is as follows, where: N1 and N2 are the number of nodes in the two layers respectively; Firstly, based on the simulation experiment, the bus voltage data of the converter before and after the line fault occurs are collected. Different fault locations and fault resistances are considered in the simulation process to obtain more available information. Secondly, the input samples and corresponding prediction labels of the neural network are constructed based on the variable-length sliding window method to obtain training data sets of different lengths to meet different prediction time requirements; Finally, the Att-GRU model is trained with the minimum prediction error as the objective function; After the model training is completed, the voltage for a period of time after fault isolation is predicted based on the real-time sampled voltage before fault isolation; the enhanced dataset can be used for subsequent feature extraction and location. To further improve the robustness and generalization ability of the prediction model, a dropout mechanism is added to the attention layer. During the calculation stage of the forward propagation of the neural network, the output of the neuron is set to zero with a certain probability.

6. The method for locating a fault in a ring DC microgrid based on local data drive according to claim 1, characterized in that: The specific steps for fault feature extraction in step 5) are as follows: The enhanced data is subjected to feature extraction using discrete wavelets, and then the total wavelet energy, relative wavelet energy, and relative wavelet energy entropy are calculated to construct a location feature vector; According to the discrete wavelet transform decomposition result, the feature vector is constructed based on the relative wavelet energy. If the signal is transformed by discrete wavelet and K detail coefficients are obtained at the decomposition level j, then the wavelet energy E of the signal at the decomposition level j is j It can be defined as: In formula (3), j = 1, 2, ... M, k = 1, 2, ... K represent different detail coefficients d at decomposition level j j,k Corresponding to the position, the total wavelet energy at all decomposition levels can be defined as: The relative wavelet energy represents the probability distribution of the wavelet energy at different decomposition levels, and can reduce the data size while retaining the characteristics of the original signal. Its definition is as follows: The relative wavelet energy is used to quantify the energy distribution difference of the signal at different scales, so as to screen out the important features of the signal. To further improve the location accuracy, the relative wavelet energy entropy is calculated to describe the uniformity information of the energy distribution of the signal at different frequencies: Combining the above feature quantities based on discrete wavelet transform, the input feature vector of the location model can be constructed as follows: In Equation (7), N is the length of the original signal, and M < N. After processing, the original signal with a length of 2N will be converted into a feature vector with a length of 2M + 4; In the feature extraction based on discrete wavelet transform, the mother wavelet function ψ(t) and the maximum decomposition level M are two important parameters affecting the extraction effect; For the maximum decomposition level M, it is usually determined jointly by the length N of the original signal and the length F of the mother wavelet filter according to the following empirical formula:

7. The method for locating a fault in a ring DC microgrid based on local data drive according to claim 1, characterized in that: The specific steps for the fault location model based on ELM-Adaboost in step 6) are as follows (1) Data selection and sample weight initialization: Randomly select m groups of training data from the input sample data and assign the tth weak locator ELM t The sample weights of are initialized as follows: D(t)=(ω t1 ,oh t1 ,L,ω tm ); oh ti =1 / m,i=1,2,Lm (9) (2) Training of the weak locator in this round and calculation of the location error: The weak locator is trained with the samples after initializing the weights. For the t-th weak locator, the location error is calculated as follows: e Li =p esti -p acti (10) In formula (10), p esti For weak locator ELM t Estimated fault location, p acti is the actual fault location; (3) Calculate the weight of the weak locator in this round: Calculate the weight coefficient α of the weak classifier t based on the positioning error t , giving greater weight to weak locators with better positioning effect: (4) Updating the weights of the training samples in the next round: Adjust the weights D(t + 1) of the training samples for the (t + 1)-th weak locator, so that the samples that are not correctly located in this round are emphasized in the next round of training. The sample weight calculation is as follows: In formula (12), is the normalization factor, since a t > 0, so if the positioning error of the i-th sample is large in this round, the weight of the sample in the next round of training will be increased; (5) Iterative training and construction of the strong classifier: Repeat steps (2)-(4) until all weak locators are trained for T rounds. Based on the weighted decision method, the final strong locator is constructed as follows:

Images