A Local Data-Driven Fault Location Method for Ring DC Microgrids

Abstract

This paper proposes a local data-driven fault location method for ring DC microgrids. Addressing the problem of insufficient fault samples and weak positive current characteristics, a precise location method based on local fault data is presented, providing a basis for maintenance work after fault isolation. To further address the issue of insufficient fault samples, a fault data augmentation method based on attention mechanism and gated recurrent unit neural network is proposed. This method expands the training data required for intelligent fault location by predicting the dynamics of fault voltage and current after fault isolation. Finally, to address the challenges of locating high-resistance and negative-ground faults caused by weak positive current characteristics, a fault location method combining discrete wavelet transform, extreme learning machine, and adaptive augmentation algorithm is proposed. This method extracts and learns implicit fault location information from fault samples with weak features, improving the accuracy of line fault location.

Description

Technical Field

[0001] This invention belongs to the field of DC microgrid fault diagnosis technology, specifically a fault location method for ring DC microgrids based on local data. Background Technology

[0002] New power systems are crucial hubs in the new energy system and key carriers for achieving "dual-carbon" goals. DC microgrids, as an important component of new power systems, can efficiently integrate various renewable power sources, energy storage units, and DC loads such as electric vehicles, and have broad application prospects. Faults in DC microgrid lines are inevitable, and the system's low inertia and weak overcurrent capability require fault diagnosis within 1-4 ms after occurrence. However, due to the significant differences in fault mechanisms and characteristics between DC microgrids and traditional AC microgrids and DC transmission systems, relevant protection specifications and industry standards are still lacking, posing challenges to the diagnosis, classification, and location of faults in DC microgrid lines characterized by fast transients, weak characteristics, and limited information.

[0003] DC microgrids can have radial, ring, and mesh topologies. Ring topologies can be simplified to radial topologies or extended to mesh topologies. Power sources in a ring microgrid can supply power in any direction to ensure stable power supply during fault conditions. These characteristics make ring microgrid applications attractive. However, in a ring structure, there are always two paths between a faulty node and any other node, which increases the difficulty of fault analysis and diagnosis.

[0004] Current fault diagnosis and location in DC microgrids largely 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 lower protection costs, but it will further reduce available fault information, bringing new challenges to protection.

[0005] The technical differences compared to existing technologies are as follows:

[0006] Comparison with patent CN117872038A "A method and device for locating instability fault sources in DC microgrids based on graph theory"

[0007] Patent CN117872038A employs a fault feature acquisition and extraction process to collect key electrical features, accurately capturing unstable oscillations. It then utilizes wavelet packet decomposition to construct fault feature vectors and applies a graph convolutional neural network for deep learning training, further improving diagnostic accuracy and efficiency. However, this method requires the collection and processing of large amounts of electrical data. Incomplete data collection or the presence of noise can negatively impact model development. Furthermore, the graph convolutional neural network used may be insufficient for rapid fault detection in large-scale power grids due to the high computational resource requirements.

[0008] The invention proposes a method for offline fault location of a ring DC microgrid based on local data-driven data of single-ended converter bus voltage and positive current. This method solves the problem of weak fault features by using intelligent fault prediction and fault feature extraction, and improves the accuracy and reliability of DC microgrid line fault location in the context of few sensors and low communication dependence.

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

[0010] Patent CN116718872A describes a method that uses intelligent electronic devices or current sensors to acquire sampled current in a line. After randomly generating fault scenarios, a transient fault current is calculated using an established system mathematical model. The Euclidean distance between the current sampled and the current sampled is then calculated to determine if it meets relevant thresholds. Finally, a genetic algorithm is used to optimize the scenarios and find the optimal solution. However, this method requires numerous sensors to acquire current information on the line, increasing construction costs. Furthermore, the algorithm has many parameters and is not suitable for low-resistance faults.

[0011] This invention is based on local data drive and collects the voltage and current at the output of the converter. It can realize fault diagnosis and location by using the bus side voltage and positive current necessary for converter control with fewer sensors, thereby ensuring the safe and stable operation of DC microgrid and promoting the application of DC microgrid technology.

[0012] III. Comparison with Patent CN116826674A "A Method for Inter-electrode Fault Protection of a Photovoltaic-Storage-Charging Cluster DC Microgrid Adapting to Control Mode Switching"

[0013] Patent CN116826674A describes a method for locating inter-pole faults in a photovoltaic-storage-charging cluster DC microgrid protection system. This method obtains protection location criteria, which are then used to initiate action and isolate the fault area. However, it utilizes data collected from power electronic transformers, circuit breakers directly connected to these transformers, and various circuit breakers within the photovoltaic-storage-charging station. The speed of data acquisition and integration directly impacts the time required for accurate fault location. Furthermore, this method is only applicable to inter-pole short-circuit faults in DC microgrids, limiting its applicability.

[0014] This invention utilizes discrete wavelet transform to construct localization features. Subsequently, a localization model is designed by combining extreme learning machine and Adaboost algorithm to mine hidden fault location information from samples with weak fault features and learn to accurately locate faults on different lines.

[0015] IV. Comparison with Patent CN118449099A "A Fault Protection Method for DC Microgrids Adapted to Multiple Network Topologies"

[0016] Patent CN118449099A determines whether a fault has occurred in a DC microgrid based on the current and voltage change rates collected at circuit breakers within the microgrid. It then calculates fault flags based on voltage and current information collected at each circuit breaker within the microgrid, and finally determines whether the circuit breaker is connected to a two-terminal or multi-terminal line, and whether it is a fault point. This method requires collecting current and voltage information from a large number of circuit breakers, and the configured global protection module needs to use information from all circuit breakers to make global calculations and judgments. The collection status of any circuit breaker information will affect the final result.

[0017] This invention proposes an Att-GRU data augmentation method that expands the training samples by predicting the voltage and current dynamics after fault isolation, thereby reducing the use of sensors, reducing dependence on fault samples, lowering construction costs, and improving fault location accuracy. Summary of the Invention

[0018] To address the aforementioned technical problems, this invention proposes a fault location method for ring DC microgrids based on local data. By using intelligent fault prediction and fault feature extraction, it solves the problems of few fault samples and weak fault features, improving the accuracy and reliability of DC microgrid line fault location in the context of few sensors and low communication dependence, and providing a reference for improving the DC fault protection system.

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

[0020] The fault location method for ring DC microgrids based on local data-driven approach comprises the following steps:

[0021] Step 1) Constructing the fault sample set;

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

[0023] Step 3) Training the fault prediction and localization model;

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

[0025] Step 5) Fault feature extraction;

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

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

[0028] A ring-shaped DC microgrid simulation model was built using MATLAB / Simulink and the 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 of the converter bus side were collected before and after the fault and compiled into an N*S raw dataset, where N is the length of the sampled data, which depends on the sampling frequency and the size of the time window; and S is the number of simulation cases considered.

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

[0030] The training datasets required for the Att-GRU prediction model and the ELM-Adaboost localization model are constructed based on different data processing methods: For the prediction model, the data is divided into multiple subsequences according to time order using the variable-length sliding window method, with the data at the beginning of the time order used as the training input and the data at the end of the time order used as the training labels; For the localization model, multiple subsequences containing voltage and current at different time periods before and after the fault are constructed using the fixed-length sliding window method, and then feature vectors are constructed based on discrete wavelet decomposition as the training input for the localization model, with the relative position of the fault point in the line used as the training labels.

[0031] We construct w new datasets containing different sampling time periods using w sliding windows, by adjusting the sliding step size n. s and window length n w The number of generated subsets and their degree of overlap can be controlled. The number of windows is calculated as follows:

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

[0033] Where N represents the time length contained in the sampled data, which depends on the sampling frequency. Taking the data processing of fault voltage as an example, the generated subset of data is represented as follows:

[0034]

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

[0036] Furthermore, step 3) training the fault prediction and localization model is as follows:

[0037] Using the aforementioned 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 no longer shows significant improvement. For the ELM-Adaboost model, the internal weights of each weak locator in the extreme learning machine are calculated based on the localization 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 augmentation in step 4) based on the Att-GRU prediction model is as follows:

[0039] A fault prediction model based on Att-GRU is used to capture the trend of fault signal changes and predict fault current and voltage data for a period of time after fault isolation.

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

[0041] First, based on simulation experiments, bus-side voltage data of the converter before and after a line fault were collected. During the simulation, different fault locations and fault resistances were considered to obtain more usable 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 datasets 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 is trained, the voltage after 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 localization. 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 forward propagation computation phase of the neural network, the output of the neuron is set to zero with a certain probability.

[0045] Furthermore, step 5) of fault feature extraction is as follows:

[0046] Discrete wavelets are 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 a localization feature vector.

[0047] Based on the discrete wavelet transform decomposition results, an eigenvector is constructed based on the relative wavelet energy. If the signal, after discrete wavelet transform, obtains K detail coefficients at decomposition level j, then the wavelet energy E of the signal at 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) Training and localization error calculation of the weak locator in this round: The weak locator is trained using samples with initialized weights. For the t-th weak locator, the localization error is calculated as follows:

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

[0066] In equation (10), p esti For weak positioner ELM t Estimated fault location, p acti This indicates the actual location of the fault.

[0067] (3) Calculate the weights of the weak locators in this round: Calculate the weight coefficient α of the weak classifier t based on the positioning error. t Give greater weight to weak locators that have better positioning performance:

[0068]

[0069] (4) Update the training sample weights for the next round: Adjust the training sample weights D(t+1) of the (t+1)th weak localizer so that the samples that were not correctly localized in this round will be given more attention in the next round of training. The sample weights are calculated as follows:

[0070]

[0071] In formula (12) As the normalization factor, since a t >0, therefore, 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 construction of strong classifier: Repeat steps (2)-(4) until all weak localizers have completed T rounds of training. Based on the weighted decision method, the final strong localizer is constructed as follows:

[0073]

[0074] Beneficial effects:

[0075] This invention proposes an offline fault location method for DC microgrids based on bus-side voltage and positive current data driven by single-ended converters. By using intelligent fault prediction and fault feature extraction, it solves the problems of few fault samples and weak fault features, and improves the accuracy and reliability of fault location in DC microgrids with few sensors and low communication dependence, providing a reference for improving the DC fault protection system. Attached Figure Description

[0076] Figure 1 This is an overall structural diagram of the present invention;

[0077] Figure 2 A schematic diagram of the fault voltage sliding window preprocessing;

[0078] Figure 3 Relative wavelet energy distribution for positive ground faults occurring 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 distributions for negative ground faults occurring 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 This is a flowchart of the Att-GRU prediction model training process of the present invention;

[0081] Figure 6 This is a topology diagram of the islanded ring DC microgrid studied in this invention.

[0082] Figure 7 For this invention, p = 0.65, R f =4,T pre Predicted data 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 The following diagram illustrates the positioning error of the BES3 test data under different fault conditions: (a) PP; (b) PPG; (c) NPG. Detailed Implementation

[0084] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0085] The technical solution adopted in this invention is:

[0086] A fault location method for a ring DC microgrid based on local data driving mainly includes two modules: fault modeling and fault location. Each module consists of three steps, as shown in the appendix. Figure 1 As shown:

[0087] Step 1) Construction of Fault Sample Set

[0088] A ring-shaped DC microgrid simulation model was built using MATLAB / Simulink and the OPAL-RT4510 real-time simulator to simulate line faults under various conditions, including different fault locations, fault resistances, fault types, and load sizes. The positive current and voltage on the converter bus side were collected before and after the fault and compiled into an N*S raw dataset, where N is the length of the sampled data, which depends on the sampling frequency and the size of the time window; and S is the number of simulation cases considered.

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

[0090] The training datasets required for the Att-GRU prediction model and the ELM-Adaboost localization model are constructed based on different data processing methods: For the prediction model, the data is divided into multiple subsequences in chronological order using the variable-length sliding window method, with the data from earlier in the chronological order used as training input and the data from later in the chronological order used as training labels; For the localization model, multiple sub-data sets containing voltage and current at different times before and after the fault are constructed using the fixed-length sliding window method, and then feature vectors are constructed based on discrete wavelet decomposition as training input for the localization model, with the relative position of the fault point in the line used as training labels.

[0091] Because the voltage and current data prior to a fault contain limited information about the fault location and its changing trends, this invention employs a sliding window method in the construction of the model training dataset to reduce the impact of useless pre-fault data on the performance of the prediction and localization models. For example... Figure 2 As shown, w new datasets containing different sampling time periods are constructed using w sliding windows. The sliding step size n is adjusted accordingly. s and window length n w This allows control over the number and overlap of generated subsets. The number of windows can be calculated as follows:

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

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

[0094]

[0095] In the above formula, i = 0, 1, 2, ..., Nn w .exist Figure 2 In this example, assuming the fault occurs at t = 0.5s and is isolated at 0.5015s, the original sampled voltage sequence contains data from [0.499, 0.5015]s. If w = 3, n... w =2ms,n s =0.25ms. After sliding window processing, three data windows can be obtained: [0.499, 0.501], [0.49925, 0.50125], and [0.4995, 0.5015], generating three new subsets. In practical applications, the window size n is determined according to specific requirements. w The values ​​can be fixed or variable, thus generating multiple subsets of the same or different sizes. Using the sliding window method can increase the amount of available data during training, reduce the impact of sampling uncertainty, and improve the model's generalization ability.

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

[0097] Using the aforementioned 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, the process continues until the prediction error no longer shows significant improvement. For the ELM-Adaboost model, the internal weights of each Extreme Learning Machine (ELM) weak locator are calculated based on the localization 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.

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

[0099] To meet the requirements of diagnostic speed, the sampling data before fault isolation may be insufficient for fault location. To expand the input information required for the location model with limited samples, this invention proposes a fault prediction model based on Att-GRU to capture the changing trend of fault signals and predict fault current and voltage data for a period of time after fault isolation.

[0100] The process of establishing the dual-hidden-layer Att-GRU model used in this invention is as follows: Figure 3As shown, N1 and N2 represent the number of nodes in the two layers, respectively. This section only presents the fault voltage prediction modeling method; the current prediction model is established similarly. First, bus-side voltage data of the converter before and after a line fault are collected based on simulation experiments. Different fault locations and fault resistances are considered during the simulation to obtain more usable information. Second, the input samples and corresponding prediction labels of the neural network are constructed based on the variable-length sliding window method, resulting in training datasets 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 model training, the voltage after fault isolation can be predicted based on the real-time sampled voltage before fault isolation. The enhanced dataset can be used for subsequent feature extraction and localization. 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 neurons to zero with a certain probability during the forward propagation computation phase of the neural network, the influence of local features on the prediction results can be avoided, and overfitting during training can also be prevented.

[0102] Step 5) Fault Feature Extraction

[0103] Discrete wavelets are used to extract features from the enhanced data. Then, the total wavelet energy, relative wavelet energy, and relative wavelet energy entropy are calculated to construct a localization feature vector, so as to solve the problem of accurate localization under weak fault characteristics.

[0104] Performing Discrete Wavelet Transform (DWT) on the voltage and current signals acquired after a fault results in a set of wavelet coefficients. Directly using these coefficients as input to the localization model would consume significant storage space and increase computational costs. Therefore, this invention constructs a feature vector based on the Relative Wavelet Energy (RWE) from the DWT decomposition results. If the signal, after DWT, yields K detail coefficients at decomposition level j, then the wavelet energy E of the signal at decomposition level j is... j It can be defined as:

[0105]

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

[0107]

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

[0109]

[0110] The relative wavelet energy can be used to quantify the differences in the energy distribution of a signal at different scales, thereby screening 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 the 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 separately and then calculating the corresponding feature quantities helps to reduce the dimension of the original data while retaining 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 the discrete wavelet transform, the mother wavelet function ψ(t) and the maximum decomposition level M are two important parameters that affect the extraction effect. Currently, 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 a DC microgrid containing a large number of transient voltage and current signals, this chapter selects the db wavelet for analysis. The maximum decomposition level is another important parameter that affects the feature extraction effect. A larger decomposition level can挖掘 more detailed information contained in the signal, but it will increase the computational 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 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 different fault cases to consider, and the sampling data for each case is data after the fault at 1ms, 1.5ms, and 2ms, then the values ​​of N are 21280, 31920, and 42560, respectively. According to Equation (8), the corresponding maximum decomposition levels after rounding are 10, 11, and 11, respectively. The frequency bands corresponding to different decomposition levels when M=11 are shown in Table 2.

[0118] Table 1 shows the filter lengths of commonly used mother wavelet in the db family.

[0119]

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

[0121]

[0122] To verify the effectiveness of the feature extraction method based on discrete wavelet transform, it is assumed that at t = 0.5s, a fault occurred at different locations (p = 0.5, 0.9) on line 56 with a resistance of R. f =5Ω positive-pole-to-ground fault and negative-pole-to-ground fault. Voltage and current of BES3 within 1.5ms after the fault are sampled at a sampling frequency of 20kHz. Feature extraction is performed using the db8 wavelet and decomposition level 11. The relative wavelet energy of positive-pole-to-ground (PPG) and negative-pole-to-ground (NPG) faults at different fault locations is as follows: Figure 3 and Figure 4 As shown, when a PPG fault occurs, the relative wavelet energy of the voltage is quite similar at different fault locations, while the relative wavelet energy of the current differs significantly, which is related to the drastic current changes under a positive ground fault. Therefore, combining the relative wavelet energy of current and voltage is more beneficial for fault location.

[0123] When an NPG fault occurs, the relative wavelet energy of the current is relatively similar at different fault locations. Although the relative wavelet energy of the voltage differs, it is not as significant as the difference in the relative wavelet energy (RWE) of the current under a PPG fault. Therefore, constructing the eigenvector shown in Equation (6) by combining the total wavelet energy (ET) and the relative wavelet energy entropy (RWEE) helps to amplify the voltage differences at different fault locations and improve the positioning accuracy.

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

[0125] ELM (Elastic Localizer Model) has many advantages, but using only a single ELM model for localization may lead to overfitting or underfitting, affecting the reliability of the localization results. Therefore, this chapter improves ELM by incorporating the Adaboost algorithm. The Adaboost algorithm is an ensemble learning method that continuously trains T weak localizers and dynamically adjusts the sample weights and weak localizer weights based on the error of each training iteration. If the error exceeds a threshold θ... e Then, more weight is assigned to samples with larger errors and weaker classifiers with better localization performance. After all weak localizers have been trained for T rounds, a strong localizer can be obtained using the Adaboost algorithm. The detailed steps are as follows:

[0126] (1) Data selection and sample weight initialization: Randomly select m sets of training data from the input sample data, and initialize the t-th weak localizer ELM. t The sample weights 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) Training and localization error calculation of the weak locator in this round: The weak locator is trained using samples with initialized weights. For the t-th weak locator, the localization error is calculated as follows:

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

[0130] In equation (10), p esti For weak positioner ELM t Estimated fault location, p acti This indicates the actual location of the fault.

[0131] (3) Calculate the weights of the weak locators in this round: Calculate the weight coefficient α of the weak classifier t based on the positioning error. t Give greater weight to weak locators that have better positioning performance:

[0132]

[0133] (4) Update the training sample weights for the next round: Adjust the weight D(t+1) of the (t+1)th weak localizer training sample so that samples that were not correctly localized in this round will be given more importance in the next round of training. The sample weights are calculated as follows:

[0134]

[0135] In formula (12) This is the normalization factor. Because a t >0, therefore, 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.

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

[0137]

[0138] To verify the effectiveness of the proposed localization method, it was tested in MATLAB / Simulink. Figure 6 A fault simulation model of the ring-shaped DC microgrid was established, using the same simulation parameters shown in Table 3. For different parts of the proposed fault location method, corresponding parameters 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 N1, N2, and N in the GRU hidden layer and ELM are specified. ELM and the positioning error threshold θ e The step size was determined through trial and error. The sliding window step size was 5 samples, and at a sampling frequency of 20kHz, the sliding step size was 0.25ms.

[0139] Table 3 shows the different parameters of the proposed positioning method.

[0140]

[0141] Table 4. Fault simulation case configuration for each line.

[0142]

[0143]

[0144] For each fault type (PP, PPG, NPG) on each line (lines 1-6), simulations were performed under different load conditions to obtain data for different fault resistances and fault locations. The cases considered in the simulation analysis are shown in Table 4. For each line fault type, 1064 cases were generated, totaling 19152 cases for all three fault types and six fault lines. Fault location is represented as the relative distance between the fault point and the end of the line in a counter-clockwise direction. For example, for fault F1, p = 0.2 indicates the fault is near the photovoltaic end. Bus-side voltage and current were sampled for 3 ms before and after the fault (1 ms before and 2 ms after the fault) as the raw dataset and then divided. 70% of the training data, after sliding window partitioning and feature extraction, was used to train the prediction and localization models, while 30% of the test data was used to test the data augmentation effect and localization accuracy.

[0145] Example 1:

[0146] Table 5 shows the different time intervals used for data augmentation.

[0147]

[0148] This embodiment illustrates the data augmentation effect. Considering different faults occurring on line 34 with a length of 1.4km, based on the diagnostic methods proposed in the previous two chapters, PP / PPG and NPG faults can be diagnosed and isolated within 1-2ms. Therefore, the analysis considers the three different prediction time intervals shown in Table 5, where T... pre The symbols indicate time intervals; "-" indicates the time before the fault, and "+" indicates the time after the fault.

[0149] The voltage and current variations are significant under different fault resistors. To mitigate the impact of numerical values ​​on the evaluation metrics, a normalized root-mean-square error (NRMSE) is defined for the test data i to evaluate the prediction accuracy.

[0150]

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

[0152] Table 6 shows the voltage and current prediction errors under different scenarios, where p acti This represents the actual fault location. To more intuitively demonstrate the prediction accuracy, Figure 7 The text presents prediction results for one specific scenario. Figure 7 As shown in Table 6, the NPG fault current prediction error is relatively large under time interval T1, but high prediction accuracy can be achieved under other scenarios. The above results indicate 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 dataset. 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 rapidly when PP and PPG fail, as shown in the prediction results under time interval T1, data including 0.5ms after the fault occurs is sufficient to train the prediction model.

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

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

[0159] (4) The prediction error decreases as the length of the post-fault data increases, which suggests that the post-fault data contains time information about voltage and current changes.

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

[0161] Based on the above analysis, when there is insufficient sampling data 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, thereby meeting the requirements of prediction accuracy and diagnosis speed.

[0162] Table 6 Prediction Errors (NRMSE) under Different Scenarios i

[0163]

[0164]

[0165] Table 7. Total Prediction Error (NRMSE) of the Test Dataset under Different Scenarios tol

[0166]

[0167] Example 2:

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

[0169] Table 8 Location error during line 56 fault

[0170]

[0171]

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

[0173]

[0174] Table 9 shows the MLE and MAPLE under different fault conditions. Based on the above results, the following conclusions can be drawn:

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

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

[0177] (3) Even under NPG faults, the positive current characteristic is weak, and 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] Example 3:

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

[0182] Table 10 MLE and MAPLE of test data under noisy conditions

[0183]

[0184] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A fault location method for a ring DC microgrid based on local data-driven methods, 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; Training datasets for the Att-GRU prediction model and the ELM-Adaboost localization model were constructed using different data processing methods: For the prediction model, the data was divided into multiple subsequences according to time sequence using a variable-length sliding window method, with the data from earlier in the time sequence used as training input and the data from later in the time sequence used as training labels; For the localization model, multiple subsets containing voltage and current at different time periods before and after the fault were constructed using a fixed-length sliding window method, and then feature vectors were constructed based on discrete wavelet decomposition as training input for the localization model, with the relative position of the fault point in the line used as training labels. Step 3) Training the fault prediction and localization model; Step 4) Data augmentation based on the Att-GRU prediction model; The data augmentation based on the Att-GRU prediction model is as follows: A fault prediction model based on Att-GRU is used to capture the trend of fault signal changes and predict fault current and voltage data for a period of time after fault isolation. The process of establishing the dual-hidden-layer Att-GRU model is as follows, where N1 and N2 are the number of nodes in the two layers, respectively. First, based on simulation experiments, bus-side voltage data of the converter before and after a line fault were collected. Different fault locations and fault resistance conditions were considered during the simulation to obtain more usable 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 datasets 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 model training, the voltage 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 localization. To further improve the robustness and generalization ability of the prediction model, a dropout mechanism is added to the attention layer, which sets the output of neurons to zero with a certain probability during the forward propagation computation phase of the neural network. Step 5) Fault feature extraction; Step 6) Fault location model based on ELM-Adaboost.

2. The fault location method for a ring DC microgrid based on local data driving according to claim 1, characterized in that: Step 1), the construction of the fault sample set, is as follows: A ring-shaped DC microgrid simulation model was built using MATLAB / Simulink and the 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 of the converter bus side were collected before and after the fault and compiled into an N*S raw dataset, where N is the length of the sampled data, which depends on the sampling frequency and the size of the time window; and S is the number of simulation cases considered.

3. The fault location method for a ring DC microgrid based on local data drive according to claim 1, characterized in that: Step 2) Fault data preprocessing based on sliding window is as follows: We construct w new datasets containing different sampling time periods using w sliding windows, by adjusting the sliding step size n. s and window length n w The number of generated subsets and their degree of overlap can be controlled. The number of windows is calculated as follows: （1） Where N represents the time length contained in the sampled data, which depends on the sampling frequency. The subset of data generated by the fault voltage data processing is represented as follows: （2） In the above formula, i = 0, 1, 2,…N-n w .

4. The fault location method for a ring DC microgrid based on local data driving according to claim 1, characterized in that: Step 3) Fault prediction and localization model training is as follows: Using the 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 no longer shows significant improvement. For the ELM-Adaboost model, the internal weights of each weak locator in the extreme learning machine are calculated based on the localization 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 fault location method for a ring DC microgrid based on local data driving according to claim 1, characterized in that: Step 5) Fault feature extraction is as follows: Discrete wavelets are 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 a localization feature vector. Based on the discrete wavelet transform decomposition results, an eigenvector is constructed based on the relative wavelet energy. If the signal, after discrete wavelet transform, obtains K detail coefficients at decomposition level j, then the wavelet energy E of the signal at decomposition level j is... j It can be defined as: （3） In equation (3), , The different detail coefficients d at decomposition level j are represented by... j,k At the corresponding position, the total wavelet energy across all decomposition levels can be defined as: （4） Relative wavelet energy The probability distribution representing wavelet energy at different decomposition levels can reduce data size while preserving the original signal characteristics. Its definition is as follows: （5） Relative wavelet energy is used to quantify the differences in energy distribution of a signal at different scales, thereby filtering out important features of the signal. To further improve positioning accuracy, the relative wavelet energy entropy is calculated. This is used to describe the uniformity of energy distribution of a signal at different frequencies. （6） Combining the features based on discrete wavelet transform, the input feature vector of the localization model can be constructed as follows: （7） 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 feature extraction based on discrete wavelet transform, the mother wavelet function The maximum decomposition level M is two important parameters that affect the extraction effect; For the maximum decomposition level M, it is determined by the original signal length N and the mother wavelet filter length F according to the following empirical formula: （8）。 6. The fault location method for a ring DC microgrid based on local data driving according to claim 1, characterized in that: Step 6) based on the ELM-Adaboost fault location model is as follows: (1) Data selection and sample weight initialization: Randomly select m sets of training data from the input sample data, and initialize the t-th weak localizer ELM. t The sample weights are initialized as follows: (9) (2) Training and localization error calculation of the weak locator in this round: The weak locator is trained using samples with initialized weights. For the t-th weak locator, the localization error is calculated as follows: (10) In equation (10), p esti For weak positioner ELM t Estimated fault location, p acti This indicates the actual location of the fault. (3) Calculate the weights of the weak locator in this round: Calculate the weight coefficients of the weak classifier t based on the positioning error. Give greater weight to weak locators that have better positioning performance: (11) (4) Update the training sample weights for the next round: Adjust the weights of the training samples of the (t+1)th weak localizer so that the samples that were not correctly localized in this round will be given more attention in the next round of training. The sample weights are calculated as follows: (12) In formula (12) As the normalization factor, since Therefore, if the i-th sample has a large positioning error in this round, the weight of that sample will be increased in the next round of training. (5) Iterative training and construction of strong classifier: Repeat steps (2)-(4) until all weak localizers have completed T rounds of training. Based on the weighted decision method, the final strong localizer is constructed as follows: (13)。

Images