Power distribution network fault diagnosis method based on hybrid neural network model

Abstract

The invention provides a hybrid neural network model-based power distribution network fault diagnosis method. The method comprises the steps of judging whether a power distribution network is in a normal state or not based on a voltage break variable signal rule; when the power distribution network is in an abnormal state, three-phase voltage signals with the fault moment delayed by three cycles are collected, and the sampling rate is unified; converting the three-phase voltage signal with the unified sampling rate into a time sequence, and performing image coding on the time sequence by using a Markov transfer field; and constructing and training a hybrid neural network fault diagnosis model, and monitoring the power distribution network in real time by using the trained diagnosis model so as to diagnose the fault type of the power distribution network. According to the method, the two-dimensional coding image features and the time sequence features of the three-phase voltage signals can be extracted at the same time, and the gating vector is introduced to perform weighted summation on the features extracted by each channel, so that the problem that feature information extracted by a single training model is incomplete is solved, higher fault diagnosis accuracy is realized, and the fault diagnosis capability of the power distribution network is improved.

Description

Technical Field

[0001] The invention relates to the technical field of distribution network fault diagnosis, and in particular to a distribution network fault diagnosis method based on a hybrid neural network model. Background Art

[0002] As the transmission link of the power system, the distribution network is responsible for delivering electric energy to the user safely and reliably. However, due to the complexity of its structure, the diversity of its equipment and the variability of its operating environment, the distribution network frequently fails, which has a serious impact on the stability of the power system and the quality of electricity consumption of users. Therefore, the development of distribution network fault diagnosis technology is an important means to improve the safe and reliable operation of the power system.

[0003] Traditional methods for fault diagnosis of distribution networks typically include empirical judgment methods: distribution management personnel judge the operating status and fault status based on experience and line history; expert system methods: using computer technology to integrate theoretical knowledge and expert experience, and organically connecting multiple components to identify faults; methods based on fuzzy theory: in power systems, due to the influence of various factors, fault diagnosis is uncertain, and fuzzy theory is good at simulating human thinking, deriving multiple possible solutions and sorting them to find the most correct solution. However, the above methods are all subjective, which affects the accuracy of diagnosis, and have poor fault tolerance when dealing with special situations.

[0004] At present, the fault diagnosis technology for distribution network systems is constantly developing. Its main methods include intelligent power operation and maintenance cloud platform monitoring and diagnosis methods based on time series convolutional residual networks. At the same time, artificial intelligence and machine learning technologies have also made significant progress in the field of distribution network fault diagnosis, which automatically identifies and classifies faults through training models. However, most of them use two-dimensional images or time series alone for model training and fault diagnosis, without combining the two, resulting in a relatively low fault diagnosis accuracy. Combining the two and prejudging the distribution network operation status can greatly improve the accuracy of distribution network fault diagnosis. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a distribution network fault diagnosis method based on a hybrid neural network model.

[0006] The technical solution of the present invention is: a distribution network fault diagnosis method based on a hybrid neural network model, comprising the following steps:

[0007] S1) monitor the three-phase voltage of the distribution network bus in real time, and determine whether the distribution network is in a normal state based on the voltage mutation signal law;

[0008] S2) When it is determined that the distribution network is in an abnormal state, a three-phase voltage signal delayed by 3 cycles from the fault time is collected at the distribution network bus, and the sampling rate is unified;

[0009] S3), converting the three-phase voltage signal with a unified sampling rate in step S2) into a one-dimensional voltage time series, and using a Markov transfer field to perform image encoding on the one-dimensional voltage time series to obtain a two-dimensional encoded image;

[0010] S4), constructing a hybrid neural network fault diagnosis model based on convolutional neural network CNN and gated recurrent unit GRU, and training it using one-dimensional voltage time series and two-dimensional coded images;

[0011] S5), the real-time monitored voltage signal is processed in steps S1)-S3) in sequence and then input into the trained hybrid neural network fault diagnosis model to diagnose the distribution network fault type.

[0012] Preferably, in step S1), real-time monitoring of the three-phase voltage of the distribution network busbar specifically includes the following steps:

[0013] S11), real-time detection of the three-phase voltage at a certain point of the distribution network bus;

[0014] S12), calculating the voltage mutation amount within 3 cycles before the current detection time t, and obtaining the voltage mutation amount matrix;

[0015] S13), transforming each column of the voltage mutation matrix to obtain a positive-sequence voltage mutation signal matrix, and calculating a fault monitoring start criterion based on the positive-sequence voltage mutation signal matrix; determining whether the distribution network is in a normal operating state by the start criterion.

[0016] As a preferred embodiment, in step S11), taking the current detection time t as the reference, the interception time window is [tt o , t], with a sampling rate of f h The three-phase voltage signal within the time window is collected; and according to the time window [tt o The three-phase voltage signals A, B, and C collected within t] constitute a three-phase voltage signal matrix T1.

[0017] As a preferred embodiment, in step S12), the time window [tt o , t] is divided into two time windows [tt o / 2, t] and [tt o ,tt o / 2]; and extract the voltage signal values ​​in the above two time windows from the three-phase voltage signal matrix T1 to form matrices T2 and T3 respectively, and make a difference between matrices T2 and T3 to obtain the time range t before the current detection time t o / 2 voltage mutation matrix T4.

[0018] Preferably, in step S13), each column of the voltage mutation matrix is ​​transformed to obtain a positive-sequence voltage mutation signal matrix, specifically:

[0019]

[0020] In the formula, are the voltage mutations of phases A, B, and C at the detection position of the distribution network bus at time t1; where the value range of t1 is [tt o / 2, t]; α is the rotation factor.

[0021] Preferably, in step 13), the calculation formula of the startup criterion ΔV of the fault monitoring is:

[0022]

[0023] In the formula, is the number of sampling points; for The element of the positive sequence voltage mutation matrix at the moment; f h is the actual sampling rate;

[0024] If ΔV ≥ ΔV set , the distribution network is in an abnormal operating state, and the current detection time is taken as the fault occurrence time t s , otherwise, the distribution network is in normal operation.

[0025] Preferably, in step S2), during the sampling process, the actual sampling frequency f is determined. h With the expected sampling frequency f s The size relationship of

[0026] If, f h <f s , then upsampling is performed,

[0027] If, f h =f s , no processing is required;

[0028] If, f h >f s , then downsampling is performed.

[0029] Preferably, in step S2), when the actual sampling frequency fh Less than the expected sampling frequency f s The sampling rate is increased by using a cubic spline interpolation method, specifically, the original sampling points are divided into multiple sub-intervals according to every two adjacent sampling points, and a cubic polynomial is set in the form of a difference between the midpoints of each sub-interval.

[0030] Preferably, in step S2), when the actual sampling frequency f h Greater than the expected sampling frequency f s Down sampling is achieved by using the decimation method.

[0031] Preferably, in step S3), the voltage value of each phase corresponding to each time point in the matrix form of the voltage signal of the sampling point after the sampling rate is unified is recorded, and each group of sequences is standardized and scaled to obtain the time series of three groups of voltages of the three phases A, B, and C.

[0032] Preferably, in step S3), the state transition probability of two adjacent time steps of the time series is calculated, a Markov state transfer matrix is ​​constructed, and the imagesc function is used to convert the element values ​​in the Markov state transfer matrix into different colors according to their sizes, and the corresponding positions on the coordinate axis are colored with corresponding colors to obtain a two-dimensional coded image of the three-phase voltage of the distribution network bus line.

[0033] Preferably, in step S4), the mathematical model of the convolutional neural network CNN is expressed as:

[0034]

[0035] In the formula, are three sets of two-dimensional coded images as input; is the output of the CNN branch; Calculated for the convolutional layer; Calculated for batch normalization layer; Calculated for activation layers; Calculated for the maximum pooling layer; It is the fully connected layer operation; Flatten layer operation.

[0036] Preferably, in step S4), the mathematical model of the gated recurrent unit GRU is:

[0037]

[0038] In the formula, are the three sets of time series input; is the output of the GRU branch; Calculated for flattened layers; Calculated for the GRU layer.

[0039] Preferably, in step S4), the calculation expression of the hybrid neural network fault diagnosis model is:

[0040]

[0041] In the formula, is the output of the GRU branch; is the output of the CNN branch; x output is the output of the hybrid neural network fault diagnosis model.

[0042] The beneficial effects of the present invention are:

[0043] 1. The present invention uses a method to determine whether the distribution network is in a normal state based on the law of voltage mutation signal to monitor the operating state of the distribution network in real time. Only when it is determined to be in an abnormal state, the next step of diagnosis is performed, thereby simplifying unnecessary workload and achieving a good preliminary judgment of the fault;

[0044] 2. The hybrid neural network fault diagnosis model constructed by the present invention can simultaneously extract the two-dimensional coded image features and time series features of the three-phase voltage signal, and introduce a gating vector to further weighted sum the features extracted from each channel, which can effectively solve the problem of incomplete feature information extracted by a single training model, achieve higher fault diagnosis accuracy, and greatly improve the fault diagnosis capability of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the process of the method of embodiment 1 of the present invention;

[0046] Figure 2 This is a circuit diagram of a power distribution network according to Embodiment 1 of the present invention;

[0047] Figure 3 A framework diagram of a hybrid neural network fault diagnosis model constructed in Example 1 of the present invention;

[0048] Figure 4 A schematic diagram of the criteria for starting fault detection in Example 1 of the present invention;

[0049] Figure 5 Schematic diagram of three-phase voltage signals in different short-circuit states in Example 1 of the present invention;

[0050] Figure 6 Schematic diagram of a two-dimensional coded image of three-phase voltages at three groups of distribution network buses M0 in Embodiment 1 of the present invention;

[0051] Figure 7 A schematic diagram of the criteria for starting fault detection in Embodiment 2 of the present invention;

[0052] Figure 8 Schematic diagram of three-phase sampling signals in Embodiment 2 of the present invention;

[0053] Fig. 9 It is a schematic diagram of the MTF two-dimensional coded image when a single phase is grounded in Example 2 of the present invention. DETAILED DESCRIPTION

[0054] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0055] Example 1

[0056] like Figure 1 As shown, this embodiment provides a distribution network fault diagnosis method based on a hybrid neural network model. The circuit diagram of the distribution network provided in this embodiment is shown in FIG. Figure 2 As shown in the figure, the three-phase voltage signal measurement point is at M0 of the busbar, and the short-circuit fault point is set at M1-M 14 Place.

[0057] The fault diagnosis method specifically comprises the following steps:

[0058] S1) real-time monitoring of the three-phase voltage of the distribution network bus, and judging whether the distribution network is in a normal state based on the voltage mutation signal; specifically comprising the following steps:

[0059] S11), real-time detection of the three-phase voltage at the distribution network bus M0; this embodiment takes the current detection time t as the reference, and the interception time window is [tt o , t], with a sampling rate of f h Acquisition time window [tt o , t] constitutes a three-phase voltage signal matrix T1, where the three-phase voltage signal matrix T1 will be updated in real time as time goes by. The three-phase voltage signal matrix T1 is expressed as:

[0060]

[0061] In the formula, r+1 is the number of sampling points contained in the time window; They are the voltage signals of the three-phase voltages A, B and C at the bus measured at time t; The three-phase voltages of A, B, and C at the busbar are o The voltage signal measured at all times;

[0062] In this embodiment, the frequency of the three-phase voltage signal is set to f d In this embodiment, t is set o for The update time window is The number of sampling points is

[0063] S12), calculating the voltage mutation amount within 3 cycles before the current detection time t, and obtaining the voltage mutation amount matrix;

[0064] In this embodiment, the time window of the sampling signal is and divide it into and two time periods; then the voltage signal values ​​in the above two time windows are extracted from the three-phase voltage signal matrix T1 to form matrices T2 and T3 respectively.

[0065]

[0066] Then, the matrix T2 and T3 are processed by difference to obtain the matrix T4, which is the voltage mutation amount within 3 cycles before the current detection time t, as follows:

[0067] T4 = T2 - T3 (4);

[0068]

[0069] In the formula, They are the voltage mutations of the three phases A, B and C at the bus at the current detection time t; For time window The number of sampling points contained in.

[0070] S13), transforming each column of the voltage mutation matrix T4, and determining whether the distribution network is in a normal operating state according to the criterion;

[0071] Each column of the voltage mutation matrix T4 is transformed as follows to obtain a matrix T5, ​​where the matrix T5 is a matrix composed of positive-sequence voltage mutation signals:

[0072]

[0073] Among them, U1 is the time window The positive sequence voltage mutation after internal transformation; α is the rotation factor; the value range of t1 is

[0074] The final matrix T5 is expressed as:

[0075]

[0076] Where V t is the positive sequence voltage mutation at time t; is the positive sequence voltage mutation at the moment before time t; is the positive sequence voltage mutation at the last sampling point within three cycles, and the number of sampling points is

[0077] The criterion is: if ΔV ≥ ΔV set , the distribution network is not in normal operation, and the current detection time is recorded as the fault occurrence time t s , otherwise, the distribution network is in normal operation; ΔV set is the default value.

[0078] in:

[0079]

[0080] In the formula, The value range is express The positive sequence voltage mutation signal at the moment;

[0081] In this embodiment, the frequency of the power distribution network is 50 Hz, and the time length of three cycles is 0.06 s. Figure 4 Shown is a schematic diagram of the start criteria.

[0082] S2) When the distribution network is judged to be in an abnormal state, the fault time t is collected. s Delay the three-phase voltage signal by 3 cycles and unify the sampling rate of the sampling signal;

[0083] The sampling rate of the collected signal in this embodiment is unified to 50kHz, denoted as f s ; From the moment of positive sequence voltage mutation t s With the initial sampling rate f h Delay the acquisition of three-phase voltages for 3 cycles to form a matrix V h :

[0084]

[0085] Where n+1 is the initial sampling rate at t s The total number of sampling points within 3 cycles of time delay, and The distribution network bus is at t s The three-phase voltage signal measured at all times, The distribution network bus is at t s The three-phase voltage signal measured at the next sampling point after the moment; It is the three-phase voltage signal measured at the last sampling point of the distribution network bus within three cycles;

[0086] Suppose the matrix form of the voltage signal at the sampling point after the sampling rate is unified is V s :

[0087]

[0088] Where m+1 is the total number of sampling points within 3 cycles after the sampling rate is unified, and The distribution network bus is at t s The three-phase voltage signal measured at all times, The distribution network bus is at t s The three-phase voltage signal measured at the next sampling point after the moment; It is the voltage signal measured at the last sampling point of the distribution network bus within 3 cycles.

[0089] First, determine the relationship between the actual sampling frequency and the expected sampling frequency:

[0090] If, f h <f s , then upsampling is performed, that is, when the actual sampling frequency is less than the expected sampling frequency, this embodiment uses a cubic spline interpolation method to increase the sampling rate, as follows:

[0091] The original sampling points are divided into multiple sub-intervals according to each two adjacent sampling points, and the midpoint of each sub-interval is used as the difference. The basic form of the cubic polynomial is:

[0092]

[0093] In the formula, a i 、b i 、c i ,d i is the unknown coefficient, which is solved by the boundary conditions, h i (x) represents the sample point x i and sample point x i+1 A cubic polynomial in x i is the horizontal coordinate value of the i-th sample point; the value range of i is 1, 2, …, n.

[0094] The boundary conditions are as follows:

[0095] 1) The endpoint of the curve passes through the data point

[0096] h i (x i )=y i ,h i (x i+1 )=y i+1(13);

[0097] In the formula, y i is the signal value at time i under the initial sampling rate, y i+1 is the signal value at time i+1 at the initial sampling rate.

[0098] 2) Continuity of adjacent curve endpoints

[0099] h i (x i+1 )=h i+1 (x i+1 ) (14);

[0100] 3) Continuity of first-order derivatives of adjacent curve endpoints

[0101] h i ′(x i+1 )=h i ' +1 (x i+1 ) (15);

[0102] 4) Continuity of second-order derivatives of adjacent curve endpoints

[0103] h″ i (x i+1 )=h″ i +1(x i+1 ) (16);

[0104] When solving the cubic spline interpolation function h(x), there are n+1 data points, n intervals, and n cubic polynomials, so there are 4n unknowns in total. Using the tridiagonal matrix algorithm to solve, the undetermined coefficients of the cubic polynomial are obtained, and then sample points are inserted in each area to obtain the sequence V after the sampling rate is unified. s .

[0105] If, f h =f s , no processing is required;

[0106] If, f h >f s , downsampling is performed, that is, when the actual sampling frequency is greater than the expected sampling frequency, this embodiment uses a decimation method to implement downsampling, and the specific operations are:

[0107] According to the interval of every D original sample points, a total of k groups of voltage matrices are divided, where k and D satisfy:

[0108]

[0109] Where D is the extraction factor; f h is the actual signal frequency; f sis the expected sampling rate;

[0110]

[0111] Start from the middle sample point and expand to both sides. Sample points, the collected sample points are sequentially formed into a new sequence V s

[0112]

[0113] In this embodiment, the frequency of the power distribution network is 50 Hz, and the time length of three cycles is 0.06 s. Through the processing of step S2), the sampling rate is unified to 50 kHz, and the total number of sampling points is 3000. Figure 5 The figure shows the three-phase voltage signals (with unified sampling rates) collected at the distribution network bus when different short-circuit faults occur in the distribution network.

[0114] S3), converting the three-phase voltage signal after the sampling rate is unified in step S2) into a one-dimensional voltage time series, and using the Markov transfer field to perform image encoding on the voltage time series to obtain a two-dimensional encoded image; specifically:

[0115] S31), converting the three-phase voltage signal after the sampling rate is unified into a one-dimensional voltage time series;

[0116] According to the sampling frequency f set in step S2) s The obtained sampling signal uses discrete digital variables to represent the continuous voltage analog signal: V s For each time point t p Record the voltage value of each phase and get:

[0117]

[0118] Where p is the index value of the sampling point, and the value range of p is 0, 1, 2, 3, ..., m; X A (t), X B (t), X C (t) are the time series of the three-phase voltages of A, B, and C respectively; V A (t0), V A (t0), V A (t0) are the voltage amplitudes of phases A, B, and C at the first sampling point at the busbar of the distribution network.

[0119] Then, each set of the above sequences is standardized and scaled to [-1, 1] according to the following formula:

[0120]

[0121] Where p ranges from 0, 1, 2, 3, ..., m, x p For X A (t), X B (t), X C (t) The index values ​​of the three time series; max(X) is X A (t), X B (t), X C (t) The maximum value of each of the three time series; min(X) is X A (t), X B (t), X C (t) The minimum value of each of the three time series; is the normalized scaling value;

[0122] After the above processing, the time series of the three groups of voltages of three phases A, B, and C are obtained as follows:

[0123]

[0124] Where m+1 is the total number of sampling points within 3 cycles after the sampling rate is unified.

[0125] S32), using the Markov transition field to perform image encoding on the three groups of voltage time series X(t) obtained in S31), and converting them into a two-dimensional coded image;

[0126] S321), construct state transfer matrix

[0127] Divide each group of voltage time series obtained in step S31) into O non-overlapping regions according to the amplitude, where O represents the number of possible states in the time series data; where m+1=O×O;

[0128] Each region of the result after the time series of each group of voltage is divided is uniformly represented as follows:

[0129] O σ =[x1 x2 x3 … x O ] 1×O (25);

[0130] Wherein, the value range of σ is 1, 2, 3, ..., O.

[0131] Change each data of each region to the serial number of its corresponding region. For each region, calculate the state transition probability of two adjacent time steps and record these probability values ​​as elements w εθ , and use w εθ Construct an O×O dimensional Markov state transfer matrix W, as shown in formula (26):

[0132]

[0133] Among them, w εθ represents the probability of transitioning from state ε to state θ;

[0134] To calculate w εθ First, construct a transfer count zero matrix C1 of size O×O, as shown in formula (28), which stores the number of voltage state transitions under different states; next, loop through each region, for each region, obtain the number of transitions from the current state to other states, and continuously update the transfer count matrix; at the same time, The corresponding probability of the state ε in the state transfer matrix is ​​set to 0; finally, as shown in formula (29), the matrix C1 is updated by counting the number of voltage state transitions in each region to obtain the matrix C2, and then these transition times are calculated and normalized to obtain the voltage transition probability matrix;

[0135]

[0136] In the formula, C εθ is the number of times the ε state transitions to the θ state.

[0137] S33), calculate the two-dimensional coded image of the three-phase voltage of the distribution network line

[0138] Use the imagesc function to convert the element values ​​in the state transfer matrix W into different colors according to their size, and color them with this color at the corresponding position on the coordinate axis; each element w in the state transfer matrix W εθ The following color mapping rules (YGB) are met: the brightness of each color of yellow, green and blue is gradually increased from level 0 to 255, and different proportions of color mixing are achieved to form corresponding colors, that is:

[0139] If εθ ≤0.5, then: YGB1=[ω εθ ×255,2ω εθ ×255,(1-ω εθ )×255] 1×3 ;

[0140] If εθ >0.5, then: YGB2=[ω εθ ×255,(-2ω εθ +2)×255,(1-ω εθ )×255] 1×3 ;

[0141] Finally, the corresponding colors are obtained by mixing the colors of the above-obtained matrix and displayed at the corresponding coordinate position of the image (ε is the horizontal coordinate and θ is the vertical coordinate). When four short-circuit faults occur in the distribution network, the following are obtained: Figure 6 The two-dimensional encoded images of the three-phase voltages at the three distribution network busbars M0 are shown.

[0142] S4), using the one-dimensional voltage time series and the two-dimensional coded image to train a hybrid neural network fault diagnosis model based on the convolutional neural network CNN and the gated recurrent unit GRU;

[0143] In this embodiment, the hybrid neural network fault diagnosis model is trained using the three groups of one-dimensional voltage time series obtained in step S31) and the two-dimensional coded image in step S33) as training data sets;

[0144] This embodiment builds a distribution network model based on the MATLAB / Simulink platform. s The different short-circuit states in the three-phase fault module are constantly changed to simulate the four abnormal states of three-phase short circuit, two-phase short circuit, single-phase grounding and two-phase grounding in the distribution network in actual operation.

[0145] For each short circuit fault state, the transition resistance and initial phase angle are set within a certain range, as shown in the following table.

[0146] Table 1 Parameter configuration

[0147]

[0148] The specific process of sample collection is as follows: first, according to step S2), the three-phase voltage signal of the distribution network bus delayed by 3 cycles at the time ts when the fault occurs is collected, and the sampling rate is unified to 50kHz; then, according to step S3), 3 groups of time series of voltage signals and 3 groups of two-dimensional coded images of voltage signals are generated, wherein 3 groups of time series and 3 groups of two-dimensional images constitute a group of sample data. For each fault state set on each branch, 50 groups of samples are collected at the distribution network bus M0, each group of samples corresponds to 3 groups of time series and 3 groups of two-dimensional coded images, each image has a dimension of 200×200×3, and the time series is a one-dimensional sequence of 1×3000. In this embodiment, each branch has 4 different short-circuit fault states, and there are 14 branches in total, so a total of 2800 groups of samples, 8400 two-dimensional coded images, and 8400 groups of time series are finally generated.

[0149] The convolutional neural network CNN of this embodiment is provided with three channels, which are respectively used to process three groups of voltage two-dimensional coded images; each channel is provided with an image input layer, a convolution layer, a batch normalization layer, an activation layer, a maximum pooling layer, a first fully connected layer, and a flattening layer; wherein the convolution layer uses a 3x3 convolution kernel, 64 filters, and a step size of 2; the batch normalization layer is used to accelerate training and improve model performance; the activation layer is used to introduce nonlinearity, and the activation function is ReLU; the maximum pooling layer uses a 3x3 pooling window with a step size of 2; the padding is 'same'; the first fully connected layer has 64 units; the flattening layer flattens the multi-dimensional input into one dimension. Its output is three groups of features of three groups of two-dimensional coded images, and the mathematical model of the convolutional neural network CNN is expressed as:

[0150]

[0151] In the formula, are three sets of two-dimensional coded images as input; is the output of the CNN branch; Calculated for the convolutional layer; Calculated for batch normalization layer; Calculated for activation layers; Calculated for the maximum pooling layer; It is the fully connected layer operation; It is the flattening layer operation;

[0152] The gated recurrent unit GRU has three channels, each of which includes a time series input layer, a flattening layer, and a GRU layer. The GRU layer controls the flow of information by introducing a gating mechanism, retaining relatively important information and removing relatively unimportant information; the flattening layer flattens the multi-dimensional input into one dimension. Its output is three groups of features of three groups of time series; the gating mechanism is a reset gate and an update gate; the mathematical model of the gated recurrent unit GRU is:

[0153]

[0154] In the formula, are the three sets of time series input; is the output of the GRU branch; Calculated for flattened layers; Calculated for the GRU layer.

[0155] The combined part of the hybrid neural network fault diagnosis model includes an attention layer, a second fully connected layer, a Softmax layer, and a classifier, wherein the second fully connected layer has 5 neurons; the Softmax layer converts the output of the second fully connected layer into a probability distribution; and the classifier outputs a predicted category label based on the probability distribution obtained in the previous layer.

[0156] The calculation expression of the hybrid neural network fault diagnosis model is:

[0157]

[0158] In the formula, is the output of the GRU branch; is the output of the CNN branch; x output is the output of the hybrid neural network fault diagnosis model.

[0159] In this embodiment, the attention layer adopts a multi-head attention mechanism, and this embodiment introduces a gate vector τ, whose dimension is set to 6, to adjust the weight ratio of each head, specifically:

[0160] τ=υ(W τ q+b τ ) (34);

[0161] In the formula, υ, W τ 、b τ is the training parameter; q is the query vector.

[0162] In this embodiment, the feature vectors extracted by the CNN branch and the GRU branch are:

[0163]

[0164] In the formula, P1, P2, and P3 are the features of three groups of two-dimensional coded images; P4, P5, and P6 represent three groups of time series features; Respectively represent the feature vectors of the qth group of three groups of two-dimensional coded images A, B, and C; They represent the eigenvectors of the δth group of the three time series of A, B, and C respectively;

[0165] Perform linear transformation on equations (35) and (36), as shown in equations (37)-(39), to map the input data to a new feature space;

[0166] Q=PZ Q =[q1,q2,…,q β ] (37);

[0167] K=PZ K =[k1,k2,…,k γ ] (38);

[0168] V=PZ V =[v1,v2,…,v γ ] (39);

[0169] In the formula, Z Q, Z K , Z V are the linear transformation parameters of query space, key space and value space respectively; Q, K, V are the query space, key space and value space respectively; P represents the characteristics of time series and two-dimensional coded images; q, k, v are the element vectors in query space, key space and value space respectively; β is the number of query vectors and γ is the number of dimensions after linear transformation.

[0170] The similarity between the query and the key is calculated by the scaled dot product model, that is:

[0171]

[0172] Where S(Q,K) is the similarity between the query vector Q and the key vector K; T is the transposition transformation; d k is the dimension of the key vector.

[0173] Apply the Softmax function to convert the above similarity scores into probability distributions, where the attention weights a(q,k i ) is expressed as:

[0174]

[0175] In the formula, is the similarity between the query vector Q and the key vector K; f Softmax It is a probability distribution operation.

[0176] The attention layer calculates the attention weights and performs weighted summation on the values ​​of each head to obtain the corresponding output Y μ , and finally perform the weighted sum of the multi-channels to obtain the output Y, that is:

[0177]

[0178] In the formula, Q, K, V, d k are the dimensions of query space, key space, value space, and key vector respectively; v j is the eigenvector in the value space; r is the number of dimensions after linear transformation; τ μ is the gate vector; μ is the index feature of the 6 outputs of the attention layer; Y μ The output for each channel.

[0179] Finally, the output of the attention layer is processed by the second fully connected layer, and converted into a probability distribution using the Softmax function to obtain the classification output of each fault.

[0180] In this embodiment, 2800 groups of samples are divided into a training set, a test set, and a validation set in a ratio of 7:1.5:1.5, wherein the two-dimensional coded image is sent to the CNN branch and the time series is sent to the GRU branch; the loss function of the hybrid neural network fault diagnosis model uses a cross entropy loss function, and is trained using the Adam optimization algorithm, wherein the number of samples in each small batch is set to 25, the initial learning rate is 0.005, and the total number of iterations is 200 times.

[0181] In each round of training iteration, the validation set is used to evaluate the model to ensure that the model is continuously optimized. After reaching the preset number of iterations and the model converges, the current model is saved. Subsequently, the diagnostic accuracy of the model is evaluated using the test set. If the accuracy is not optimal, the neural network hyperparameters are adjusted and the training step is returned to retrain until the diagnostic accuracy on the test set reaches the standard. When the optimal standard is reached, the model is put into step 5 for online monitoring and diagnosis.

[0182] S5) Using the hybrid neural network fault diagnosis model trained offline to monitor the distribution network in real time, so as to perform online diagnosis of the distribution network fault type, specifically:

[0183] When the distribution network system is actually running, firstly, according to step S1), it is determined whether the distribution network is in a normal state. When it is determined that the distribution network is in an abnormal state, the fault occurrence time t is obtained. s ;

[0184] Then, according to step S2), t is collected at the distribution network bus M0. s Delay the three-phase voltage signal by 3 cycles and unify the sampling rate to 50kHz;

[0185] Then, according to the method of step S3), the time series of the detection voltage signal is obtained and the three groups of voltage time series are converted into three groups of two-dimensional coded images using the Markov transfer field;

[0186] Finally, the time series and two-dimensional coded image obtained in step S3) are sent to the hybrid neural network fault diagnosis model trained in step 4 to diagnose the distribution network fault type, so as to realize real-time monitoring and diagnosis of abnormal status of the distribution network.

[0187] Example 2

[0188] S1), assuming that a single-phase grounding fault occurs at M1 in the distribution network system, through the processing of step S1, the start criterion ΔV is calculated to be 2001.5V, where ΔV set Take 200V and the distribution network frequency as 50Hz, then the time length of 3 cycles is 0.06s. At this time, the start criterion is met, and the time of fault occurrence is t s , go to step S2, such as Figure 7 shown.

[0189] Step S2), collect the fault occurrence time t at the distribution network bus s Delay the three-phase voltage signal by three cycles; then unify the sampling rate to 50kHz; Figure 8 shown.

[0190] Step S3), obtaining the time series of the above-mentioned collected voltage signal, and converting the time series into a two-dimensional coded image through the Markov transition field, wherein the matrix form of the time series is shown in (44);

[0191]

[0192] Finally, the three sets of time series and three sets of two-dimensional coded images are sent to the fault diagnosis model trained offline, and the probability of predicting a single-phase grounding fault is 95.6%. Fig. 9 shown.

[0193] The above embodiments and descriptions are only for illustrating the principles and best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, all of which fall within the scope of the present invention to be protected.

Claims

1. A distribution network fault diagnosis method based on a hybrid neural network model, characterized in that: The steps include: S1) monitor the three-phase voltage of the distribution network bus in real time, and determine whether the distribution network is in a normal state based on the voltage mutation signal law; S2), when it is determined that the distribution network is in an abnormal state, the three-phase voltage signal of the distribution network bus delayed by 3 cycles at the fault time is collected at the distribution network bus, and the sampling rate is unified; S3), converting the three-phase voltage signal with a unified sampling rate in step S2) into a one-dimensional voltage time series, and using a Markov transfer field to perform image encoding on the one-dimensional voltage time series to obtain a two-dimensional encoded image; S4), constructing a hybrid neural network fault diagnosis model based on convolutional neural network CNN and gated recurrent unit GRU, and training it using one-dimensional voltage time series and two-dimensional coded images; S5), the real-time monitored voltage signal is processed in steps S1)-S3) in sequence and then input into the trained hybrid neural network fault diagnosis model to diagnose the distribution network fault type.

2. A distribution network fault diagnosis method based on a hybrid neural network model according to claim 1, characterized in that: In step S1), the three-phase voltage of the distribution network bus is monitored in real time, and based on the voltage mutation signal law, it is determined whether the distribution network is in a normal state, which specifically includes the following steps: S11), real-time detection of the three-phase voltage at a certain point of the distribution network bus; S12), calculating the voltage mutation amount within 3 cycles before the current detection time t, and obtaining the voltage mutation amount matrix; S13), transforming each column of the voltage mutation matrix to obtain a positive-sequence voltage mutation signal matrix, and calculating a start criterion for fault monitoring based on the positive-sequence voltage mutation signal matrix; determining whether the distribution network is in a normal operating state by the start criterion.

3. A distribution network fault diagnosis method based on a hybrid neural network model according to claim 2, characterized in that: In step S11), the current detection time t is taken as the reference, and the time window is [tt o , t], with a sampling rate of f h The three-phase voltage signal within the time window is collected; and according to the time window [tt o The three-phase voltage signals A, B, and C collected within t] constitute a three-phase voltage signal matrix T1.

4. A distribution network fault diagnosis method based on a hybrid neural network model according to claim 3, characterized in that: In step S12), the time window [tt o , t] is divided into two time windows [tt o / 2, t] and [tt o ,tt o / 2]; and extract the voltage signal values ​​in the above two time windows from the three-phase voltage signal matrix T1 to form matrices T2 and T3 respectively, and make a difference between matrices T2 and T3 to obtain the time range t before the current detection time t o / 2 voltage mutation matrix T4.

5. A distribution network fault diagnosis method based on a hybrid neural network model according to claim 4, characterized in that: In step 13), each column of the voltage mutation matrix T4 is transformed as follows to obtain a positive-sequence voltage mutation signal matrix. The specific expression of the transformation is: In the formula, are the voltage mutations of phases A, B, and C at the detection position of the distribution network bus at time t1; where the value range of t1 is [tt o / 2,t]; α is the rotation factor; The calculation formula of the startup criterion ΔV of the fault monitoring is: In the formula, is the number of sampling points; V θ is the element of the positive sequence voltage mutation matrix at time θ; h is the actual sampling rate; If ΔV ≥ ΔV set , the distribution network is in an abnormal operating state, and the current detection time is taken as the fault occurrence time t s , otherwise, the distribution network is in normal operation, where ΔV set is the default value.

6. A distribution network fault diagnosis method based on a hybrid neural network model according to claim 1, characterized in that: In step S2), during the sampling process, the actual sampling frequency f is determined. h With the expected sampling frequency f s The size relationship of If, f h <f s , then upsampling is performed, that is, when the actual sampling frequency f h Less than the expected sampling frequency f s The sampling rate is increased by using the cubic spline interpolation method, specifically, the original sampling points are divided into multiple sub-intervals according to each two adjacent sampling points, and a cubic polynomial is set in the form of a difference between the midpoints of each sub-interval; If, f h =f s , no processing is required; If, f h >f s , then downsampling is performed; that is, when the actual sampling frequency f h Greater than the expected sampling frequency f s Down sampling is achieved by using the decimation method.

7. A distribution network fault diagnosis method based on a hybrid neural network model according to claim 1, characterized in that: In step S3), the voltage value of each phase at each time point in the matrix of the voltage signal of the sampling point after the sampling rate is unified is recorded, and each group of sequences is standardized and scaled to [-1, 1], so as to obtain three groups of voltage time series of the three phases A, B, and C.

8. A distribution network fault diagnosis method based on a hybrid neural network model according to claim 7, characterized in that: In step S3), the state transition probability of two adjacent time steps of the time series is calculated, a Markov state transfer matrix is ​​constructed, and the element values ​​in the Markov state transfer matrix are converted into different colors according to their sizes using the imagesc function, and the corresponding positions on the coordinate axis are colored with corresponding colors to obtain a two-dimensional coded image of the three-phase voltage signal of the distribution network bus line.

9. A distribution network fault diagnosis method based on a hybrid neural network model according to claim 1, characterized in that: In step S4), the convolutional neural network CNN is provided with three channels, which are respectively used to process three groups of two-dimensional coded images of voltage; each channel is provided with an image input layer, a convolution layer, a batch normalization layer, an activation layer, a maximum pooling layer, a first fully connected layer, and a flattening layer; the mathematical model of the convolutional neural network CNN is expressed as: In the formula, are three sets of two-dimensional coded images as input; is the output of the CNN branch; Calculated for the convolutional layer; Calculated for batch normalization layer; Calculated for activation layers; Calculated for the maximum pooling layer; It is the fully connected layer operation; It is the flattening layer operation; The gated recurrent unit GRU is provided with three channels, each of which includes a time series input layer, a flattening layer, and a GRU layer. The GRU layer controls the flow of information by introducing a gating mechanism, retains relatively important information, and removes relatively unimportant information; the flattening layer flattens the multi-dimensional input into one dimension; the gated recurrent unit GRU outputs three groups of features of three groups of time series; the mathematical model of the gated recurrent unit GRU is: In the formula, are the three sets of time series input; is the output of the GRU branch; Calculated for flattened layers; Calculated for the GRU layer.

10. A distribution network fault diagnosis method based on a hybrid neural network model according to claim 9, characterized in that: In step S4), the combined part of the hybrid neural network fault diagnosis model includes an attention layer, a second fully connected layer, a Softmax layer, and a classifier; the mathematical model of the hybrid neural network fault diagnosis model is: In the formula, is the output of the GRU branch; is the output of the CNN branch; x output is the output of the hybrid neural network fault diagnosis model.

Images