A power distribution network fault diagnosis method based on a hybrid neural network model

Abstract

The application provides a power distribution network fault diagnosis method based on a hybrid neural network model, comprising judging whether the power distribution network is in a normal state based on a voltage mutation signal rule; when the power distribution network is in an abnormal state, collecting three-phase voltage signals delayed by 3 cycles at a fault moment, and unifying a sampling rate; converting the three-phase voltage signals with the unified sampling rate into time series, and then using a Markov transition field to perform image coding; constructing and training a hybrid neural network fault diagnosis model, and using the trained diagnosis model to perform real-time monitoring on the power distribution network to diagnose the fault type of the power distribution network. The application can simultaneously extract two-dimensional coding image features and time series features of three-phase voltage signals, and introduce a gating vector to weight and sum the features extracted by each channel, solve the problem of incomplete feature information extraction of a single training model, realize higher fault diagnosis accuracy, and improve the fault diagnosis capability of the power distribution network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution network fault diagnosis, and in particular to a power distribution network fault diagnosis method based on a hybrid neural network model. BACKGROUND

[0002] As a transmission link of the power system, the power distribution network undertakes the heavy responsibility of safely and reliably transmitting power to the user end. However, due to the complexity of the structure, the diversity of the equipment and the variability of the operating environment, the power distribution network is prone to frequent faults, which seriously affects the stability of the power system and the power quality of the users. Therefore, developing a power distribution network fault diagnosis technology is an important means to improve the safe and reliable operation of the power system.

[0003] The traditional fault diagnosis methods for the power distribution network include the experienced judgment method: the power distribution management personnel judge the operating condition and the fault state according to experience and the line history, the expert system method: the computer technology is used to integrate theoretical knowledge and expert experience, and the faults are identified through the organic connection of multiple components, the method based on fuzzy theory: due to the influence of various factors, the fault diagnosis has uncertainty, and the fuzzy theory is good at simulating human thinking, and multiple possible solutions are sorted out to find the most correct solution. However, the above methods have a certain subjectivity, which affects the accuracy of the diagnosis, and the fault tolerance is poor when dealing with special cases.

[0004] At present, the fault diagnosis technology for the power distribution network system is developing, and the main methods include the intelligent power operation and maintenance cloud platform monitoring, the diagnosis method based on the time sequence convolution residual network. At the same time, the artificial intelligence and machine learning technology has made remarkable progress in the field of power distribution network fault diagnosis, which automatically identifies and classifies faults through model training. However, most of them use two-dimensional images or time series for model training and fault diagnosis, and do not combine the two, so that the fault diagnosis accuracy is relatively low. Therefore, the two are combined for use, and the running state of the power distribution network is pre-judged, which can well improve the accuracy of the power distribution network fault diagnosis. SUMMARY

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

[0006] The technical scheme of the present application is as follows: a power distribution network fault diagnosis method based on a hybrid neural network model, comprising the following steps:

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

[0008] S2), when judging that the power distribution network is in an abnormal state, collecting three-phase voltage signals delayed by 3 cycles at the fault moment at the power distribution network bus, and unifying the sampling rate;

[0009] S3), converting the three-phase voltage signals with the unified sampling rate in step S2) into one-dimensional voltage time series, and using a Markov transition field to image encode 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 a convolutional neural network (CNN) and a gated recurrent unit (GRU), and training the hybrid neural network fault diagnosis model using the one-dimensional voltage time series and the two-dimensional encoded image;

[0011] S5), inputting the real-time monitored voltage signals processed in steps S1)-S3) in sequence into the trained hybrid neural network fault diagnosis model to diagnose the fault type of the power distribution network.

[0012] As preferred, in step S1), the three-phase voltage of the power distribution network bus is monitored in real time, specifically including the following steps:

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

[0014] S12), calculation of the voltage mutation variable within 3 cycles before the current detection moment t to obtain a voltage mutation variable matrix;

[0015] S13), transformation processing of each column of the voltage mutation variable matrix to obtain a positive sequence voltage mutation variable signal matrix, and calculation of a fault monitoring start criterion according to the positive sequence voltage mutation variable signal matrix; determination of whether the power distribution network is in a normal operating state through the start criterion.

[0016] As preferred, in step S11), the voltage signal in the time window [t-t o , t] is intercepted with the current detection moment t as the reference, the sampling rate f h is used to collect the three-phase voltage signals in the time window; and the A, B, and C three-phase voltage signals collected in the time window [t-t o , t] are used to form a three-phase voltage signal matrix T1.

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

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

[0019]

[0020] wherein, are the voltage mutations of the three phases A, B and C at the detection position of the bus of the distribution network at time t1, respectively; the value range of t1 is [t-t o / 2, t]; and a is a rotation factor.

[0021] As preferred, in step 13), the calculation formula of the starting criterion ΔV of the fault monitoring is:

[0022]

[0023] wherein, is the number of sampling points; is an element of the positive sequence voltage mutation matrix at time t; and f h is the actual sampling rate;

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

[0025] As preferred, in step S2), during the sampling process, the size relationship between the actual sampling frequency f h and the expected sampling frequency f s is judged.

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

[0027] If f h = f s , no processing is needed.

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

[0029] As preferred, in step S2), when the actual sampling frequency fh less than the expected sampling frequency f s When the sampling rate is increased, the method of cubic spline interpolation is used, specifically, the original sampling points are divided into multiple subintervals according to every two adjacent sampling points, and a cubic polynomial is set in the form of the midpoint difference of each subinterval.

[0030] As preferred, in step S2), when the actual sampling frequency f h greater than the expected sampling frequency f s When the sampling rate is increased, the method of cubic spline interpolation is used, specifically, the original sampling points are divided into multiple subintervals according to every two adjacent sampling points, and a cubic polynomial is set in the form of the midpoint difference of each subinterval.

[0031] As preferred, in step S3), the voltage signal matrix of the sampling points after the sampling rate is unified is recorded for each time point and each phase voltage value, and each group of sequences is normalized and scaled, so as to obtain the time sequence of the three-phase 3 groups of voltages A, B and C.

[0032] As preferred, in step S3), the state transition probability of the adjacent two time steps of the time sequence is calculated, the Markov state transition matrix is constructed, the element values in the Markov state transition matrix are converted into different colors according to the size by using the imagesc function, and the corresponding color is colored at the corresponding position of the coordinate axis, so as to obtain the two-dimensional encoding image of the three-phase voltage of the distribution network bus line.

[0033] As preferred, in step S4), the mathematical model of the convolutional neural network CNN is represented as:

[0034]

[0035] In the formula, is the input of the 3 groups of two-dimensional encoding images; is the output of the CNN branch; is the convolution layer calculation; is the batch normalization layer calculation; is the activation layer calculation; is the maximum pooling layer calculation; is the full connection layer operation; is the flattening layer operation.

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

[0037]

[0038] In the formula, is the input of the 3 groups of time sequences; is the output of the GRU branch; is the flattening layer calculation; is the GRU layer calculation.

[0039] As preferred, in step S4), the calculation expression of the mixed 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 mixed neural network fault diagnosis model.

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

[0043] 1. The present application can simplify unnecessary workload and well achieve preliminary judgment of faults by judging whether the power distribution network is in a normal state based on the voltage mutation signal law, and only when it is judged to be in an abnormal state, the next step of diagnosis is performed.

[0044] 2. The mixed neural network fault diagnosis model constructed by the present application can extract two-dimensional code image features and time sequence features of three-phase voltage signals at the same time, and introduce a gating vector to further weight and sum the features extracted by each channel, which can well solve the problem of incomplete feature information extraction of a single training model, realize higher fault diagnosis accuracy, and greatly improve the fault diagnosis capability of the power distribution network. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a flowchart of the method of embodiment 1 of the present application;

[0046] Figure 2 is a circuit schematic diagram of the power distribution network of embodiment 1 of the present application;

[0047] Figure 3 is a mixed neural network fault diagnosis model framework constructed by embodiment 1 of the present application;

[0048] Figure 4 is a criterion diagram for starting fault detection in embodiment 1 of the present application;

[0049] Figure 5 is a three-phase voltage signal diagram of different short-circuit states in embodiment 1 of the present application;

[0050] Figure 6 is a two-dimensional code image diagram of three-phase voltage at M0 of the power distribution network in embodiment 1 of the present application;

[0051] Figure 7 is a criterion diagram for starting fault detection in embodiment 2 of the present application;

[0052] Figure 8 This is a schematic diagram of the three-phase sampling signal in Embodiment 2 of the present invention;

[0053] Figure 9 This is a schematic diagram of the MTF two-dimensional encoded image when a single phase is grounded in Embodiment 2 of the present invention. Detailed Implementation

[0054] The specific embodiments of the present invention will be further described below with reference to 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 below. Figure 2 As shown, the three-phase voltage signal measurement point is at M0 on the busbar, and the short-circuit fault point is set at M1-M2. 14 Place.

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

[0058] S1) Real-time monitoring of the three-phase voltage of the distribution network bus, and determination of whether the distribution network is in a normal state based on voltage change signals; specifically including the following steps:

[0059] S11) Real-time detection of the three-phase voltage at the distribution network bus M0; in this embodiment, the current detection time t is used as the reference, and the time window is [tt] o The voltage signal of t] is sampled at a rate f. h Acquisition time window [tt] o The three-phase voltage signals A, B, and C within [t] constitute a three-phase voltage signal matrix T1, which is updated in real time over time. The three-phase voltage signal matrix T1 is represented as:

[0060]

[0061] In the formula, r+1 is the number of sampling points contained in the time window; These are the voltage signals of the three-phase voltages A, B, and C at the busbar measured at time t; The three-phase voltages A, B, and C at the busbar are respectively at t o The voltage signal obtained by constant measurement;

[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 Then update the time window as follows The number of sampling points is

[0063] S12), the voltage mutation quantity within 3 cycles before the current detection time t is calculated to obtain a voltage mutation quantity matrix;

[0064] The time window of the sampled signal in the embodiment is and is divided into and two time periods; then the voltage signal values within 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 matrices T2 and T3 are processed by difference to obtain a matrix T4, i.e., the voltage mutation quantity within 3 cycles before the current detection time t, which is specifically as follows:

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

[0068]

[0069] In the formula, are the voltage mutation quantities of the A, B and C three phases at the current detection time t at the busbar, respectively; is the number of sampling points contained in the time window .

[0070] S13), each column of the voltage mutation quantity matrix T4 is processed by transformation, and whether the distribution network is in a normal operation state is determined according to a criterion;

[0071] Each column of the voltage mutation quantity matrix T4 is processed by transformation as follows to obtain a matrix T5, wherein the matrix T5 is a matrix composed of positive sequence voltage mutation signals:

[0072]

[0073] In the formula, U1 is the transformed positive sequence voltage mutation within the time window ; α is a rotation factor; the value range of t1 is

[0074] The finally obtained matrix T5 is represented as:

[0075]

[0076] In the formula, V t is the positive sequence voltage mutation at the time t; is the positive sequence voltage mutation at the time t; ΔV is the positive sequence voltage mutation variable of the last sampling point in 3 cycles, and the number of sampling points is

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

[0078] Wherein:

[0079]

[0080] In the formula, the value range of ΔV is represents the positive sequence voltage mutation variable signal at the time t

[0081] In this embodiment, the frequency of the power distribution network is 50 Hz, so the time length of 3 cycles is 0.06 s, as shown in the schematic diagram of starting criterion. Figure 4

[0082] S2), when it is judged that the power distribution network is in an abnormal state, the three-phase voltage signals 3 cycles later than the fault occurrence time t s are collected, and the sampling rates of the sampling signals are unified;

[0083] The sampling rate of the collected signals in this embodiment is unified to 50 kHz, denoted as f s ; 3 cycles later than the positive sequence voltage mutation time t s is collected at the initial sampling rate f h to form a matrix V h :

[0084]

[0085] In the formula, n+1 is the total number of sampling points in 3 cycles later than t s at the initial sampling rate, and is the three-phase voltage signal measured at t s at the bus of the power distribution network, is the three-phase voltage signal measured at the next sampling point at t s at the bus of the power distribution network; is the three-phase voltage signal measured at the last sampling point in 3 cycles at the bus of the power distribution network;

[0086] ​​Let the matrix form of the voltage signal of the sampling point after sampling rate unification be V s :

[0087]

[0088] In the formula, m+1 is the total number of sampling points in 3 cycles after sampling rate unification, and is the three-phase voltage signal measured at the bus of the power distribution network at t s , is the three-phase voltage signal measured at the next sampling point of the bus of the power distribution network at t s ; is the voltage signal measured at the last sampling point in 3 cycles at the bus of the power distribution network.

[0089] First, the size relationship between the actual sampling frequency and the expected sampling frequency is judged:

[0090] If f h < f s , upsampling processing is performed, that is, when the actual sampling frequency is less than the expected sampling frequency, the method of cubic spline interpolation is used to increase the sampling rate, and the specific process is as follows:

[0091] The original sampling points are divided into multiple subintervals according to every two adjacent sampling points, and the midpoint of each subinterval is set in the form of difference, and the basic form of the cubic polynomial is:

[0092]

[0093] In the formula, a i , b i , c i , and d i are undetermined coefficients, which are solved by boundary conditions, h i (x) represents the cubic polynomial in the sample point x i and the sample point x i+1 ; 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 curve end point passes through the data point

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

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

[0098] 2) continuity of the first derivative of the adjacent curve end points

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

[0100] 3) continuity of the first derivative of the adjacent curve end points

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

[0102] 4) continuity of the second derivative of the adjacent curve end points

[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 three-diagonal matrix algorithm, the undetermined coefficients of the cubic polynomial are obtained, so that the sample points are inserted in each region to obtain the sequence V s after the sampling rate is unified.

[0105] If f h =f s , no processing is needed.

[0106] If f h >f s , down-sampling processing is performed, that is, when the actual sampling frequency is greater than the expected sampling frequency, the embodiment uses the decimation method to realize down-sampling, and the specific operation is as follows:

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

[0108]

[0109] wherein D is a decimation factor; f h is an actual signal frequency; f sfor the expected sampling rate;

[0110]

[0111] Starting from the middle sample point, select sample points, and form a new sequence V s

[0112]

[0113] In this embodiment, the power distribution network frequency is 50 Hz, and the time length of 3 cycles is 0.06 s. Through the processing of step S2), the sampling rate is unified to 50 kHz, and the total number of sample points is 3000. As shown in Figure 5 the three-phase voltage signals collected at the bus of the power distribution network when different short-circuit faults occur in the power distribution network (after sampling rate unification).

[0114] S3), convert the three-phase voltage signals after sampling rate unification in step S2) into a one-dimensional voltage time sequence, and use Markov transition field to image encode the voltage time sequence to obtain a two-dimensional encoded image; specifically:

[0115] S31), convert the three-phase voltage signals after sampling rate unification into a one-dimensional voltage time sequence;

[0116] According to the sampling frequency f s obtained, the continuous voltage analog signal is represented by discrete digital variables: V s (t) is recorded to obtain: p

[0117]

[0118] In the formula, p is the index value of the sample 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 sequences of the A, B, and C three-phase voltages, respectively; V A (t0), V A (t0), V A (t0) are the voltage amplitudes of the A, B, and C three-phase at the first sample point at the bus of the power distribution network, respectively.

[0119] Then, each group of sequences is normalized and scaled to [-1, 1] according to the following formula, that is:

[0120] ​

[0121] where p is in the range of 0, 1, 2, 3,..., m, x p is X A (t), X B (t), X C (t) are the index values of the three groups of time series; max(X) is X A (t), X B (t), X C (t) are the maximum values of the three groups of time series respectively; min(X) is X A (t), X B (t), X C (t) are the minimum values of the three groups of time series respectively; is the normalized scaling value;

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

[0123]

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

[0125] S32), using the Markov transition field to image encode the three groups of voltage time series X(t) obtained in S31) to convert them into two-dimensional encoded images;

[0126] S321), constructing a state transition matrix

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

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

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

[0130] where σ is in the range of 1, 2, 3,..., O.

[0131] Each data in each region is changed to the serial number of its corresponding region, and for each region, the state transition probability of adjacent two time steps is calculated, and these probability values are recorded as elements w εθ , and an O×O-dimensional Markov state transition matrix W is constructed using w εθ , as shown in equation (26):

[0132]

[0133] where w εθ denotes the probability of transition from state ε to state θ;

[0134] To calculate w εθ , first construct a transition count zero matrix C1 of size O x O, as shown in equation (28), which stores the transition times of voltage states at different states; next, loop through each region, for each region, get the transition times from the current state to other states, and continuously update the transition count matrix; at the same time, set the probability corresponding to state ε in the state transition matrix to 0; finally, as shown in equation (29), update the matrix C1 by counting the voltage state transition times of each region to obtain the matrix C2, and then calculate and normalize these transition times to obtain the voltage transition probability matrix;

[0135]

[0136] where C εθ is the number of times of transition from state ε to state θ.

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

[0138] The element values in the state transition matrix W are converted into different colors according to their sizes using the imagesc function, and the corresponding position on the coordinate axis is colored with this color; each element w εθ in the state transition matrix W satisfies the following color mapping rule (YGB): the brightness of each color of yellow-green-blue gradually increases from level 0 to 255, realizing different proportions of color mixing to form the corresponding color, that is:

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

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

[0141] Finally, the corresponding color is obtained by color mixing of the above obtained matrix, and is displayed at the coordinate position corresponding to 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 images are obtained​Figure 6 a two-dimensional encoding image of three-phase voltages at the power distribution network bus M0.

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

[0143] The embodiment trains the hybrid neural network fault diagnosis model using the three groups of one-dimensional voltage time series obtained in step S31) and the two-dimensional encoding image in step S33) as training data sets.

[0144] The embodiment builds a power distribution network model based on a MATLAB / Simulink platform, simulates four abnormal states of three-phase short circuit, two-phase short circuit, single-phase grounding, and two-phase grounding existing in the power distribution network in actual operation by changing different short circuit states in the three-phase fault module at time t s

[0145] For each short circuit fault state, the transition resistance and the 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, collect the three-phase voltage signals of the power distribution network bus at the fault occurrence time ts delayed by 3 cycles according to step S2), and unify the sampling rate to 50 kHz; then generate three groups of voltage signal time series and three groups of two-dimensional encoding images of the voltage signals according to step S3), wherein the three groups of time series and the three 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 power distribution network bus M0, each group of samples corresponds to three groups of time series and three groups of two-dimensional encoding images, each image has a dimension of 200x200x3, and the time series is a one-dimensional sequence of 1x3000. In this embodiment, there are four different short circuit fault states for each branch, and there are 14 branches in total, so a total of 2800 groups of samples, 8400 two-dimensional encoding images, and 8400 groups of time series are generated.

[0149] ​The convolutional neural network CNN of the embodiment is provided with three channels for processing three groups of voltage two-dimensional encoding images respectively; each channel is provided with an image input layer, a convolutional layer, a batch normalization layer, an activation layer, a maximum pooling layer, a first full connection layer, and a flattening layer; wherein the convolutional layer uses a 3x3 convolutional kernel, 64 filters, and a step 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, a step of 2, and padding of'same'; the first full connection layer has 64 units; and the flattening layer flattens the multi-dimensional input into one dimension. The output is three groups of features of the three groups of two-dimensional encoding images, and the mathematical model of the convolutional neural network CNN is represented as:

[0150]

[0151] wherein, is the input of the three groups of two-dimensional encoding images; is the output of the CNN branch; is the convolutional layer calculation; is the batch normalization layer calculation; is the activation layer calculation; is the maximum pooling layer calculation; is the full connection layer operation; is the flattening layer operation;

[0152] The gated recurrent unit GRU is provided with three channels, each including a time series input layer, a flattening layer, and a GRU layer, wherein the GRU layer controls the flow of information by introducing a gating mechanism to retain relatively important information and remove relatively unimportant information; and the flattening layer flattens the multi-dimensional input into one dimension. The output is three groups of features of three groups of time series; wherein the gating mechanism is a reset gate and an update gate; and the mathematical model of the gated recurrent unit GRU is:

[0153]

[0154] wherein, is the input of the three groups of time series; is the output of the GRU branch; is the flattening layer calculation; is the GRU layer calculation.

[0155] The combination part of the hybrid neural network fault diagnosis model includes an attention layer, a second full connection layer, a Softmax layer, and a classifier, wherein the second full connection layer has 5 neurons; the Softmax layer converts the output of the second full connection layer into a probability distribution; and the classifier outputs the predicted class label according to the probability distribution obtained from the previous layer.

[0156] The calculation expression of the mixed 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 mixed neural network fault diagnosis model.

[0159] In this embodiment, the attention layer adopts a multi-head attention mechanism, and a gating vector τ with a dimension of 6 is introduced 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 are training parameters; and q is a 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 respectively the features of three groups of two-dimensional code images; P4, P5, and P6 respectively represent three groups of time sequence features; respectively represent the feature vectors of the qth group of two-dimensional code images in groups A, B, and C; respectively represent the feature vectors of the δth group of time sequences in groups A, B, and C;

[0165] Linear transformation is performed on formulas (35) and (36), and the input data is mapped to a new feature space, as shown in formulas (37)-(39):

[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 linear transformation parameters of query space, key space and value space, respectively; Q, K, V are query space, key space and value space, respectively; P represents the features of time series and two-dimensional encoding images; q, k, v are element vectors in query space, key space and value space, respectively; β is the number of query vectors, and γ is the dimension number after linear transformation.

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

[0171]

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

[0173] The above similarity score is converted into a probability distribution by applying the Softmax function, wherein the attention weight a(q, k i ) corresponding to the query Q and K is represented as:

[0174]

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

[0176] The attention layer performs weighted sum processing on each value of each head by calculating the attention weight, to obtain the corresponding output Y μ , and finally performs weight sum on the multi-head channel to obtain the output Y, that is:

[0177]

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

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

[0180] In this embodiment, 2800 groups of samples are divided into training set, test set and validation set according to the ratio of 7:1.5:1.5, wherein the two-dimensional code image is sent to the CNN branch, and the time sequence is sent to the GRU branch; the loss function of the mixed neural network fault diagnosis model uses the cross-entropy loss function, and is trained by 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 iteration number is 200 times.

[0181] In each round of training iteration, the validation set is used to evaluate the model to ensure continuous optimization of the model, and after the model converges to the preset iteration number, the current model is saved. Subsequently, the test set is used to evaluate the diagnostic accuracy of the model, and if the accuracy does not reach the optimal value, 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 offline trained mixed neural network fault diagnosis model to monitor the power distribution network in real time to diagnose the fault type of the power distribution network, specifically:

[0183] When the power distribution network system is actually running, first, whether the power distribution network is in a normal state is judged according to step S1), and when it is judged that the power distribution network is in an abnormal state, the time t s of fault occurrence is obtained.

[0184] Then, according to step S2), three-phase voltage signals at time t s delayed by 3 cycles are collected at the bus M0 of the power distribution network, and the sampling rate is unified to 50 kHz.

[0185] Then, according to the method of step S3), the time sequence of the detection voltage signal is obtained, and the three groups of time sequences of the voltage are converted into three groups of two-dimensional code images by using the Markov transition field.

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

[0187] Embodiment 2

[0188] S1), assuming that a single-phase ground fault occurs at M1 in the power distribution network system, through the processing of step S1, the starting criterion ΔV result is calculated as 2001.5V, wherein ΔV set is 200V, and the frequency of the power distribution network is 50Hz, so the time length of 3 cycles is 0.06s. At this time, the starting criterion is satisfied, and the time t s of fault occurrence is recorded, and step S2 is entered.Figure 7 as shown.

[0189] Step S2), collecting the moment t of fault occurrence at the distribution network bus s the three-phase voltage signal delayed by three cycles; then the sampling rate is unified to 50 kHz; as Figure 8 as shown.

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

[0191]

[0192] Finally, the obtained three groups of time series and three groups of two-dimensional encoding images are sent into the offline trained fault diagnosis model, and the probability of being predicted as a single-phase ground fault is 95.6%, and the three groups of two-dimensional encoding images obtained in the embodiment are shown in Figure 9 as shown.

[0193] The above embodiments and descriptions described in the specification are only to illustrate the principles and best embodiments of the present application, and various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application.

Claims

1. A power distribution network fault diagnosis method based on a hybrid neural network model, characterized in that, It comprises the following steps: 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 jump signal rule; S2), when judging that the distribution network is in an abnormal state, collecting the three-phase voltage signal of the distribution network bus at the fault time delayed by 3 cycles at the distribution network bus, and unifying the sampling rate; S3), converting the three-phase voltage signal with a unified sampling rate in step S2) into a one-dimensional voltage time sequence, and using a Markov transition field to image encode the one-dimensional voltage time sequence to obtain a two-dimensional encoding image; S4), constructing a hybrid neural network fault diagnosis model based on a convolutional neural network (CNN) and a gated recurrent unit (GRU), and training it using the one-dimensional voltage time sequence and the two-dimensional encoding image; The gated recurrent unit GRU has three channels, each channel including a time sequence input layer, a flattening layer, and a GRU layer, wherein the GRU layer controls the flow of information by introducing a gating mechanism to retain relatively important information and remove relatively unimportant information; the flattening layer flattens the multi-dimensional input into one dimension; the gated recurrent unit GRU outputs 3 groups of features for 3 groups of time sequences; and the mathematical model of the gated recurrent unit GRU is: ； wherein is the input 3 groups of time series; is the output of the GRU branch; is the flattened layer calculation; is the GRU layer calculation; The combination part of the hybrid neural network fault diagnosis model includes an attention layer, a second fully connected layer, a Softmax layer, and a classifier; and 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; is the output of the hybrid neural network fault diagnosis model; S5), inputting the real-time monitored voltage signal into the trained hybrid neural network fault diagnosis model after processing through steps S1)-S3) in sequence to diagnose the fault type of the distribution network.

2. The power distribution network fault diagnosis method based on a hybrid neural network model according to claim 1, characterized in that: In step 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 jump signal rule, comprises the following steps: S11), real-time detection of the three-phase voltage at a certain place of the distribution network bus; S12), calculating the voltage jump within 3 cycles before the current detection time t to obtain a voltage jump matrix; S13), transforming each column of the voltage jump matrix to obtain a positive sequence voltage jump signal matrix, and calculating a fault monitoring start criterion according to the positive sequence voltage jump signal matrix; determining whether the distribution network is in a normal operating state through the start criterion.

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

4. The power 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 [t - t o , t] is divided into two time windows [t - t o / 2, t] and [t - t o , t - t o / 2]; and the voltage signal values in the above two time windows are extracted in the three-phase voltage signal matrix T1 to form matrices T2 and T3, respectively, and the difference between matrices T2 and T3 is obtained to obtain the voltage mutation matrix T4 in the time range of t o / 2 before the current detection time t.

5. The power 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 jump matrix T4 is transformed as follows to obtain a positive sequence voltage jump signal matrix, and the transformation expression is specifically: ； In the formula, , , are voltage jump variables of the A, B, C three-phase at the detection position of the bus of the power distribution network at time t1, respectively; wherein the value range of t1 is [t-t o / 2, t]; is a rotation factor; The start criterion of the failure monitoring The calculation formula is: ； ； In the formula, is the number of sampling points; is is the element of the positive sequence voltage mutation matrix at the moment f h is the actual sampling rate; If , the power distribution network is in an abnormal operation state, and the current detection time is taken as the fault occurrence time t s , otherwise, the power distribution network is in a normal operation state, wherein is a preset value.

6. The power 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, it is judged whether the actual sampling frequency is greater than or less than the expected sampling frequency f s . If , upsampling processing is performed, i.e. when the actual sampling frequency is less than the expected sampling frequency f s , the increase of the sampling rate is realized by using the method of cubic spline interpolation, 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 the difference of the midpoints of each sub-interval; If then no processing is required; If down-sampling is performed; that is, when the actual sampling frequency is greater than the expected sampling frequency f s Down-sampling is achieved using decimation when the actual sampling frequency 7. The power 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 corresponding to each time point in the matrix of the voltage signal after unifying the sampling rate is recorded, and each group of sequences is standardized and scaled to [-1, 1], thereby obtaining 3 groups of voltage time sequences of phases A, B, and C.

8. The power 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 in the time series is calculated, a Markov state transition matrix is constructed, the element values in the Markov state transition matrix are converted into different colors according to the size by using the imagesc function, and the corresponding color is colored at the corresponding position of the coordinate axis, so as to obtain a two-dimensional encoding image of the three-phase voltage signal of the power distribution network bus line.

9. The power 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 for processing three groups of two-dimensional encoding images of the voltage; each channel is provided with an image input layer, a convolutional layer, a batch normalization layer, an activation layer, a maximum pooling layer, a first full connection layer and a flattening layer; and a mathematical model of the convolutional neural network CNN is represented as: ； wherein is the input of the 3 groups of 2D encoded images; is the output of the CNN branch; is the convolution layer computation; is the batch normalization layer computation; is the activation layer computation; is the max pooling layer computation; is the fully connected layer operation; is the flatten layer operation.

Images