Fault identification method for AC / DC power transmission system based on space-time diagram convolution

Abstract

The invention provides an AC / DC power transmission system fault identification method based on space-time diagram convolution, and the method comprises the steps: carrying out the modeling of a diagram structure, and constructing a diagram data set; performing graph convolution on the graph data to obtain a matrix Mt; extracting time dynamic characteristics of the matrix Mt; outputting a predicted value of each fault category by adopting an activation function; solving a loss function of the fault identification model; updating the parameters of the fault recognition model according to back propagation until the loss function value meets the requirement, and obtaining a trained fault recognition model; and selecting DC system fault event data outside the graph data set, and inputting the DC system fault event data into the trained fault identification model for fault identification. According to the technical scheme, automatic diagnosis of various electrical faults of the direct-current power transmission system can be realized, and the efficiency of fault handling work is improved.

Description

Technical Field

[0001] The present invention relates to the field of power system fault diagnosis, and in particular to an AC / DC power transmission system fault identification method based on spatiotemporal graph convolution. Background Art

[0002] As one of the most typical applications of power electronics technology in power systems, DC transmission projects not only solve the static stability limit problem of AC transmission in long-distance transmission, but also realize the asynchronous networking of regional power grids. It is an important step in the intelligent transformation of my country's power grid. At present, domestic DC projects include end-to-end, back-to-back and multi-terminal DC transmission projects. The rapid application of new technology projects and complex architecture make the fault characteristic analysis more complicated. In particular, most of the domestic DC projects are not equipped with DC circuit breakers and do not have the ability to self-clear DC faults. Fault identification and processing are particularly important.

[0003] Once a DC transmission system fails, it may cause a large-scale power outage, severely affecting industrial production and residents' lives, and leading to huge economic losses. Through fault analysis and identification technology, it is possible to prevent the occurrence of faults or locate and deal with fault events as soon as possible, reducing losses during project operation.

[0004] At present, the fault diagnosis method of HVDC transmission system based on machine learning usually adopts multi-layer neural network (MLP), such as convolutional neural network (CNN), which directly selects single electrical quantity data of certain points in AC / DC system for feature extraction and fault probability calculation. However, the electrical data of a single measurement point cannot reflect the fault characteristics of the entire power system. Such algorithms do not consider the interactive influence relationship in the entire AC / DC system, and cannot extract the spatial or temporal characteristics of electrical data in the AC / DC system, so their practical applicability is low. Summary of the invention

[0005] To solve the above problems, the present invention implements a fault identification method for AC / DC power transmission systems based on spatiotemporal graph convolution, which can effectively process non-Euclidean structure data in the DC system, and utilize the time and space relationship between nodes in the DC transmission system to simultaneously identify various types of faults in multiple areas such as the AC system, DC system and converter valves.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for fault identification of AC and DC power transmission systems based on spatiotemporal graph convolution comprises the following steps:

[0008] S1. Perform graph structure modeling, collect DC system fault data under various working conditions, construct a graph data set, and divide the graph data set into a training data set and a test data set;

[0009] S2. Obtain the matrix M by performing graph convolution on the graph data in the graph dataset t ;

[0010] S3, the matrix M after the convolution operation of S2 t After being processed by the gated recurrent neural network, the temporal dynamic features are extracted to obtain K t , K t After performing graph convolution again, the Softmax activation function is used to output the predicted value Z of each fault category. lf ;

[0011] S4, combining the fault category prediction process of steps S1-S3 into a fault identification model, and obtaining a loss function of the fault identification model;

[0012] S5, take the training data set and update the parameters of the fault identification model in S4 through back propagation;

[0013] S6. After updating the parameters of the fault identification model, take a test data set and repeat steps S2 to S5 until the loss function value of the test data set is lower than a preset threshold, then end the model training and obtain a trained fault identification model;

[0014] S7. Select DC system fault event data outside the data set in S1, and input the fault identification model trained in S6 to perform fault identification.

[0015] Furthermore, step S1 specifically includes:

[0016] Set each key electrical measurement point in the DC system to N x The connection relationship of each node is determined according to the DC system topology and fault characteristics, and the graph structure is modeled; various types of faults are set in the DC simulation system, fault parameters are adjusted, and fault simulation is performed under different working conditions; data is extracted from the simulation waveform, that is, the electrical quantity vector data of each measurement node within time T in n faults or normal working conditions are combined into a characteristic matrix {X0, X1···, X t}, forming a graph data set G{G1, G2···, G n}, according to the simulation fault settings or actual engineering fault results, each graph data G n The faulty graph is labeled, and no label is added when there is no fault; data in the graph dataset is randomly selected as the training dataset and the test dataset respectively.

[0017] Furthermore, step S2 specifically includes:

[0018] Graph convolution is performed based on the node features of the modeled graph structure. The features of the graph structure include the node adjacency matrix A T(i，j) and A TThe degree matrix, A T If the i node and the j node are connected, then the element A in the adjacency matrix T(i，j) If it is not connected, it is 1, and if it is not connected, it is 0. Each node is not connected to itself;

[0019] The adjacency matrix is ​​added with the identity matrix I to increase the node's own characteristics, resulting in the following matrix A:

[0020] A=A T +I (1)

[0021] Let D be A T The degree matrix is ​​a diagonal matrix diag(), and the elements on the diagonal of the degree matrix are the number of edges d(v) connecting each node to other nodes in the graph structure. i ), the expression is as follows:

[0022] D=diag(d(v1),d(v2),···,d(v N )) (2)

[0023] The node feature data X from the initial time to time t t , adjacency matrix A T(i，j) , degree matrix D to get the matrix M at each moment t :

[0024] M t =σ ReLU( D -1 / 2 AD -1 / 2 X t W 1 ) (3)

[0025] where σ ReLU is the ReLU nonlinear activation function, W 1 , W 2 is the weight parameter matrix;

[0026] The ReLU nonlinear activation function expression is as follows:

[0027]

[0028] Furthermore, step S3 specifically includes:

[0029] The temporal dynamic feature extraction method is based on the “z” in the recurrent neural unit from the initial moment, t-1 to time t. t 、r t 、h t , K t " is composed of the following calculation process:

[0030] z t =σ sig (Wz1 M t +W z2 K t-1 +b z ) (5)

[0031] r t =σ sig (W r1 M t +W r2 K t-1 +b r ) (6)

[0032]

[0033] The tanh activation function expression is as follows:

[0034]

[0035] Hadamard product The matrix calculation method is as follows:

[0036]

[0037] where z t is the update gate output, r t is the reset gate output, h t is the candidate hidden state, K t is the output of the recurrent neural unit at time t; K t-1 represents the output of the recurrent neural unit at time t-1, M t represents the node feature data input at time t, σ sig is the sigmoid activation function; W z1 , W z2 , W r1 , W r2 , W h1 , W h2 is the weight parameter, b z 、b r 、b h is the bias parameter;

[0038] set up It is expressed by the following formula:

[0039]

[0040] After the above calculations, we get the output K of the recurrent neural unit at time t. t , and again with Perform graph convolution operation and use Softmax activation function σ softmax Output the predicted value Z of each fault category lf:

[0041]

[0042] The S1, S2, and S3 fault category prediction process is the fault identification model.

[0043] Furthermore, step S4 specifically includes:

[0044] The cross entropy loss function expression of the fault recognition model is as follows:

[0045]

[0046] Where F represents the number of fault categories and l represents the graph structure y L The graph data with fault labels, Y lf represents the various fault label values ​​in S1 training or test data, Y lf When it is 1, the fault occurs, which is a positive value; when it is 0, the fault does not occur, which is a negative value. lf Represents the fault category prediction value of graph data.

[0047] Furthermore, the back propagation refers to a method for calculating the parameter gradient of each layer of the neural network, using the chain rule to find the partial derivatives of the loss function with respect to the weights and biases of each neuron layer by layer, forming the gradient of the loss function with respect to the weight and bias vectors as the basis for modifying the weights.

[0048] Furthermore, step S7 specifically includes: selecting a newly occurring event that requires fault identification, extracting data from the DC system from the initial moment of the event to the end moment of time T, converting the data into a feature matrix format according to S1, inputting the fault identification model trained in S6, repeating steps S2 and S3, obtaining various fault probability values, and selecting the maximum probability as the identified fault category.

[0049] The beneficial effects of the present invention are as follows:

[0050] A multi-category electrical fault identification method is proposed for the DC transmission system. A DC fault cause identification method is proposed, which can automatically diagnose the cause of the fault. Considering the dynamic nature of the DC fault process, an update mechanism based on the time state sequence is added to the static graph convolution architecture to form a dynamic graph convolution layer. After being processed by multi-layer graph convolution and gated recurrent neural network, the data extracted by the multi-layer fusion model takes into account the spatial characteristics and temporal characteristics of multi-region data of the DC system, and the fault identification of the entire DC system is more comprehensive. Existing similar methods can only sample limited electrical data and realize limited category fault identification. The present invention can simultaneously identify multiple types of faults in multiple regions such as AC systems, DC systems and converter valves. In addition, the graph convolution method adopted by the present invention can construct a corresponding graph structure model according to the topological structure of different power systems, realize fault identification of different projects, and provide strong support for the operation, maintenance and fault analysis of DC transmission systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of a method for identifying faults in an AC / DC power transmission system based on spatiotemporal graph convolution provided by the present invention;

[0052] Figure 2 A direct current transmission simulation system and a fault setting point diagram according to an embodiment of the present invention;

[0053] Figure 3 It is a schematic diagram of a graph structure based on a DC system topology diagram and fault characteristics according to an embodiment of the present invention;

[0054] Figure 4 is an electrical quantity diagram of a fault measurement point of a direct current transmission system according to an embodiment of the present invention;

[0055] Figure 5 It is a graph showing changes in the loss function value after multiple trainings of the fault identification model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] See also Figure 1 The embodiment of the present invention provides a method for identifying faults in an AC / DC power transmission system based on spatiotemporal graph convolution, comprising the following steps:

[0058] S1. Taking the end-to-end flexible DC transmission system as an example, graph structure modeling is carried out to collect fault data under various working conditions and perform data preprocessing, that is, extracting data from the simulation waveform, and then composing the electrical quantity vector data of each measurement node into a graph data set G{G1, G2···, G n} and marked with a label. Preferably, step S1 specifically includes the following steps:

[0059] exist Figure 2 As shown, F1-F8 are respectively set with single-phase grounding short-circuit fault on the converter transformer grid side, three-phase grounding short-circuit fault on the converter transformer grid side, single-phase grounding fault on the AC connection bus, three-phase grounding fault on the AC connection bus, short-circuit grounding fault on the DC line, short-circuit fault between DC lines, short-circuit fault on the converter valve, and short-circuit fault on the converter valve submodule. The fault causes are divided into the above 8 categories according to the fault settings. Adjust the fault parameters, such as the grounding impedance size, fault duration, etc., and perform multiple fault simulations under different working conditions.

[0060] The key electrical measurement points in the DC system are set as ten nodes N1-N10, and the connection relationship of each node is determined according to the DC system topology diagram and fault characteristics to form the following Figure 3 The structure of the diagram is shown in Figure 1. The 10 key electrical measurement points selected in the DC system converter station are as follows: Figure 2 As shown, they are: AC system bus (node ​​N1), AC connection bus (node ​​N2), ABC three-phase upper bridge arm (nodes N3, N4, N5), ABC three-phase lower bridge arm (nodes N6, N7, N8), positive DC bus (node ​​N9), negative DC bus (node ​​N10). The measurement data of each node are as follows: instantaneous voltage value, instantaneous current value, voltage effective value, current effective value, active power, reactive power or zero-sequence current (measured only at the AC system bus).

[0061] Extract electrical data from the fault simulation waveform (electrical data can be found in Figure 4 The electrical waveform diagram of some nodes shown in the figure) is that in each fault that occurs, starting from the initial moment of 5s, the electrical quantity time series data of each measurement node at moments t-1, t, t+1, etc. are extracted at intervals of 0.01s within time T, and the moment t series data is represented in the form of a feature vector.

[0062] The electrical quantity vector data of each measurement node from the initial time to time t above constitutes the characteristic matrix {X0, X1···, X t}, n faults or normal operating conditions {X0, X1···, X t} composed of graph datasets G{G1, G2..., G n According to the simulation fault setting or the actual engineering fault result, each graph data G nLabel the graphs with eight types of labels, such as "single-phase ground short circuit fault on the converter transformer side, ground fault on the DC transmission line, and short circuit fault on the converter valve". If there is no such fault, no label is added. 90% of the data in the graph dataset is randomly selected as the training dataset, and the remaining 10% is used as the test dataset.

[0063] S2. Obtain the matrix M by performing graph convolution on the graph data t Preferably, step S2 specifically includes the following steps:

[0064] According to the connection relationship between the nodes in the graph structure, the adjacency matrix A of the nodes in the graph structure is formed T(i，j) . A T If the i node and the j node are connected, then the element A in the adjacency matrix T(i，j) If connected, it is 1, and if not connected, it is 0. Each node is not connected to itself.

[0065] The adjacency matrix is ​​added to the identity matrix I to increase the node's own characteristics, resulting in the following matrix A.

[0066] A=A T +I (1)

[0067] Let D be A T The degree matrix is ​​a diagonal matrix diag(). The elements on the diagonal of the degree matrix are the number of edges d(v) connecting each node to other nodes in the graph structure. i ), the expression is as follows:

[0068] D=diag(d(v1),d(v2),···,d(v N )) (2)

[0069] From the initial time to time t, the node feature data {X0, X1···, X t}, adjacency matrix, degree matrix to obtain the matrix M at each moment t :

[0070] M t =σ ReLU (D -1 / 2 AD -1 / 2 X t W 1 ) (3)

[0071] where σ ReLU is the ReLU nonlinear activation function, W 1 , W 2 is the weight parameter matrix.

[0072] The ReLU nonlinear activation function expression is as follows:

[0073]

[0074] S3, the data after the convolution of S2 graph is processed by the gated recurrent neural network, and the time dynamic features are extracted to obtain K t , K t After performing graph convolution again, the Softmax activation function is used to output the predicted value Z of each fault category. lf Preferably, the temporal dynamic feature extraction method is based on the “z t 、r t 、h t , K t " is composed of the operation process.

[0075] Specifically, z t To update the gate output, its expression is as follows:

[0076] z t =σ sig (W z1 M t +W z2 K t-1 +b z ) (5)

[0077] K t-1 represents the output of the recurrent neural unit at time t-1, M t Represents the node feature data input at time t. sig is the sigmoid activation function.

[0078] The activation function expression is as follows:

[0079]

[0080] r t To reset the gate output, its expression is as follows:

[0081] r t =σ sig (W r1 M t +W r2 K t-1 +b r ) (7)

[0082] h t is the candidate hidden state, the expression is as follows:

[0083]

[0084] The tanh activation function expression is as follows:

[0085]

[0086] Finally, the candidate hidden state, reset gate and update gate are used to obtain the recurrent neural unit output K at time t. t :

[0087]

[0088] In the above formula, W z1 , W z2 , W r1 , W r2 , W h1 , W h2 is the weight parameter, b z 、b r 、b h is the bias parameter. The Hadamard product The matrix calculation method is as follows:

[0089]

[0090] set up It is expressed by the following formula:

[0091]

[0092] After the above calculations, we get the output K of the recurrent neural unit at time t. t , and again with Perform graph convolution operation and use Softmax activation function σ softmax Output the predicted value Z of each fault category (and no-fault condition) lf , the process expression is as follows:

[0093]

[0094] The above S1, S2, and S3 fault category prediction process is the fault identification model.

[0095] The ith output of the Softmax activation function can be expressed as follows, where H is the output category (including no fault), z is the input vector with dimensions H, z i is the i-th element in the input vector, and the denominator is the exponential sum of all elements in the input vector:

[0096]

[0097] S4. Obtain the loss function of the fault identification model. Specifically, the cross entropy loss function of the fault identification model is:

[0098]

[0099] Where F represents the number of fault categories and l represents the graph structure y LThe graph data with fault labels in Y, that is, only the loss function of the graph with faults is calculated. lf represents the fault label value of each type in the S1 training or test data (when it is 1, the fault of this type occurs, which is a positive value, and when it is 0, the fault of this type does not occur, which is a negative value), Z lf Represents the predicted value of the fault category of the graph data. This expression represents the fault probability prediction error value of the model after removing the unlabeled graph data.

[0100] S5. Update the parameters of the fault identification model according to back propagation. Preferably, step S5 specifically includes the following steps:

[0101] Take the training data set and update the parameters in the model through back propagation. One of the weight parameters W h1 The following are examples of update methods:

[0102] According to the chain rule, we can find the loss function for the weight parameter W. h1 The partial derivative of .

[0103]

[0104] Update parameters:

[0105]

[0106] S6. After updating the parameters in the fault identification model, take the test data set and repeat steps S2, S3, and S4. If the loss function value does not meet the threshold requirement, continue to repeat steps S2 to S5 to update the parameters until the loss function drops below the threshold, and then end the model training. The preset threshold can be set to 0.3. The loss function value after repeated training can be seen Figure 5 shown.

[0107] S7, select a newly occurring fault event, extract the electrical data of each node of the DC system from the initial time of 1.4s to the end time of 1.9s of the event from the fault waveform at intervals of 0.01s, convert the data into the format of a feature matrix according to S1, input the trained model in S6, repeat steps S2 and S3, and obtain the maximum probability calculated value of the short-circuit fault of the converter valve submodule among various types of faults, which is the identified fault category. The present invention extracts the fault electrical quantity characteristics of multiple points in the power system, such as Figure 4 The waveform diagrams of electrical sampling of some nodes are illustrated, namely: "A phase bridge arm current, U1 submodule voltage" in the A phase upper bridge arm N3 node, "positive DC bus to ground voltage, DC bus current" measured electrical quantities in the positive DC bus N6 node, and "AC voltage instantaneous value, AC current instantaneous value" measured electrical quantities in the AC connection bus node N2.

[0108] The DC power transmission system fault identification method based on graph convolution proposed in the present invention decomposes the nodes of each part of the DC power transmission system into a graph structure, and analyzes the complex relationship between the nodes in the graph structure by efficiently processing multiple graph data, thereby obtaining the fault characteristics. The present invention also has good scalability and sparse data processing capabilities, and the fault identification nodes can also be expanded when the power system structure changes.

[0109] Table 1 below is a comparison of the GCN-GRU (temporal graph convolutional neural network), CNN (convolutional neural network) and DNN (deep neural network) models used in the present invention for the recognition capabilities of various fault categories including short-circuit faults of converter valve submodules, and the dimensions of data sampling types. The existing neural network model cannot simultaneously identify faults in AC / DC systems and converter valves, and its universality is relatively weak; the present invention integrates the multi-point electrical data features of the power system to simultaneously identify multiple types of faults in multiple areas such as AC systems, DC systems and converter valves.

[0110] Table 1

[0111]

[0112]

[0113] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for fault identification of AC and DC power transmission systems based on spatiotemporal graph convolution, characterized in that: The steps include: S1. Perform graph structure modeling, collect DC system fault data under various working conditions, construct a graph data set, and divide the graph data set into a training data set and a test data set; S2. Obtain the matrix M by performing graph convolution on the graph data in the graph dataset t ; S3, the matrix M after the convolution operation of S2 t After being processed by the gated recurrent neural network, the temporal dynamic features are extracted to obtain K t , K t After performing graph convolution again, the Softmax activation function is used to output the predicted value Z of each fault category. lf ; S4, combining the fault category prediction process of steps S1-S3 into a fault identification model, and obtaining a loss function of the fault identification model; S5, take the training data set and update the parameters of the fault identification model in S4 through back propagation; S6. After updating the parameters of the fault identification model, take a test data set and repeat steps S2 to S5 until the loss function value of the test data set is lower than a preset threshold, then end the model training and obtain a trained fault identification model; S7. Select DC system fault event data outside the data set in S1, and input the fault identification model trained in S6 to perform fault identification.

2. The method for identifying AC / DC power transmission system faults based on spatiotemporal graph convolution according to claim 1, characterized in that: Step S1 specifically includes: Set each key electrical measurement point in the DC system to N x The connection relationship of each node is determined according to the DC system topology and fault characteristics, and the graph structure is modeled; various types of faults are set in the DC simulation system, fault parameters are adjusted, and fault simulation is performed under different working conditions; data is extracted from the simulation waveform, that is, the electrical quantity vector data of each measurement node within time T in n faults or normal working conditions are combined into a characteristic matrix {X0, X1···, X t }, forming a graph data set G{G1, G2···, G n }, according to the simulation fault settings or actual engineering fault results, each graph data G n The faulty graph is labeled, and no label is added when there is no fault; data in the graph dataset is randomly selected as the training dataset and the test dataset respectively.

3. The method for identifying AC / DC power transmission system faults based on spatiotemporal graph convolution according to claim 1, characterized in that: Step S2 specifically includes: The graph convolution operation is performed based on the node features of the modeled graph structure. The features of the graph structure include the node adjacency matrix A T(i，j) and A T The degree matrix, A T If the i node and the j node are connected, then the element A in the adjacency matrix T(i，j) If it is not connected, it is 1, and if it is not connected, it is 0. Each node is not connected to itself; The adjacency matrix is ​​added with the identity matrix I to increase the node's own characteristics, resulting in the following matrix A: A=A T +I (1) Let D be A T The degree matrix is ​​a diagonal matrix diag(), and the elements on the diagonal of the degree matrix are the number of edges d(v) connecting each node to other nodes in the graph structure. i ), the expression is as follows: <h2 style=";text-align:left;direction:ltr">D = diag(dv1),d(v2),···,d(v<h2 style=";text-align:left;direction:ltr"> N <h2 style=";text-align:left;direction:ltr"> )) (2) The node feature data X from the initial time to time t t , adjacency matrix A T(i，j) , degree matrix D to get the matrix M at each moment t : M t =σ ReLU (D -1 / 2 AD -1 / 2 X t W 1 ) (3) where σ ReLU is the ReLU nonlinear activation function, W 1 , W 2 is the weight parameter matrix; The expression of ReLU nonlinear activation function is as follows:

4. The method for identifying faults in AC and DC power transmission systems based on spatiotemporal graph convolution according to claim 3, characterized in that: Step S3 specifically includes: The temporal dynamic feature extraction method is based on the "z" in the recurrent neural unit from the initial moment, t-1 to time t. t 、r t 、h t , K t " is composed of the following calculation process: z t =σ sig (W z1 M t +W z2 K t-1 +b z ) (5) r t =σ sig (W r1 M t +W r2 K t-1 +b r ) (6) The tanh activation function expression is as follows: Hadamard product The matrix calculation method is as follows: where z t is the update gate output, r t is the reset gate output, h t is the candidate hidden state, K t is the output of the recurrent neural unit at time t; K t-1 represents the output of the recurrent neural unit at time t-1, M t represents the node feature data input at time t, σ sig is the sigmoid activation function; W z1 , W z2 , W r1 , W r2 , W h1 , W h2 is the weight parameter, b z , b r , b h is the bias parameter; set up It is expressed by the following formula: After the above calculations, we get the output K of the recurrent neural unit at time t. t , and again with Perform graph convolution operation and use Softmax activation function σ softmax Output the predicted value Z of each fault category lf : The S1, S2, and S3 fault category prediction process is the fault identification model.

5. The method for identifying faults in AC and DC power transmission systems based on spatiotemporal graph convolution according to claim 4, characterized in that: Step S4 specifically includes: The cross entropy loss function expression of the fault recognition model is as follows: Where F represents the number of fault categories and l represents the graph structure y L The graph data with fault labels, Y lf represents the various fault label values ​​in S1 training or test data, Y lf When it is 1, the fault occurs, which is a positive value; when it is 0, the fault does not occur, which is a negative value. lf Represents the fault category prediction value of graph data.

6. The method for identifying faults in AC and DC power transmission systems based on spatiotemporal graph convolution according to claim 1, characterized in that: The back propagation refers to a method for calculating the parameter gradient of each layer of the neural network. The chain rule is used to find the partial derivatives of the loss function with respect to the weights and biases of each neuron layer by layer, and the gradient of the loss function with respect to the weight and bias vector is constructed as the basis for modifying the weights.

7. The method for identifying faults in AC and DC power transmission systems based on spatiotemporal graph convolution according to claim 1, characterized in that: Step S7 specifically includes: selecting a newly occurring event that requires fault identification, extracting data from the DC system from the initial moment of the event to the end moment of time T, converting the data into a feature matrix format according to S1, inputting the fault identification model trained in S6, repeating steps S2 and S3, obtaining various fault probability values, and selecting the maximum probability as the identified fault category.

Images