Power system cascading failure prediction method and equipment based on dynamic data driving

Abstract

The invention provides a power system cascading failure prediction method and device based on dynamic data driving, and the method comprises the steps: building a hybrid differential algebraic equation set of a power system, and generating a cascading failure original sample through dynamic simulation; performing feature extraction on the samples to obtain time sequence node features and time sequence edge features, and constructing a feature data set; sample label processing: coding a next-level fault line number set according to the fault chain information, and combining the next-level fault line number set with the characteristic data to form a cascading fault data set; constructing a time graph edge convolutional network encoder of a coupling edge condition convolutional graph neural network and a gating circulation unit; and carrying out time graph edge convolutional network encoder training by adopting a contrast learning architecture, and finally realizing cascading failure path prediction. According to the method, data generation and prediction of the cascading failure of the power system can be realized; on one hand, a traditional cascading failure prediction framework based on a steady-state load flow calculation method is broken through; and on the other hand, the problem of non-uniform sample distribution caused by weak links of the power system is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system cascading failure prediction, and in particular relates to a novel power system cascading failure prediction technical solution based on dynamic data drive. Background Art

[0002] Cascading failures in power systems are one of the main causes of major blackouts. Their prediction and prevention have become a critical issue for the safe and stable operation of new power systems. The North American Electric Reliability Corporation (NERC) defines a cascading failure as the uncontrolled, sequential failure of system components triggered by a disturbance at any location. The integration of high levels of renewable energy significantly increases system operational uncertainty while reducing system robustness, leading to a surge in the risk of cascading failures. Globally, incidents such as the 2019 UK blackout and the 2023 Brazilian blackout demonstrate that even minimal initial disturbances can propagate through multiple spatial and temporal couplings and trigger catastrophic cascading effects. Accurately predicting the subsequent fault propagation paths at the early stages of a cascading failure can greatly enhance the ability of grid operators to promptly prevent cascading failures and minimize system losses. However, with the increasing prominence of the "dual high" characteristics of new power systems, grids are exhibiting increasingly robust nonlinearities and multiple time constants. This makes the development of cascading failure propagation paths difficult to predict, posing a significant challenge for operators in preventing the propagation of faults in a timely manner.

[0003] Cascading failure simulation is the main means of analyzing cascading failures in power systems and is also the basis for studying cascading failure prediction methods. For cascading failure simulation, scholars have adopted methods based on power physics, simulation-based technologies, probability-based models, and graph-based modeling and analysis. Among them, methods based on complex system theory, represented by the OPA model, have made great progress due to their excellent physical interpretability. It is worth noting that the dynamic simulation-based methods proposed in recent years, such as COSMIC, have broken through the traditional power flow calculation framework and fully characterized the transient characteristics during the propagation of cascading failures. Compared with the cascading failure samples obtained by methods based on the steady-state power flow calculation framework, they are closer to the actual situation. However, the huge amount of data also places higher demands on the cascading failure prediction model.

[0004] In recent years, machine learning technology, with its ability to model massive amounts of data, has provided new insights into cascading failure prediction. Currently, academic research on cascading failure prediction focuses on three dimensions: fault scale prediction, fault stage classification, and fault path deduction. In the field of path prediction, some studies have constructed Bayesian belief networks and multi-attribute decision models to achieve cascading path prediction through line vulnerability scoring; some have innovatively integrated knowledge graphs with machine learning to establish a fault correlation analysis framework for AC / DC hybrid power grids; and some have explored complete fault path prediction using Markov decision processes and time-varying graph recurrent neural networks. However, it is worth noting that existing research results still have two major limitations: on the one hand, the data used for model training mostly comes from static power flow simulation. Although this type of method has significant computational efficiency advantages in the data set preparation stage, its lack of modeling of key spatiotemporal characteristics may lead to inaccurate characterization of the fault process; on the other hand, whether it is an actual power grid or a simulation system, there are bound to be weak links due to the grid structure, protection setting settings and even the actual age of the equipment. After the initial disturbance of the system, the failure rate of the weak link will be significantly higher than that of the relatively more robust components. The uneven sample distribution caused by this phenomenon has not been considered in existing research, which may lead to the model's insufficient ability to recognize fault modes with low occurrence frequency but higher accident risk. Summary of the Invention

[0005] The present invention mainly solves two technical problems existing in the existing technology. First, it realizes the prediction of cascading failures based on dynamic data-driven method based on the coupling method of graph convolutional network and time series learning neural network, breaking through the framework of traditional steady-state power flow calculation method. Second, it overcomes the problem of uneven sample distribution caused by weak links in the power system through data enhancement method and contrastive learning architecture.

[0006] The technical solution of the present invention provides a method for predicting power system cascading failures based on dynamic data drive, which performs the following process: Establish a hybrid differential algebraic equation system for the power system and generate original samples of cascading failures through dynamic simulation; Extract features from samples to obtain time series node features and time series edge features, and construct a feature dataset; Sample label processing, including encoding the next-level fault line number set according to the fault chain information and combining it with the feature data to form a chain fault dataset; Construct a temporal graph edge convolutional network encoder that couples edge-conditional convolutional graph neural networks with gated recurrent units; A contrastive learning architecture is used to train the temporal graph edge convolutional network encoder, ultimately achieving cascading failure path prediction.

[0007] Moreover, the mixed differential algebraic equations include continuous state variables, continuous algebraic variables and relay action binary vectors in the power system.

[0008] Furthermore, the method of generating the original sample of cascading failure by dynamic simulation includes: Randomly set node loads and maximum renewable energy power to calculate the AC optimal power flow; initialize differential equation state variables and set simulation step sizes; randomly disconnect two lines as the initial fault, and the fault clearing time follows a normal distribution; When the system is decoupled, an independent simulation of the subsystem is generated, the mixed differential algebraic equations are solved simultaneously, the protection action is detected and the system status is updated; the simulation is iterated until the maximum simulation time is reached.

[0009] Moreover, the timing node features include voltage amplitude / phase angle, frequency change rate, load power and power generation equipment power; the timing edge features include line active / reactive power, equivalent impedance parameters and transformer ratio.

[0010] Moreover, in the temporal graph edge convolutional network encoder, an edge-conditional convolutional graph neural network is first set up, and node feature convolution based on edge feature values ​​is performed according to the features corresponding to each moment in the temporal node features and temporal edge features; the output of the edge-conditional convolutional graph neural network at each moment is input into the corresponding gated recurrent unit in chronological order, and the output of the last gated recurrent unit is used as the corresponding latent space embedding vector of the chain failure sample output by the encoder.

[0011] Moreover, the described contrastive learning architecture is used to train the temporal graph edge convolutional network encoder, including taking temporal node features, temporal edge features and sample labels as input, constructing positive and negative sample pairs through data enhancement, optimizing encoder parameters through pre-training, and using a contrastive loss function in the embedding space during pre-training; optimizing classifier parameters through fine-tuning, and using multi-label binary cross entropy loss as the classification loss function during fine-tuning.

[0012] Moreover, the encoder parameters are also fine-tuned while optimizing the classifier parameters.

[0013] On the other hand, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for predicting power system chain failures driven by dynamic data as described above is implemented.

[0014] On the other hand, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the power system cascading failure prediction method based on dynamic data driving as described above is implemented.

[0015] On the other hand, the present invention provides a computer program product, including a computer program, characterized in that when the computer program is executed by a processor, the power system cascading failure prediction method based on dynamic data driving as described above is implemented.

[0016] This paper proposes a power system cascading failure prediction scheme based on a time-graph edge-conditional convolutional network using dynamic simulation and contrastive learning. Its innovation lies in two aspects: first, it establishes a dynamic data-driven paradigm based on hybrid differential-algebraic equations, overcoming the information limitations of traditional static simulation; second, it constructs a self-supervised training mechanism embedded in contrastive learning, which enhances the model's ability to identify small-sample failure modes through sample augmentation and latent space mapping.

[0017] The solution of the present invention is simple and convenient to implement and has strong practicality. It solves the problems of low practicality and inconvenience in actual application existing in related technologies, can improve user experience, and has important market value. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of dynamic simulation of cascading failures according to an embodiment of the present invention.

[0019] Figure 2 This is a structural diagram of a time graph edge convolutional network encoder according to an embodiment of the present invention.

[0020] Figure 3 This is a structural diagram of a multi-label prediction task classifier according to an embodiment of the present invention.

[0021] Figure 4 This is a diagram of the comparative learning architecture of a time graph edge conditional convolutional network according to an embodiment of the present invention.

[0022] Figure 5 This is a dynamic response diagram of the frequency change rate at the fault node during a cascading failure in an embodiment of the present invention.

[0023] Figure 6 4 is a diagram showing a dynamic response of the load rate at the fault line during a cascading fault in an embodiment of the present invention.

[0024] Figure 7 4 is a diagram showing the dynamic response of voltage at a fault node during a cascading fault in an embodiment of the present invention.

[0025] Figure 8 The prediction results of each line of the original model of the embodiment of the present invention are F 1 Score graph.

[0026] Figure 9 The prediction results of each line when verifying the influence of the comparative learning module on the model prediction performance in the embodiment of the present invention are F 1Score diagram. DETAILED DESCRIPTION

[0027] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings and embodiments to fully understand the purpose, characteristics and effects of the present invention.

[0028] The method for predicting cascading failures in a power system proposed in the present invention includes constructing a hybrid differential-algebraic equation mathematical model for the power system, designing a repeated cascading failure simulation process to obtain a cascading failure sample set based on dynamic data, and preprocessing the raw data obtained from the simulation to form a normalized feature data set. Furthermore, an encoder is constructed that couples boundary-conditional convolutional graph neural networks and gated recurrent units, and a learning architecture for a time-boundary-conditional convolutional graph neural network based on data enhancement and contrastive learning is designed to achieve fault path prediction for cascading failures. The present invention enables data generation and prediction of cascading failures in power systems, breaking through the traditional cascading failure prediction framework based on steady-state power flow calculation methods and overcoming the uneven sample distribution problem caused by weak links in the power system.

[0029] The present invention provides a method for predicting power system cascading failures based on dynamic data. The implementation process includes the following steps: Step 1, dynamic simulation of cascading failures: The technical solution of the present invention is particularly suitable for new power systems and can also be used in other traditional power systems.

[0030] In the embodiment, a new power system is taken as an example for explanation. First, a mixed differential algebraic equation of the new power system is established: (1) (2) (3) Where, Indicates the current moment, x represents the continuous state variables in the power system, such as generator speed or angle, which are expressed according to the differential equation in formula (1) f () changes with time, express x right Seek derivation, Represents the continuous state variable at the current moment x ; y Represents continuous algebraic variables in power systems, such as the amplitude or phase angle of node voltages, which usually have only algebraic relationships with other variables. g ( , Continuous algebraic variable representing the current moment y ; zis a set of binary vectors representing the relay action, which is composed of discrete logic equations h ()Sure, A binary vector representing the relay action status at the current moment.

[0031] By configuring its component parameters and protection parameters, repeated simulation of new power system cascading failures is carried out to obtain the original samples of new power system cascading failures.

[0032] See also Figure 1 In step 1 of the embodiment, the specific operation method for performing repeated simulation of a new type of power system cascading failure is preferably as follows: Step 1.1: Randomly assign the operating status. You can randomly assign the total load size of each node and the maximum power of the new energy source, and then proceed to step 1.2; Step 1.2: Calculate the optimal power flow. Based on the randomly assigned load level, the AC optimal power flow is calculated with the goal of minimizing network losses. The AC optimal power flow (OPF) calculation can be performed using existing techniques and will not be described in detail in this invention. During the calculation, the distributed generation outputs power in a fixed proportion to the load size of the node where it is located, and is not used as a decision variable in the optimal power flow problem. The process then proceeds to step 1.3. Step 1.3: Initialize the system state. According to the optimal power flow calculation results obtained in step 1.2, initialize the state variables of differential equation (1). In addition, initialize the current moment 0s, time domain simulation step 0.01 s, free to set the maximum simulation time , then proceed to step 1.4; Step 1.4: Initial fault selection. Randomly select two transmission line disconnections as the initial faults. The fault clearing time is sampled using a normal distribution with a mean of 0.15s and a standard deviation of 0.1s. Then proceed to step 1.5. Step 1.5: Check system connectivity. If the current fault causes the system to be disconnected, proceed to step 1.6; otherwise, proceed directly to step 1.7. Step 1.6: Generate data for each subsystem and simulate the subsequent process of cascading failures within each subsystem, update the current system's admittance matrix, and solve the algebraic equation system (2). The admittance matrix is ​​a matrix composed of the conductance values ​​(or resistance and admittance values) between each node in the power grid. It is the process variable for power flow calculation. When the line is in the disconnected state, directly remove the component representing the corresponding line from the system admittance matrix, and then proceed to step 1.7; Step 1.7: Solve the algebraic differential equations simultaneously. If convergence occurs, go to step 1.9; otherwise, determine the step size. Is it greater than the minimum step size? If it is greater than, let and solve again; otherwise, go to step 1.8; Step 1.8: Set the current subsystem to be unstable, all loads are cut off, end the current subsystem simulation, and continue the simulation of other subsystems, and go to step 1.9; Step 1.9: Calculate the relay protection logic equation (3) based on the system operating status calculated in step 1.7, and update the protection trigger status. If there is a protection action, go to step 1.10, otherwise go to step 1.11; Step 1.10: Update the system status and current time (i.e., current moment) based on the protection action obtained in step 1.9. Then proceed to step 1.5; Step 1.11: Update the current time (i.e. the current moment) Then proceed to step 1.12; Step 1.12: If the current time Greater than the preset dynamic simulation end time , go to step 1.14, otherwise go to step 1.13; Step 1.13: Increase the simulation step size to , to improve the simulation efficiency when no new discrete events occur, and then go to step 1.7; Step 1.14: After the cascading failure simulation is complete, calculate the load loss caused by this failure and output the fault record. Go to step 1.1 to start the next repeated simulation.

[0033] Step 2, feature selection and feature data set construction: For each new power system original sample obtained in step 1, key variables are selected to form time series node feature vectors and time series edge feature vectors.

[0034] The time series node features extracted by the present invention are related to the grid nodes, including voltage amplitude / phase angle, frequency change rate, load power and power of power generation equipment; The time series edge features extracted by the present invention are related to the power grid lines, including line active / reactive power, equivalent impedance parameters and transformer ratio; The preferred solution in the embodiment is: The variables included in the time series node feature vector are: node voltage amplitude, node voltage phase angle, node frequency, node frequency change rate, node load active power, node load reactive power, node synchronous machine active power, node synchronous machine reactive power, node synchronous machine power angle, node synchronous machine speed, node wind turbine active power, node wind turbine reactive power, node photovoltaic active power, node photovoltaic reactive power, node distributed active power, node distributed reactive power; The variables included in the time series edge feature vector are: active power from the beginning to the end of the line, active power from the end to the beginning of the line, reactive power from the beginning to the end of the line, reactive power from the end to the beginning of the line, line equivalent resistance, line equivalent reactance, line equivalent conductance, line equivalent susceptance, and line transformer ratio.

[0035] Interpolation, normalization, and masking steps are then performed in sequence to form the final feature dataset for neural network training.

[0036] In specific implementation, it is recommended to set the time series node feature vector and time series edge feature vector according to Table 1. Table 1 Feature vectors for cascading failure prediction

[0037] In step 2 of the embodiment, the specific operation method of preferably forming the feature data set ultimately used for neural network training by using the selected time series node feature vectors and time series edge feature vectors is as follows: Step 2.1: Interpolation. For each feature in each sample, use the cubic spline interpolation method with a fixed time interval of 0.02s to reconstruct it into a sequence of equally spaced timestamps.

[0038] Step 2.2: Normalization. Take the maximum and minimum values ​​of each feature within the entire sample range, and normalize all corresponding features based on the maximum and minimum values.

[0039] Step 2.3: Masking: For each feature that does not exist on a node, set it to 0.

[0040] Step 3: Sample label processing: For each new power system original sample obtained in step 1, the line number set of the next-level line fault after the initial disturbance occurs is obtained based on the fault chain information, encoded as a label, and combined with the corresponding time series node features and time series edge features to form a new power system cascading failure dataset.

[0041] Step 4, temporal graph edge convolutional network encoder construction: define an encoder that couples edge-conditional convolutional graph neural network and gated recurrent unit (GRU).

[0042] See also Figure 2 , the encoder uses the temporal node features obtained in step 2 X node ∈ R T×nn×dn and temporal edge features X edge ∈ R T×ne×de is the input, where R represents the field of real numbers,T Indicates the number of timestamps. nn Indicates the number of network nodes, ne represents the number of network edges, dn represents the number of node features, de Indicates the number of edge features. It should be noted that the time series node feature X node and temporal edge features X edge They are all three-dimensional arrays, the first dimension of which is the time dimension, the second dimension is the order dimension of the node or edge, and the third dimension is the order dimension of the feature quantity of the node or edge. For the convenience of description, the subscripts of the time series node or edge features in the subsequent description are t Indicates time, i and j Indicates the node number. If a single subscript appears, it indicates a slice of the feature array in the corresponding dimension, such as Indicates that the time series node feature is in the time dimension t The feature matrix at each moment has the dimension nn × dn .akin, Indicates that the time series edge feature is in the time dimension t The feature matrix at each moment.

[0043] For the temporal graph edge convolutional network encoder, first, an edge-conditional convolutional graph neural network is set, in which a multi-layer edge-conditional graph convolutional network is set. The features corresponding to each moment in the temporal node features and temporal edge features are convolved with the node features based on the edge feature values ​​using the following formula:

[0044] In the formula Representative l Layer-wise edge-conditional graph convolutional network i Nodes at time t Similarly, the eigenvector of Representative l -1 layer edge conditional graph convolutional network i Nodes at time t The eigenvector of () represents the activation function, Representation node i Neighborhood, Representation node j For nodes i Neighbor nodes of Representative l -1 layer edge conditional graph convolutional network by edge The convolution kernel weight matrix determined by the features is calculated as follows

[0045] In the formula represents a parameterized function, which in this invention is a multilayer perceptron, e ij For the edge E ij The eigenvector of θ are the learnable parameters in the parameterized function.

[0046] Then, the edge conditional convolutional graph neural network outputs each moment x 1. x 2. ... x t 、…、 x T Will be input into the corresponding gated recurrent units GRU1, GRU2, ..., GRU in chronological order t , ..., GRU T , and each GRU will output an embedding vector in the latent space h 1. h 2. ... h t 、…、 h T , the last GRU T Output h T It will be used as the output of the encoder, which is also called the latent space embedding vector of the chain failure sample in the present invention.

[0047] Step 5: Training of temporal graph edge-conditional convolutional network based on contrastive learning: The present invention proposes to construct a Figure 2 The temporal graph edge convolutional network shown is the contrastive learning architecture of the encoder, and its structure is as follows Figure 4 As shown. This architecture uses the time series node features and time series edge features obtained in step 2 and the sample labels obtained in step 3 Y The encoder and classifier are trained by taking the input data as input, constructing positive and negative sample pairs through data augmentation, optimizing the encoder parameters through pre-training, optimizing the classifier parameters through fine-tuning, and making the fine-tuning of the encoder parameters optional. Finally, the effects of the trained encoder and classifier can be verified through the test step.

[0048] To address the issue of skewed sample distribution, this paper preferentially utilizes contrastive representation learning techniques in conjunction with data augmentation to enhance the neural network model's ability to learn small sample events. The SimCLR (simple framework for contrastive learning of visual representations) framework has significant potential in scenarios where the number of labels is insufficient. It can address the imbalanced sample size of power system cascading failures and support the prediction of power system cascading failures based on dynamic simulation data.

[0049] In step 5 of the embodiment, the specific operation method preferably adopted for training the encoder and the classifier is as follows: Step 5.1, data enhancement: For each set of time series node features and time series edge features obtained in step 2 x k , the subscript here k Represents the number of the sample in the sample set, and forms a new set of time series feature vectors through four sub-steps: amplitude scaling, length scaling, noise addition, and smoothing. x' k , this new set of time series feature vectors x' k With the original feature vector group x k is called a pair of positive samples. Furthermore, step 5.1 contains the following four sub-steps: Step 5.1.1, amplitude scaling: randomly select subsequences from the original time series node features and time series edge features, and randomly adjust the amplitudes of these subsequences.

[0050] Step 5.1.2: Length Scaling: Select new random subsequences from the amplitude-scaled time series features, scale the lengths of these subsequences, and reconstruct them, ensuring that the total length of the sequence remains unchanged.

[0051] Step 5.1.3: Add noise: Add random Gaussian noise to the time series features after amplitude and length scaling.

[0052] Step 5.1.4: Smoothing: Randomly select some sample points from the noise-added features and perform forward and backward interpolation.

[0053] Step 5.2, Pre-training: The encoder receives the original sample x k , enhanced samples x' k and other samples x g , and x k and x'k As a positive pair, x k and x g As a negative pair. The encoder then maps all samples into the embedding space based on the working process in step 4 to form the corresponding latent space embedding vector h g 、 h' k 、 h g , contrastive loss is used to measure the relative distance between positive and negative pairs in the embedding space:

[0054] In the formula e is a natural constant, N is the total number of samples, s(·) represents the cosine similarity between two vectors, τ represents the temperature coefficient, and the denominator represents the sum of the cosine similarities of all items, where the upper bound of the sum is 2N-1 This is because in the batch process, each batch has N The original samples and N An enhanced sample.

[0055] During the pre-training phase, the contrastive loss function trains the trainable parameters in the encoder through back-propagation.

[0056] Step 5.3, fine-tuning: see Figure 3 , the embodiment connects a fully connected layer after the pre-trained encoder, and sets an activation layer sigmoid and an output layer as a classifier for each label category. At this time, the model will receive the original sample x k And generate predicted labels , the subscript here k Represents the number of the sample in the sample set. Using the predicted label and the true label Y k Calculate the classification loss, which will be used to optimize the classifier parameters. The present invention further proposes that the encoder parameters can be fine-tuned while optimizing the classifier parameters. It should be emphasized that in a cascading failure of the power system, a disturbance may cause more than one next-level failure. Therefore, cascading failure prediction is not a multi-classification task but a multi-label classification task. Therefore, the predicted label and the true label are actually a label vector. The true label vector Y k Contains elements y 1 、y 2 、...、y dn .

[0057] In the fine-tuning stage, multi-label binary cross entropy loss is used as the classification loss function:

[0058] In the formula num_l Indicates the total number of tags. num_s represents the total number of samples, Indicates the n_s The first sample n_l The true value of the label, Indicates the n_s The first sample n_l The predicted value of the label.

[0059] Step 5.4: Test samples on which the model has neither been enhanced nor trained x test Input into the fine-tuned encoder to generate the embedding vector of the test sample h test , and generate predicted labels through the classifier , compare the predicted labels and the true labels to test the model performance.

[0060] To verify the technical effectiveness of this invention, a dynamic simulation was conducted on a modified IEEE 39-node system. The synchronous generator at node 34 in the original system was replaced with a photovoltaic power plant, and the synchronous generators at nodes 36 and 37 were replaced with wind farms. Distributed photovoltaics were connected at nodes 4, 8, 15, 20, 24, and 29 to simulate the characteristics of the new power system. After the modification, the installed capacity of renewable energy in the system accounted for 27.45%, and the distributed power generation capacity accounted for approximately 10% of the load capacity.

[0061] Calculation results: 1) Dynamic simulation results of cascading failures Based on step 1, a large number of new power system cascading failure samples can be obtained. This calculation result selects a representative sample to show the complete fault chain as shown in Table 2. The changes of fault-related electrical quantities during the cascading failure process are shown in Table 2. Figure 5-7 shown.

[0062] Table 2 Examples of fault chains obtained based on dynamic simulation

[0063] In this cascading fault simulation, the initial fault was set as lines 13-14 and 23-24, which failed and disconnected 1 second after the simulation started. The fault was cleared 0.12 seconds later, and the two lines were reclosed.

[0064] As can be seen, while the initial disturbance did not directly trigger a power flow shift sufficient to trip subsequent lines, it did cause distributed renewable energy to disconnect from the grid due to the sustained frequency fluctuations. This in turn indirectly triggered the overload of heavily loaded lines 10-13 and 13-14, ultimately triggering line protection and disconnection. Following this subsequent disconnection, voltage at some nodes began to drop, triggering the low-voltage load shedding device.

[0065] 2) Cascading failure prediction results After repeated dynamic simulation of cascading failures, a total of 1,000 valid cascading failure samples were generated. 20%, or 200 samples, were selected as the test set and excluded from model pre-training and fine-tuning. The remaining 800 samples were input into the constructed cascading failure prediction model as the training set. The test set samples were then input into the trained model, and the output prediction labels were compared with the true labels. The overall accuracy reached 95%. The prediction results for each line are shown below. Figure 8 shown. Figure 8 in F The formula for calculating the score is:

[0066]

[0067]

[0068] In the formula TP is the number of samples whose predicted value is 1 and whose true value is also 1, FP The predicted value is 1 but the true value is 0. FN is the number of samples whose predicted value is 0 but the true value is 1; precision is the precision rate, which represents the proportion of samples predicted to be correct among all samples predicted to be positive. recall It is the recall rate, which represents the proportion of samples predicted by the model among all actual positive examples.

[0069] Furthermore, to comprehensively evaluate the performance of the proposed cascading failure prediction model, and to compare the performance of the contrastive learning method with that without data augmentation, we directly used the temporal graph edge convolutional network to learn the temporal graph structural features in the samples on the same dataset. The experimental results show that the final accuracy of the prediction using the temporal graph edge convolutional network reached 78%. However, the performance of this learning strategy on the small sample labels was not ideal. Figure 9 Shows the lines F 1 score comparison results show that except for a few routes with large sample sizes, the prediction results of the remaining routes have declined. F 1 score is even lower than 0.5.

[0070] Furthermore, in order to verify the superiority of the proposed temporal graph edge convolutional network model in considering network topology information, this paper designed an ablation experiment to compare the performance of the temporal graph edge convolutional network, the temporal graph neural network model without considering edge conditional convolution, and the temporal learning model without considering graph properties on the same batch of data sets and presented the results in Appendix 3.

[0071] Table 3 F1 score and overall accuracy of each model’s prediction results for each route

[0072] The results show that the prediction performance of the temporal graph edge convolutional network is significantly better than that of a temporal graph neural network that does not consider edge information, both for individual lines and overall prediction performance. The temporal graph neural network model also significantly outperforms a temporal learning model that does not consider graph topology. This demonstrates that considering network topology and information contained in system lines can significantly improve the performance of the prediction model in the cascading failure prediction scenario, validating the necessity of the graph convolution module in the proposed method.

[0073] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0074] The following describes the new type of power system cascading failure prediction electronic device based on dynamic data drive provided by the present invention. The new type of power system cascading failure prediction electronic device based on dynamic data drive described below and the new type of power system cascading failure prediction method based on dynamic data drive described above can be referenced to each other.

[0075] The electronic device may include a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may invoke logic instructions in the memory to execute a novel power system cascading failure prediction method based on dynamic data drive, which primarily includes the software processing portion of the aforementioned steps.

[0076] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0077] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the software processing part of the new power system cascading failure prediction method based on dynamic data driving provided by the above methods.

[0078] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the software processing part of the new power system chain failure prediction method based on dynamic data drive provided by the above methods.

[0079] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0080] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for predicting cascading failures in power systems based on dynamic data, characterized by: Carry out the following process, Establish a hybrid differential algebraic equation system for the power system and generate original samples of cascading failures through dynamic simulation; Extract features from samples to obtain time series node features and time series edge features, and construct a feature dataset; Sample label processing, including encoding the next-level fault line number set according to the fault chain information and combining it with the feature data to form a chain fault dataset; Construct a temporal graph edge convolutional network encoder that couples edge-conditional convolutional graph neural networks with gated recurrent units; A contrastive learning architecture is used to train the temporal graph edge convolutional network encoder, ultimately achieving cascading failure path prediction.

2. The method for predicting power system cascading failures based on dynamic data drive according to claim 1, characterized in that: The mixed differential algebraic equation group includes continuous state variables, continuous algebraic variables and relay action binary vectors in the power system.

3. The method for predicting power system cascading failures based on dynamic data drive according to claim 1, characterized in that: The method of generating the original sample of cascading failure by dynamic simulation includes: Randomly set node loads and maximum renewable energy power to calculate the AC optimal power flow; initialize differential equation state variables and set simulation step sizes; randomly disconnect two lines as the initial fault, and the fault clearing time follows a normal distribution; When the system is decoupled, an independent simulation of the subsystem is generated, the mixed differential algebraic equations are solved simultaneously, the protection action is detected and the system status is updated; the simulation is iterated until the maximum simulation time is reached.

4. The method for predicting power system cascading failures based on dynamic data drive according to claim 1, characterized in that: The timing node features include voltage amplitude / phase angle, frequency change rate, load power and power generation equipment power; the timing edge features include line active / reactive power, equivalent impedance parameters and transformer ratio.

5. The method for predicting power system cascading failures based on dynamic data drive according to claim 1, characterized in that: In the temporal graph edge convolutional network encoder, an edge-conditional convolutional graph neural network is first set up, and node feature convolution based on edge feature values ​​is performed according to the features corresponding to each moment in the temporal node features and temporal edge features; the output of the edge-conditional convolutional graph neural network at each moment is input into the corresponding gated recurrent unit in chronological order, and the output of the last gated recurrent unit is used as the corresponding latent space embedding vector of the chain failure sample output by the encoder.

6. The method for predicting power system cascading failures based on dynamic data drive according to claim 1, characterized in that: The temporal graph edge convolutional network encoder is trained using a contrastive learning architecture, including taking temporal node features, temporal edge features, and sample labels as input, constructing positive and negative sample pairs through data augmentation, optimizing encoder parameters through pre-training, and using a contrastive loss function in the embedding space during pre-training; and optimizing classifier parameters through fine-tuning, using a multi-label binary cross entropy loss as the classification loss function during fine-tuning.

7. The method for predicting power system cascading failures based on dynamic data drive according to claim 1, characterized in that: While optimizing the classifier parameters, the encoder parameters are also fine-tuned.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the power system cascading failure prediction method based on dynamic data driving as described in any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting power system cascading failures based on dynamic data driving as claimed in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program, characterized in that: When the computer program is executed by a processor, the method for predicting power system cascading failures based on dynamic data driving as claimed in any one of claims 1 to 7 is implemented.

Images