A power system stability evaluation method and system based on space-time features

Abstract

The present application relates to the technical field of power grid, especially to a power system stability evaluation method and system based on space-time characteristics. First evaluation data of the power system is obtained, including bus voltage amplitude and phase angle data; the first evaluation data is input into a preset RNN-AGCN evaluation model; based on the output result of the RNN-AGCN evaluation model, the stability of the power system is evaluated. The first evaluation data includes pre-fault steady-state data, fault data and post-fault clearing data, which are collected by PMU. Before inputting the data, the data is normalized by min-max and processed to a specified dimension by 1x1 convolution. The RNN-AGCN evaluation model includes recurrent neural network RNN and adaptive graph convolution network AGCN, which are respectively used to extract the time sequence characteristics of transient data and the spatial structure characteristics of power grid topology. Through the dual extraction of time sequence and spatial characteristics, the accuracy and noise immunity of the evaluation are improved, and the model has strong topology adaptability, which provides a strong guarantee for the safe and stable operation of the power grid.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid, and particularly relates to a power system stability evaluation method and system based on space-time features. BACKGROUND

[0002] Transient stability assessment (TSA) of power systems is a key technology to ensure the safe and stable operation of power grids. Traditional TSA methods mainly include time-domain simulation method, direct method and energy function method, etc. Although the time-domain simulation method has high accuracy, it is time-consuming and difficult to meet the real-time evaluation requirements; the direct method and the energy function method are fast in calculation, but they are strongly dependent on system models and parameters, and the evaluation results are conservative. With the development of artificial intelligence technology, TSA methods based on deep learning have gradually become a research hotspot. Existing deep learning evaluation methods mainly use convolutional neural network (CNN), long short-term memory network (LSTM) and other algorithms to construct evaluation models, and realize the rapid evaluation of system stability by learning the features of transient process data of power systems.

[0003] However, the existing TSA methods based on deep learning still have the following problems: first, most of the methods only focus on the time sequence features of transient data, ignoring the inherent topological structure features of power systems, resulting in insufficient adaptability of the model to system structure changes; second, in the feature extraction process, a single neural network structure is often used, which is difficult to consider both the time sequence correlation and the spatial structure characteristics of the data, limiting the feature expression ability of the model; third, due to the large difference in the number of stable samples and unstable samples in the transient samples of power systems, and the fact that some samples are in the stable boundary region, the model is prone to evaluation deviation, especially in the presence of measurement noise, the evaluation accuracy is significantly reduced; finally, the existing evaluation methods generally lack effective preprocessing mechanisms for PMU collected data, and do not fully consider the dimensional differences of different features at different stages, affecting the effect of subsequent feature extraction. These problems seriously restrict the practical application effect of deep learning methods in the transient stability evaluation of power systems. SUMMARY

[0004] In view of the problems existing in the prior art, the present application is proposed.

[0005] Therefore, the problem to be solved by the present application is how to solve the problem that most methods only focus on the time sequence characteristics of transient data, ignore the inherent topological structure characteristics of the power system, and lead to insufficient adaptability of the model to the structural changes of the system; secondly, in the feature extraction process, a single neural network structure is often used, which is difficult to simultaneously consider the time sequence correlation and spatial structure characteristics of the data, limiting the feature expression ability of the model; thirdly, due to the large difference in the number of stable samples and unstable samples in the transient samples of the power system, and part of the samples are in the stable boundary region, which makes the model prone to evaluation deviation, especially in the presence of measurement noise, the evaluation accuracy is significantly reduced; finally, the existing evaluation methods generally lack effective preprocessing mechanism for PMU collected data, and do not fully consider the dimension difference of different features at different stages, which affects the effect of subsequent feature extraction.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a power system stability evaluation method based on space-time features, which comprises obtaining first evaluation data of the power system, the first evaluation data comprising bus voltage amplitude and phase angle data;

[0008] The first evaluation data is input into a preset RNN-AGCN evaluation model;

[0009] Based on the output result of the RNN-AGCN evaluation model, the stability of the power system is evaluated.

[0010] As a preferred scheme of the power system stability evaluation method based on space-time features, the first evaluation data further comprises pre-fault steady-state data, fault data and post-fault clearing data; the first evaluation data is collected by a preset phasor measurement unit PMU.

[0011] As a preferred scheme of the power system stability evaluation method based on space-time features, before the first evaluation data is input, the first evaluation data is subjected to min-max normalization processing; the normalized data is processed to a specified dimension by 1x1 convolution.

[0012] As a preferred scheme of the power system stability evaluation method based on space-time features, the RNN-AGCN evaluation model comprises a recurrent neural network RNN for extracting time sequence characteristics of transient data; an adaptive graph convolutional network AGCN for extracting spatial structure characteristics of the power grid topology; and the RNN and AGCN are interleaved to form a space-time feature extraction layer.

[0013] As a preferred embodiment of the power system stability assessment method based on spatiotemporal characteristics described in this invention, the adaptive graph convolutional network (AGCN) includes: establishing a graph network structure containing a set of nodes and a set of edges; automatically learning the adjacency matrix of the graph through an internal graph learning layer; and performing self-connection and symmetric normalization processing on the adjacency matrix.

[0014] As a preferred embodiment of the power system stability assessment method based on spatiotemporal features described in this invention, the method further includes: adding a skip connection to the output of each RNN, summing the skip connections at each layer sequentially; and using the residual connection between the output of the RNN and the output of the AGCN as the input to the next spatiotemporal feature extraction layer.

[0015] As a preferred embodiment of the power system stability assessment method based on spatiotemporal characteristics described in this invention, the power system stability assessment based on the output results of the RNN-AGCN assessment model includes: calculating the maximum power angle difference between any two generators; substituting the calculated maximum power angle difference into the transient stability index (TSI) formula; and determining the system stability based on the sign of the TSI.

[0016] Secondly, embodiments of the present invention provide a power system stability assessment system based on spatiotemporal characteristics, which includes a data acquisition module for acquiring first assessment data of the power system, wherein the first assessment data includes bus voltage amplitude and phase angle data;

[0017] The evaluation module inputs the first evaluation data into the preset RNN-AGCN evaluation model;

[0018] The results output module evaluates the stability of the power system based on the output results of the RNN-AGCN evaluation model.

[0019] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, they implement the steps of the power system stability assessment method based on spatiotemporal characteristics as described in the first aspect of the present invention.

[0020] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of the power system stability assessment method based on spatiotemporal characteristics as described in the first aspect of the present invention.

[0021] The application has the advantages that: the application realizes the dynamic characteristic capture of the whole process of the power system by acquiring complete time sequence characteristics including pre-fault steady-state data, fault data and post-fault data, and combining the PMU real-time acquisition technology. The problem of large dimension difference of different characteristic quantities is effectively solved by min-max normalization processing and 1x1 convolution dimension adjustment of the acquired data. The space-time feature extraction layer composed of RNN and AGCN is adopted to realize the double extraction of the transient data time sequence characteristics and the power grid topology spatial structure characteristics, and the limitation of only considering single feature dimension in the prior art is overcome. The adaptive ability of the model to the system topology change is improved by automatically learning the adjacency matrix in the internal graph learning layer of AGCN. The gradient disappearance problem in deep network training is effectively avoided by the skip connection and residual connection optimization design of the RNN output. Finally, the clear evaluation standard is established by calculating the maximum power angle difference of the generator and combining the TSI transient stability index for comprehensive judgment. The application realizes the full-process automation of the power system stability evaluation, significantly improves the accuracy and efficiency of the evaluation, and has strong anti-noise ability and topology adaptability, which provides a strong guarantee for the safe and stable operation of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Figure 1 Flowchart of the power system stability evaluation method based on space-time characteristics;

[0024] Figure 2 Computer device diagram of the power system stability evaluation method based on space-time characteristics;

[0025] Figure 3 Overall framework diagram of the power system transient stability evaluation of the power system stability evaluation method based on space-time characteristics;

[0026] Figure 4 Cyclic time network principle diagram of the power system stability evaluation method based on space-time characteristics;

[0027] Figure 5 RNN-AGCN-based power system transient stability evaluation model diagram of the power system stability evaluation method based on space-time characteristics;

[0028] Figure 6 Confusion matrix diagram of different loss functions of the power system stability evaluation method based on space-time characteristics;

[0029] Figure 7 FIG. 1 is a diagram of evaluation performance comparison of each model of the power system stability evaluation method based on spatio-temporal features under different noise levels;

[0030] Figure 8 FIG. 2 is a diagram of evaluation performance comparison of each model of the power system stability evaluation method based on spatio-temporal features under different topologies. DETAILED DESCRIPTION

[0031] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0032] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from the description, and those skilled in the art can make similar generalizations without departing from the concept of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0033] Secondly, the "one embodiment" or "embodiment" referred to herein means that a specific feature, structure or characteristic can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor does it mean an embodiment that is separate or selectively excluded from other embodiments.

[0034] Embodiment 1

[0035] Reference Figures 1-2 For the first embodiment of the present application, the embodiment provides a power system stability evaluation method based on spatio-temporal features, comprising,

[0036] S100: obtaining first evaluation data of the power system, the first evaluation data comprising bus voltage amplitude and phase angle data;

[0037] Specifically, the bus voltage amplitude data can be real-time voltage value, average voltage value, maximum voltage value and other data of the key nodes of the power system, the phase angle data can be bus phase angle, phase difference, power factor angle and other data, the power device junction temperature data can be temperature rise value, temperature threshold value, thermal model parameter and other data. The sampling time window data comprises steady-state data 2 seconds before the fault, data during the fault duration and data 3 seconds after the fault is cleared.

[0038] In an optional embodiment, the first evaluation data can be any combination of data obtained by various collection devices such as PMU devices, intelligent terminals, SCADA systems, etc. For example, when the bus voltage amplitude data is high-precision real-time data collected by the PMU device, the phase angle data can be phase angle data collected by the intelligent terminal.

[0039] In an optional embodiment, the first evaluation data can also increase or decrease other monitoring data according to different evaluation targets. For example, when a more comprehensive evaluation of system stability is required, the generator speed and power data (i.e., the third type of monitoring data) can be increased to capture the dynamic response characteristics of the unit, thereby further improving the accuracy and comprehensiveness of the evaluation. In addition, considering the diversity of power system operating states, an adaptive sampling algorithm can also be introduced to dynamically adjust the collected monitoring data, ensuring that high-quality system data can be obtained under various operating conditions.

[0040] In an optional embodiment, if the phase angle data is not required to be considered, the relevant monitoring data can be omitted to reduce the workload of data collection and processing. However, in the present application, such data is still retained to ensure the comprehensiveness of the evaluation.

[0041] It should be noted that the acquisition of the above monitoring data can comprehensively cover the key nodes of the power system, including but not limited to important busbars, generator units, load centers, etc., thereby ensuring the accuracy and comprehensiveness of the evaluation results. At the same time, by introducing various measurement devices and data collection systems such as PMU devices, intelligent terminals, SCADA systems, etc., the advantages of various types of data can be fully utilized, complementing each other to improve the efficiency and accuracy of the evaluation. In addition, for different evaluation needs, the types and quantities of data can be flexibly adjusted to meet the needs of actual evaluation work. This multi-source data fusion strategy not only improves the flexibility and adaptability of the evaluation system, but also provides rich data sources for subsequent intelligent analysis.

[0042] S101: The first evaluation data further includes pre-fault steady-state data, fault data, and post-fault clearing data; and the first evaluation data is collected by a preset phasor measurement unit (PMU).

[0043] S200: inputting the first evaluation data into a preset RNN-AGCN evaluation model;

[0044] Specifically, the RNN-AGCN evaluation model includes an input layer, a time series feature extraction layer, a spatial feature extraction layer, and an output layer. The input layer receives voltage amplitude and phase angle data after min-max normalization processing. The time series feature extraction layer is composed of multiple layers of RNN networks, which are used to capture system dynamic characteristics. The spatial feature extraction layer uses an AGCN network to analyze the topology of the power grid. The output layer gives the evaluation result of system stability.

[0045] In an optional embodiment, the network structure of the RNN-AGCN evaluation model can include any combination of design schemes with different depths and widths. For example, when the system scale is small, a simplified structure of 3-layer RNN and 2-layer AGCN can be adopted; when facing large-scale systems, a deep structure of 5-layer RNN and 4-layer AGCN can be expanded to extract more complex features.

[0046] In an optional embodiment, the RNN-AGCN evaluation model can also adjust network parameters according to different application scenarios. For example, in real-time evaluation scenarios, the number of network layers can be reduced and the calculation structure can be simplified (i.e. lightweight model) to improve the calculation speed; in offline analysis scenarios, a deeper network structure can be used to improve the expression ability of the model by increasing the number of hidden layers and neurons. In addition, considering the diversity of power system operation characteristics, an adaptive learning mechanism can also be introduced to dynamically adjust network parameters to ensure that the model maintains good performance under different operating conditions.

[0047] In an optional embodiment, if the evaluation task is simple, only RNN or AGCN single network structure can be used to simplify the model complexity. However, in this application, a complete RNN-AGCN dual network structure is used to fully integrate time and space features.

[0048] It should be noted that the design of the RNN-AGCN evaluation model needs to consider multiple factors, including but not limited to data features, calculation efficiency, evaluation accuracy, etc., to ensure the practicality and reliability of the model. At the same time, by reasonably configuring the network structure and parameters such as the number of hidden layers, the number of neurons, the type of activation function, etc., the feature extraction capability of the model can be fully utilized to improve the accuracy of the evaluation. In addition, for power systems of different scales, the network structure can also be flexibly adjusted to balance the relationship between calculation efficiency and evaluation accuracy. This adaptive network design strategy not only improves the generalization ability of the model, but also provides more optimization space for practical applications.

[0049] S201: Before inputting the first evaluation data, further comprising: performing min-max normalization processing on the first evaluation data; and processing the normalized data to a specified dimension through 1x1 convolution.

[0050] S300: Based on the output result of the RNN-AGCN evaluation model, the stability of the power system is evaluated.

[0051] Specifically, the stability evaluation process includes core steps such as feature extraction, data fusion, state judgment, etc. Among them, the feature extraction stage respectively captures the dynamic response characteristics of the system by the RNN network, and analyzes the topological structure characteristics by the AGCN network; the data fusion stage comprehensively processes the time sequence and spatial characteristics; the state judgment stage makes a final evaluation on the stable state of the system based on the transient stability index TSI.

[0052] In an optional embodiment, the stability evaluation result can include any combination of multiple output forms. For example, when a quick decision is needed, only the binary judgment result of stability / instability can be output; when detailed analysis is needed, quantitative indicators such as stability margin and critical fault time can be given at the same time, providing more reference for system operation optimization.

[0053] In an optional embodiment, the evaluation process can also add auxiliary functions according to different application requirements. For example, in the monitoring and early warning scene, the early warning level division function can be added to subdivide the system state into multiple levels such as safe, attention, warning, and danger; in the fault analysis scene, the fault feature recognition function can be added to assist the operation personnel to quickly locate the problem. In addition, considering the diversity of power system disturbances, an adaptive evaluation mechanism can also be introduced to dynamically adjust the evaluation strategy according to the disturbance type and degree.

[0054] In an optional embodiment, for simple evaluation tasks, only the TSI index can be used for judgment. However, in this application, a multi-index comprehensive evaluation scheme is adopted to provide more comprehensive evaluation information.

[0055] It should be noted that the generation of the stability evaluation result needs to consider multiple technical links, including but not limited to feature fusion method, criterion selection, threshold setting, etc., so as to ensure the reliability of the evaluation result. At the same time, by reasonably designing the evaluation process, such as the optimization configuration of the links of feature extraction, data fusion, state judgment, etc., the extracted feature information of the model can be fully utilized to improve the accuracy of the evaluation. In addition, for different application scenarios, the output form and evaluation strategy can also be flexibly adjusted to meet the actual engineering requirements. This comprehensive evaluation scheme not only improves the reliability of the evaluation result, but also provides strong support for system operation management.

[0056] S301: The RNN-AGCN evaluation model includes: a recurrent neural network RNN for extracting time sequence features of transient data; an adaptive graph convolution network AGCN for extracting spatial structure features of power grid topology; and the RNN and the AGCN are interleaved to form a spatio-temporal feature extraction layer.

[0057] S302: The adaptive graph convolutional network AGCN includes: establishing a graph network structure containing a node set and an edge set; automatically learning the adjacency matrix of the graph through an internal graph learning layer; and performing self-connection and symmetric normalization processing on the adjacency matrix.

[0058] S303: Further comprising: adding a skip connection to the output of each RNN, sequentially summing each layer of the skip connection; and connecting the output of the RNN and the output of the AGCN in a residual manner, and taking the result as the input of the next spatio-temporal feature extraction layer.

[0059] S304: The output result of the RNN-AGCN evaluation model is used to evaluate the stability of the power system, including: calculating the maximum power angle difference of any two generators; substituting the calculated maximum power angle difference into the transient stability index TSI formula; and judging the system stability according to the positive and negative of TSI.

[0060] Further, the embodiment also provides a power system stability evaluation system based on spatio-temporal features, comprising,

[0061] A data acquisition module acquires first evaluation data of the power system, and the first evaluation data includes bus voltage amplitude and phase angle data;

[0062] An evaluation module inputs the first evaluation data into a preset RNN-AGCN evaluation model;

[0063] A result output module evaluates the stability of the power system according to the output result of the RNN-AGCN evaluation model.

[0064] In summary, by collecting complete time series data including pre-fault steady-state data, fault data and post-fault data, and combining with real-time acquisition by the PMU phasor measurement unit, the dynamic characteristics of the entire process of the power system are captured, thereby providing a complete and reliable data basis for subsequent stability evaluation, and effectively avoiding the evaluation deviation caused by incomplete data acquisition in the traditional method.

[0065] By performing min-max normalization processing on the input data and using 1x1 convolution to process the data to a specified dimension, the problem of large dimension difference of different features in different stages is solved, the accuracy of subsequent feature extraction is improved, and the complexity of model training is reduced.

[0066] Through the innovative design of interweaving the RNN recurrent neural network and the AGCN adaptive graph convolutional network to form the spatio-temporal feature extraction layer, the transient data time series features and the power grid topology spatial structure features are extracted, the limitation of only considering a single feature dimension in the prior art is overcome, and the comprehensiveness of feature extraction is improved.

[0067] By establishing a graph network structure of node set and edge set in AGCN, and automatically learning the adjacency matrix through the internal graph learning layer, adaptive learning of the power grid topology structure is realized, and the adaptability of the model to the change of system topology and the generalization performance are improved.

[0068] By adding a skip connection to the RNN output and summing it up, and by optimizing the residual connection between the RNN output and the AGCN output, the gradient vanishing problem in deep network training is effectively avoided, and the training stability and convergence speed of the model are improved.

[0069] By calculating the maximum power angle difference of the generator and combining it with the TSI transient stability index for comprehensive judgment, a clear evaluation standard is established, and the interpretability and reliability of the evaluation result are improved.

[0070] The entire evaluation method realizes the full-process automation from data acquisition, feature extraction to stability evaluation, significantly improves the efficiency and accuracy of power system stability evaluation, and provides a strong guarantee for the safe and stable operation of the power grid.

[0071] Embodiment 2

[0072] Reference Figure 2 - Figure 8 As a second embodiment of the present application, the embodiment provides a power system stability evaluation method based on spatio-temporal features. In order to verify the beneficial effects of the present application, economic benefit calculation and simulation experiments are used for scientific demonstration.

[0073] The overall framework of power system transient stability evaluation includes two stages, namely offline training and online evaluation. In the offline training stage, the training set is used to train the RNN-AGCN model, and the validation set is used to find the best loss value in the model training process. The test set is completely unknown in the offline training process. In online evaluation, the test set comes from the measurement data of PMU. As shown in Figure 3 .

[0074] Recurrent neural network is a neural network structure specially used for processing sequence data, which can process time series data and predict future data trends. RNN can handle time dependencies in sequences, so it is very suitable for time series prediction.

[0075] The basic principle of RNN is to introduce the concept of time in neural networks, so that the network can process sequence data. The basic structure of RNN is a cycle unit, which includes an input layer, a hidden layer and an output layer. At each time step, the network receives an input vector and a hidden state vector, which are combined through a nonlinear function to produce an output vector and a new hidden state vector as the input and hidden state of the next time step. This feedback mechanism allows the network to remember previous information and consider historical information when processing sequence data. The principle of recurrent temporal network is shown in Figure 4 .

[0076] The RNN calculates the hidden state by the following formula:

[0077] h t =f(W·x t +U·h t-1 +b) (1)

[0078] Where W and U are the weight matrices of input to hidden and hidden to hidden, respectively, b is the bias term, and f is the activation function.

[0079] The transient process of power system can be divided into wave process, electromagnetic transient process and electromechanical transient process from the time sequence. In this process, the physical quantities of power system such as bus voltage amplitude and phase angle have obvious time-varying characteristics, which is essentially a time series. RNN and LSTM have great advantages in processing time series data, but RNN has no complex gating mechanism and loop structure inside, and compared with LSTM, the calculation cost of RNN is smaller and the training stability is better. Therefore, according to the time sequence characteristics of power system transient data, RNN is used to fully extract the data to ensure the evaluation performance of the model.

[0080] Adaptive graph convolution network:

[0081] In the power system, each bus is connected through the line, and the buses influence each other, which is essentially a graph network structure. If each bus is regarded as a node and the connecting line is regarded as an edge, the topological structure of the power system can be described as G=(V,E), where V represents the node set, E represents the edge set. The connection relationship between each node is represented by an adjacency matrix A∈R N×N , N represents the number of nodes, and the elements in the matrix A are valued according to the following principle:

[0082]

[0083] The graph convolution network (CN) takes the node feature matrix and the adjacency matrix as input, and outputs high-dimensional node feature representation after multiple convolution layers. The interlayer propagation can be represented as:

[0084] H(l+1) = f(H (l) ,A), l = 0, 1, 2,...., L (2)

[0085] In the formula, f(·) represents a nonlinear mapping function, L is the number of hidden layers, H (l) represents the node feature matrix of the output of the lth hidden layer. If l = 0, then H (0) = X.

[0086] For each layer of graph convolution layer, the process is to multiply the adjacency matrix A and the node feature matrix H (l) , and then perform linear transformation using the trainable parameter matrix W (l) , that is, the interlayer propagation can be written as:

[0087] H (l+1) = f(H (l) ,A) = σ(AH (l) W (l) ) (3)

[0088] In the formula, σ(·) represents an activation function. Since there is no self-connection of the node itself when constructing the adjacency matrix in the topological graph, the feature of the node itself cannot be noticed, which is not conducive to the expression of the node feature. Therefore, the node feature matrix is added to the N-order unit matrix, and the diagonal elements of the adjacency matrix are all 1, which is equivalent to adding self-connection to each node, that is:

[0089]

[0090] In the formula, is the adjacency matrix after adding self-connection, and I is an N-order unit matrix. It is noted that is not normalized, which will introduce numerical instability during interlayer propagation, so further symmetric normalization needs to be performed on , that is:

[0091]

[0092] In the formula, is the degree matrix of the adjacency matrix , which is a diagonal matrix, and the diagonal elements satisfy Thus, the interlayer propagation process of GCN can be described as:

[0093]

[0094] It is crucial to construct the adjacency matrix of the graph by using the GCN network to mine the spatial structure features of the power grid topology. For the power system, the degree of mutual influence between buses is different, and there may be mutual influence between two buses without actual connection relationship. If the value of formula (2) is taken (0, 1), the potential problem is that the mutual influence relationship between nodes cannot be fully represented. Therefore, the application adopts an adaptive graph convolution network (AGCN) to mine the spatial structure features of the power grid topology, and the adaptivity is reflected in that the internal graph learning layer automatically learns the adjacency matrix of the graph from the data, and the specific process is as follows:

[0095] M1=tanh(αE1Θ1) (7)

[0096] M2=tanh(αE2Θ2) (8)

[0097]

[0098]

[0099] In the formula, E1 and E2 are randomly initialized node embedding features, Θ1 and Θ2 are model trainable parameters, and α is a hyperparameter for controlling the saturation rate of the activation function. By calculating The similarity matrix between nodes can be obtained, and the adjacency matrix can be obtained after nonlinear operation of the activation function. The value of the element in the matrix is in (0, 1), and the value of the node pair with high mutual influence in the influence matrix is closer to 1. argtopm(·) represents a vector of the nearest m neighbor nodes to the center node, and the elements of the remaining positions are set to zero. The purpose is to sparsify the adjacency matrix, thereby reducing the computational complexity.

[0100] In addition, in order to prevent the AGCN network from converging to a single point during interlayer propagation and causing information loss, it is considered to retain part of the original state of the nodes during propagation, and the propagation process can be described as:

[0101]

[0102] In the formula, H in represents the original node feature matrix input to the AGCN network, represents the adjacency matrix after self-connection and symmetric normalization processing of A adp , and β is the feature retention rate, which is used as a hyperparameter to control the retention degree of the original node features.

[0103] TSA model based on RNN-AGCN:

[0104] The construction of input features needs to be able to reflect the transient process of the power system and have obvious time-varying characteristics. The present application selects all bus voltage amplitudes and phase angles as initial input features, and the sampling window covers the pre-fault steady state, fault and post-fault clearing. Due to the large dimension difference of different features at different stages, the min-max normalization needs to be performed before inputting into the neural network. The power system transient stability evaluation model based on RNN-AGCN is as shown in Figure 5

[0105] The transient data after min-max normalization is first processed by 1x1 convolution to the specified dimension, and then input into the RNN-AGCN module. The RNN and AGCN are interleaved to constitute a space-time feature extraction layer, which extracts the time sequence features and spatial structure features of the transient data respectively. The graph learning layer calculates the similarity between nodes through randomly initialized node embedding features, and adaptively generates an adjacency matrix with sparsity. After self-connection and symmetric normalization processing, it is combined with the output of the RNN and input into the AGCN for graph convolution operation to realize the extraction of space-time features of the transient data. In addition, in order to prevent the problem of gradient disappearance, the input of the RNN and the output of the AGCN are connected in residual connection and then used as the input of the next space-time feature extraction layer. At the same time, a skip connection is added to the output of each RNN, and the essence is 1xL i convolution, L i represents the input sequence length of the i-th skip connection, and the sum is taken for each layer of skip connection in turn until the last layer of skip connection is finished. Then, the space-time features extracted by the model are subjected to nonlinear operation and full connection operation, and then combined with the transient stability index (TSI) to obtain the stability / instability evaluation result. The expression of TSI is as follows:

[0106]

[0107] Δδ max is the maximum power angle difference of any two generators. For each sample, if η TSI is positive, the system is stable, and the sample is labeled as 1, otherwise the sample is unstable and labeled as 0. The essence is to judge the stability of the system by calculating whether the maximum power angle difference of any two generators in the system is greater than 180°.

[0108] Focal loss function:

[0109] FL is improved on the basis of the weighted cross-entropy loss function. The expression of the weighted cross-entropy loss function can be described as:

[0110]

[0111] ​where y is the true label of the sample, 1 represents stable, and 0 represents unstable; p i is the probability of the model predicting the class; w us , w s respectively represent the weight factors of unstable samples and stable samples, and the values thereof are usually determined according to the proportion of unstable samples and stable samples in the sample set. In addition to the sample class imbalance problem in the actual power system, there are also some samples at the stable boundary, and the classification of such samples is difficult, so when designing the loss function, the loss proportion of the samples should be increased, and the loss proportion of the easy-to-classify samples should be reduced. FL introduces a difficult-easy sample modulation factor on the basis of the weighted cross-entropy to improve the easy-to-mistake problem of the difficult-to-classify samples at the stable boundary, and the expression of FL is as follows:

[0112]

[0113] where γ is the difficult-easy sample modulation factor; when p i is closer to 0.5, it means that the sample classification difficulty is greater, and when p i is closer to 0 or 1, it means that the sample classification difficulty is smaller. By adjusting the value of γ, the loss proportion of the difficult-to-classify samples in the model training process is increased, and the loss proportion of the easy-to-classify samples is reduced.

[0114] Evaluation index:

[0115] The transient stability assessment is essentially a binary classification problem, and the transient samples are in a class imbalance state. If only the global evaluation accuracy is used as the evaluation index of the model, the performance of the model cannot be comprehensively reflected. In addition, considering that the cost of judging the unstable sample as stable is greater than that of judging the stable sample as unstable, the evaluation index should pay more attention to the evaluation accuracy of the unstable sample. Therefore, the confusion matrix in the machine learning classification task is used to comprehensively evaluate the evaluation performance of the model by using the accuracy (A cc ), recall (R ec ), precision (P re ), and F1 score, and the confusion matrix is shown in Table 1.

[0116] Table 1 TSA confusion matrix

[0117]

[0118] where T s , T us respectively represent the number of stable samples and unstable samples that are correctly evaluated, F s represents the number of unstable samples that are evaluated as stable samples (missed judgment), and F usThe mathematical definition of each evaluation index indicating that a stable sample is evaluated as an unstable sample (misjudgment) is shown in formula (16) ((1)).

[0119]

[0120] In the formula, A cc characterizes the proportion of samples correctly evaluated to the total samples, reflecting the global accuracy of model evaluation; R ec characterizes the proportion of unstable samples correctly evaluated to the actual unstable samples, R ec is higher, indicating that the number of unstable samples misjudged as stable samples is less; P re characterizes the proportion of unstable samples correctly evaluated to the unstable samples evaluated, F1 is R ec and P re is the harmonic mean of R

[0121] Example analysis

[0122] To verify the effectiveness of the method, the PSD-BPA simulation software is used to generate a sample set, and the RNN-AGCN evaluation model is built based on the pytorch framework under python3.8. Since the hyperparameters of the deep neural network have a great influence on the performance of the model, and there is no perfect theory to guide the selection of hyperparameters at present, the network layer number of the spatio-temporal feature extraction layer is set to 3 after multiple experimental tests; dropout is added after each hidden layer of RNN to prevent overfitting, the value of dropout is set to 0.3, and the expansion factor is set to 2; the depth of AGCN network is set to 3, the number of neighbor nodes is set to 8, the saturation rate of the activation function is α = 3, and the feature retention rate is β = 0.05; the unstable sample weight factor w us = 0.75, the stable sample weight factor w s = 0.25, and the difficulty sample modulation factor γ = 2; the training batch size is set to 128, the training round is set to 120, the initial learning rate is set to 0.001, and the optimization algorithm is Adam algorithm.

[0123] Loss function effect comparison

[0124] Binary cross-entropy and weighted cross-entropy loss functions are commonly used to guide deep learning evaluation model training. The two and the focal loss function are used to guide model training, and the performance indicators of the obtained training model on the test set are shown in Table 2 and Figure 7 .

[0125] Table 2 Test results of different loss functions

[0126] Loss function A cc / %]]> [R ec / %]]> P re / %]]> [F1 / %] Binary cross-entropy 98.85 97.08 97.79 97.44 Weighted cross-entropy 99.26 98.18 98.53 98.35 Focal loss 99.51 98.54 99.26 98.90

[0127] Table 2 and Figure 7 It can be seen that the TSA model trained by the focal loss function has the best performance, and the number of missed samples and misjudged samples is the least. This is because the sample category weight factor is introduced in the focal loss function, which reduces the frequency of stable samples participating in model parameter adjustment, improves the problem that the model tends to evaluate unknown samples as stable samples, and at the same time, the existence of the difficulty sample modulation factor makes the model pay more attention to the difficult-to-classify samples on the stable boundary during the training process. In the following analysis, the RNN-AGCN model refers to the model trained by the focal loss function.

[0128] 2.2 Model evaluation performance comparison

[0129] To verify the effectiveness of the model proposed in the application, the same sample set is used to construct DT, SVM, CNN, LSTM, TCN, AGCN and other deep learning evaluation models for comparison. Among them, the C5.0 algorithm is selected for DT, and the default value 0.25 is used for the confidence factor; the radial basis function is used for SVM, and the grid search method and five-fold cross-validation are used to find the optimal value of the related parameters; the CNN is set to have 2 convolutional layers and 2 fully connected layers, and the convolution kernel size is set to 3; the number of hidden layer neurons of the LSTM is {64, 128, 32}, and a fully connected layer is set. The evaluation performance of each model on the same test set is shown in Table 3.

[0130] Table 3 Test results of different evaluation models

[0131] Model A cc / %]]> [R ec / %]]> P re / %]]> [F1 / %] DT 94.92 86.13 90.77 88.39 SVM 95.41 87.59 91.60 89.55 CNN 96.72 90.51 94.66 92.54 LSTM 98.03 94.89 96.30 95.59 TCN 98.28 95.62 96.68 96.15 AGCN 98.93 97.08 98.15 97.61 RNN-AGCN 99.51 98.54 99.26 98.90

[0132] As can be seen from Table 3, the shallow neural networks such as DT and SVM have limited understanding of the underlying data, with the worst accuracy and F1 scores that do not reach 90%. The feature extraction capability of deep learning algorithms such as CNN and LSTM is better than that of shallow neural networks, so the evaluation performance of the corresponding models is also higher. The RNN-AGCN evaluation model proposed in the application fully integrates the time sequence features of transient data and the spatial structure features of power grid topology, and its evaluation indicators are better than those of RNN and AGCN, with an accuracy of 99.51% and an F1 score of 98.9%. Especially, the recall rate reaches 98.54%, which means that the number of unstable samples missed as stable samples by the proposed model is the least, ensuring the evaluation accuracy of unstable samples.

[0133] 2.3 Evaluation result analysis under the condition of PMU measurement data with noise

[0134] In actual power systems, the data collected by the phasor measurement unit (PMU) may contain measurement errors. According to the IEEE power system synchronization data transmission standard

[25] , the PMU measurement error should be less than 1%. The conversion formula between measurement error and signal-to-noise ratio is as follows:

[0135]

[0136] P s P represents the effective power of the signal. n Let η be the effective noise power and η be the signal-to-noise ratio (SNR). A lower SNR indicates more severe noise interference. This invention considers measurement errors of 0.2%, 0.4%, 0.6%, and 0.8%, corresponding to noise SNRs of 26.99, 23.98, 22.22, and 20.97 dB, respectively. The test set with added noise was used to test the evaluation performance of each model. The test results are as follows: Figure 7 As shown.

[0137] from Figure 7 As can be seen, the evaluation performance of all models decreased after noise was introduced, and the more severe the noise, the greater the decrease in accuracy and F1 score. The RNN-AGCN model proposed in this invention maintains an accuracy of 98.17% at a noise level of 20.97 dB, a decrease of only 1.34% compared to the noise-free case, and the F1 score still reaches 96%, a decrease of 2.89% compared to the noise-free case. Compared with the other six comparison models, the proposed evaluation model exhibits the best performance at all noise levels, indicating that the proposed model has strong robustness to noise.

[0138] 2.5 Generalization ability test of the model under the new topology

[0139] In power system operation, line maintenance or faults often cause the system to deviate from its full topology. If a model possesses strong generalization ability to these topology changes—meaning it maintains good evaluation performance even after the power system topology changes—the model update process can be avoided, thus reducing computational costs. Therefore, this section will compare the generalization abilities of different models under new topologies to verify the performance advantages of the proposed evaluation model after system topology changes.

[0140] New topology considers randomly breaking 1—strip transmission line in IEEE 39 node system, forms N-1, N-2 topology, and ensures that the number of nodes of the system does not change. Because the operation mode combination of N-1, N-2 is too much, the cost of complete traversal is too large, therefore, the present application randomly selects 40 kinds of topology combinations and sets the fault to carry out time domain simulation. The fault setting condition refers to section 2.1, and a total of 4080 simulation samples are generated, of which the number of stable samples is 2448, and the number of unstable samples is 1632. The sample set is used as a test set to test the evaluation performance of each model, and the test results are as shown in Figure 8

[0141] It can be seen that the RNN-AGCN model proposed in the present application has the best evaluation performance under the new topology, and the accuracy and F1 score are 98.43% and 98.03% respectively, which are better than the other 6 groups of comparison models, which shows that the proposed model has good generalization ability to the topology change of the system. The performance of the AGCN model under the new topology is only slightly worse than the RNN-AGCN, which is better than the DT, SVM, CNN, LSTM and TCN models, because the AGCN extracts the spatial structure features of the power grid topology, and when the system topology changes, it will be reflected in the adjacency matrix input to the model, so its performance is less affected by the topology change. RNN-AGCN further considers the time sequence features of the transient data on the basis of AGCN, so the performance on the original topology and the new topology test set is better than AGCN. Other conventional deep learning and shallow neural network models cannot reveal the change of the system topology structure, so the performance on the new topology test set is not good.

[0142] 2.6 Model time consumption analysis

[0143] In order to reflect the time efficiency of the RNN-AGCN evaluation model, the training time consumption and online evaluation time consumption of the conventional deep learning model (CNN and LSTM) are compared and analyzed. Among them, the offline training time consumption is the sum of the training set training model and the validation set finding the best loss value. The test results are shown in Table 4.

[0144] Table 4 Comparison of time efficiency of different models

[0145] Model Training time consumption / s Evaluation time consumption / s Single sample evaluation time consumption / ms CNN 97.32 0.035 0.029 LSTM 100.26 0.040 0.033 RNN-AGCN 1462.47 0.390 0.320

[0146] ​It can be seen that the offline training time of the RNN-AGCN is much longer than that of the CNN and LSTM models, which is due to the feature representation update through the connection between nodes in the RNN-AGCN model, which increases the amount of calculation in the information transmission process, thereby causing the increase of the model training time. Notably, the online evaluation of the RNN-AGCN model on 1220 samples only takes 0.39s, and the average evaluation time of a single sample is only 0.32ms, which is slightly higher than that of the conventional deep learning model, but is negligible compared with the time-domain simulation method. Therefore, the RNN-AGCN model of the present application is still a TSA model that can meet the online evaluation requirements.

[0147] Embodiment 3

[0148] The embodiment also provides a computer device suitable for the case of the power system stability evaluation method based on space-time features, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the power distribution network forced oscillation detection and positioning method proposed in the above embodiment.

[0149] The embodiment also provides a storage medium having a computer program stored thereon, and the program is executed by a processor to realize the power distribution network forced oscillation detection and positioning method proposed in the above embodiment.

[0150] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0151] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part 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 to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0153] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0154] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example, by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following technology, known in the art, or combinations thereof, can be used: discrete logic circuitry having logic gates for implementing logic functions upon an application of data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

[0155] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A method for power system stability assessment based on spatio-temporal features, characterized in that: The method comprises the following steps: acquiring first evaluation data of a power system, wherein the first evaluation data comprises bus voltage amplitude and phase angle data; inputting the first evaluation data into a preset RNN-AGCN evaluation model; the RNN-AGCN evaluation model comprises a recurrent neural network (RNN) for extracting time sequence features of transient data and an adaptive graph convolutional network (AGCN) for extracting spatial structure features of a power grid topology; the RNN and the AGCN are interleaved to form a space-time feature extraction layer; the AGCN comprises the following steps: establishing a graph network structure comprising a node set and an edge set; automatically learning an adjacency matrix of the graph through an internal graph learning layer; and performing self-connection and symmetric normalization processing on the adjacency matrix; based on an output result of the RNN-AGCN evaluation model, the stability of the power system is evaluated.

2. The method for power system stability assessment based on spatio-temporal features as claimed in claim 1, wherein: The first evaluation data further comprises pre-fault steady-state data, fault data and post-fault clearing data; the first evaluation data is collected by a preset phasor measurement unit (PMU).

3. The method for power system stability assessment based on spatio-temporal features as claimed in claim 2, wherein: Before the first evaluation data is inputted, the first evaluation data is subjected to min-max normalization processing; and the normalized data is processed to a specified dimension through 1x1 convolution.

4. The method for power system stability assessment based on spatio-temporal features as claimed in claim 3, wherein: The method further comprises the following steps: a skip connection is added to the output of each RNN, and each layer of the skip connection is sequentially summed; the output of the RNN and the output of the AGCN are connected in residual, and the result is used as the input of the next space-time feature extraction layer.

5. The method for power system stability assessment based on spatio-temporal features as claimed in claim 4, wherein: The step of evaluating the stability of the power system based on the output result of the RNN-AGCN evaluation model comprises the following steps: calculating the maximum power angle difference between any two generators; substituting the calculated maximum power angle difference into a transient stability index (TSI) formula; and judging the system stability according to the positive and negative of the TSI.

6. A spatiotemporal feature based power system stability assessment system based on the spatiotemporal feature based power system stability assessment method of any one of claims 1-5, characterized in that: The method further comprises the following steps: an evaluation module for inputting the first evaluation data into a preset RNN-AGCN evaluation model; a result output module for evaluating the stability of the power system based on an output result of the RNN-AGCN evaluation model. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the power system stability evaluation method based on space-time features according to any one of claims 1-5.

8. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the power system stability evaluation method based on space-time features according to any one of claims 1-5.

Images