Data driving method suitable for multi-task transient stability evaluation of power system
Through the maximum Lyapunov index and dynamic heterogeneous graph model, the bottleneck problem of criterion selection and evaluation algorithm in multi-task transient stability evaluation of power system is solved, and more accurate transient stable state evaluation and model performance improvement are achieved, and the The application capability of online stable evaluation.
Patent Information
- Application Number
- CN202411002458.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-08
AI Technical Summary
The existing data-driven methods have bottlenecks in the selection and evaluation algorithm methods of criterion selection and evaluation in the multi-tasking transient stability evaluation of power systems, and cannot effectively consider the spatial characteristics and node heterogeneity of the grid topology, resulting in limited applications of complex power systems.
Using the maximum Lyapunov index (MLE) as the stable criterion, combined with the dynamic heterogeneous graph model, the system's spatiotemporal dynamic characteristics are extracted through the heterogeneous graph transformation architecture and the time attention-gated recursive unit to build a multi-tasking temporary installation suitable for power systems. State stability evaluation method, considering node heterogeneity and topological changes.
A more accurate transient stable state evaluation is achieved, which improves the performance indicators of the model, especially to achieve optimal indicators, and provides interpretability of decision rules through the Shapley method, enhancing the application capabilities of online stable evaluation.
Smart Images

Figure CN118917194A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of transient stability assessment of power systems, and in particular to a data-driven method suitable for multi-task transient stability assessment of power systems, which adopts a maximum Lyapunov exponent criterion-dynamic heterogeneous graph model analysis framework. Background Art
[0002] Transient stability state assessment is an important basis for taking effective control measures in time after the power grid is disturbed. With the increase in the proportion of new energy grid access and the proportion of power electronic converter access, the new power system with stronger new energy absorption capacity has heterogeneous components, complex morphological structures, nonlinear control processes, and diversified coupling forms. The characteristics of the complex system of the power grid are highlighted (Ju Ping, Jiang Tingyu, Huang Hua. On the "three self" properties of the new power system [J]. Proceedings of the CSEE, 2023, 43(07): 2598-2608), which puts forward higher requirements for rapid and accurate transient stability assessment. The combination of complexity science theory and data-driven thinking provides new ideas for the stability analysis and control of large power grid systems.
[0003] Most of the existing data-driven methods use the transient stability index as the power angle criterion and the voltage criterion based on the voltage measurement threshold information of a single binary meter. The wide access of wide area measurement systems (WAMS) and synchronized phasor measurement units (PMUs) provides data support for obtaining various response trajectory information about system dynamics, which can provide a richer basis for transient stability judgment. The literature (Zhu Cunhao, Ma Shiying, Zheng Chao, et al. Identification of transient instability of power systems based on measured response trajectories [J]. Proceedings of the CSEE, 2021, 41(15): 5077-5091) points out that during the disturbance process, the phase trajectory characteristic analysis of different stable states shows different geometric concave-convex characteristics. The literature (Wei Shaopan, Yang Ming, Han Xueshan, et al. Online identification of transient power angle stability based on phase trajectory MLE index [J]. Automation of Electric Power Systems, 2017, 41(16): 71-79) proposed a transient power angle stability identification method based on the maximum Lyapunov index of the speed difference-power angle difference phase trajectory. The above research can characterize the dynamic characteristics of complex systems to a certain extent, but since new energy units do not have traditional rotor angle characteristics, it is difficult to analyze the transient stability characteristics of the system from the power angle characteristics. There are many forms of instability in the transient disturbance process. At present, the analysis of transient power angle stability problems is mostly based on the power supply side, and transient voltage stability problems are mostly based on the load side, but in essence, both are caused by power imbalance in the transient process (Li Zonghan, Sun Huadong, Zheng Chao, et al. Identification of the dominant instability form based on the reactive power distribution characteristics of wide-area branch response [J]. Proceedings of the CSEE, 2023, 43(13): 4883-4898). Considering that the selection of input features should follow the two principles of criticality and easy accessibility, important measurement information characterizing the power angle and voltage stability can be screened from the power side and the load side respectively.
[0004] On the other hand, various data-driven models have been widely used in system transient stability assessment tasks, such as random forests (Wu Yuxiang, Han Xiaoqing, Niu Zhewen, et al. Transient stability assessment method based on variable weight random forest and its interpretability analysis [J]. Automation of Electric Power Systems, 2023, 47(14): 93-104), long short-term memory networks (Jie Zhijun, Zhang Dongxia, Han Xiaoqing, et al. Research on power system transient stability assessment method based on improved long short-term memory network [J]. Power System Technology, 2024, 48(03): 998-1010), ensemble learning models, etc. (Zhou Ting, Yang Jun, Zhan Xiangpeng, et al. A data-driven transient voltage stability assessment method and its interpretability research [J]. Power System Technology, 2021, 45(11): 4416-4425). The above models can effectively extract the time series characteristics of electrical quantities and establish a mapping relationship between input characteristics and stability, but do not consider the impact of the spatial characteristics of the power grid topology on the system stability. The power system is a network formed by interconnected generators, loads and control elements, and its time-varying topology has a great impact on the stability of the system.
[0005] The graph neural network based on the message passing mechanism is adapted to the structural characteristics of the power system graph. The aggregation relationship of nodes can be constrained by the network topology, so the spatial characteristics of the system can be extracted with the help of the graph neural network. Reference (Jiang Tao, Dong Yu, Wang Changjiang, et al. Transient voltage stability assessment of receiving power system based on graph convolution and bidirectional long short-term memory network [J]. Power System Technology, 2023, 47(12): 4937-4951) A transient voltage assessment model based on the combination of graph convolution neural network and long short-term memory network mines the spatiotemporal distribution law of the power grid; Reference (Zhang Liang, An Jun, Zhou Yibo. Transient stability assessment of power system based on time convolution and graph attention network [J]. Automation of Electric Power Systems, 2023, 47(7): 114-122) A graph attention network was built to extract the spatial characteristics of transient data. The graph model considered in the above study belongs to the category of homogeneous graphs with single node and edge types. The input and output feature dimensions of each node should be consistent. It has not yet been considered that various types of graph data may be naturally heterogeneous. Different types of nodes have different attributes and should be regarded as heterogeneous nodes when forming graph topology. Taking different types of bus nodes in the power system as an example, in addition to the common bus node voltage amplitude and phase angle characteristics, the synchronous generator node also has unit-level feature quantities such as power angle and generator terminal voltage. The wind turbine node has equipment-level variables such as grid-side and rotor-side converters that characterize the dynamic characteristics of the wind turbine itself. At this time, the power grid topology belongs to the category of heterogeneous graphs. In addition, the power grid topology will continue to change on the time section due to different fault clearing modes, and it cannot be accurately represented using a simple static isomorphic graph.
[0006] In summary, the current research on multi-task transient stability assessment methods based on data-driven methods has bottlenecks in the selection of criteria and the applicability of assessment algorithms, which limits its application in complex power systems. Summary of the invention
[0007] In order to solve the problems existing in the criterion selection and evaluation algorithm method of the transient stability assessment method of the power system based on the data-driven method, the present invention constructs a dynamic heterogeneous graph model of multi-task transient stability assessment with the maximum Lyapunov exponent (MLE) as the stability criterion, and provides a data-driven method suitable for multi-task transient stability assessment of the power system from the perspective of the combination of complex systems and data-driven.
[0008] The present invention (1) in terms of mechanism criterion selection, considering that different types of voltage measurement information can reflect different stability mechanisms, the dynamic characteristics of the generator terminal voltage amplitude can reflect the change of power, and then indirectly reflect the stability of the system power angle, starting from the terminal voltage and the load bus voltage, respectively, with the help of chaotic system stability theory, the transient power angle and voltage stability problems are incorporated into a unified analysis framework to complete the complex characteristics characterization and mechanism analysis of the new power system. The biggest benefit of this research idea is that the voltage can be directly obtained by the synchronous phasor measurement device (PMU), and there is no need to determine the power angle reference system between generator sets; (2) in terms of transient stability assessment algorithm, due to the non-enumerability of the operating mode under the transient stability assessment problem of the power system, the current various data-driven methods are limited by the way the model itself processes input data, and generally do not take into account the differences in the properties of the node type itself and the changes in the power grid topology, resulting in the model being unable to better extract the spatiotemporal dynamic characteristics of the system. The dynamic heterogeneous graph model has certain advantages in processing non-Euclidean graph data of power grids with both node heterogeneity and topological dynamics.
[0009] The present invention is implemented by the following technical solution: a data-driven method suitable for multi-task transient stability assessment of power systems, comprising the following steps:
[0010] Phase 1: First, verify the feasibility of the “voltage response information-phase trajectory concave-convex characteristics-maximum Lyapunov exponent” criterion framework, and use the maximum Lyapunov exponents of the terminal voltage and load bus voltage trajectory as the power angle and voltage transient stability criteria. The details are as follows:
[0011] S1-1: Phase trajectory physical characteristics analysis:
[0012] The v'-Δv phase trajectory is plotted with the generator terminal voltage and the load bus voltage deviation as the horizontal coordinates and the voltage amplitude change rate as the vertical coordinate, as shown in formula (1):
[0013]
[0014] Among them, v kΔt and v (k-1)Δt are the voltage amplitude measurements at time kΔt and time (k-1)Δt, Δv kΔt is the deviation between the voltage and the reference voltage at time kΔt, v' kΔt is the voltage amplitude change rate at the current moment relative to the previous moment, f s is the sampling frequency;
[0015] This step analyzes the phase plane composed of different transient voltage response information and observes the physical characteristics of the phase trajectory in the two-dimensional plane. Based on this characteristic, the system is analyzed in detail during the transient disturbance process. The stable and unstable scenarios show different divergent or contraction trends on the phase plane, so as to characterize different transient stability types and states.
[0016] S1-2: Construction of the maximum Lyapunov exponent criterion:
[0017] The maximum Lyapunov index is an important indicator for measuring the dynamic characteristics of a nonlinear chaotic system, and it characterizes the average exponential rate of convergence or divergence between adjacent orbits of the system. If the value of the maximum Lyapunov index is greater than zero, it is considered that the distance between adjacent orbits in the phase space of the system increases exponentially with time, and the system presents a chaotic phenomenon. If it is less than zero, it is considered that the system tends to be stable. The calculation method is given by formula (2):
[0018]
[0019] The subscript values represent the sampling moments corresponding to the voltage values.
[0020] Phase II: Considering three types of nodes, namely synchronous generator nodes, doubly-fed wind turbine nodes and busbar nodes, the time series electrical quantities and topological information of multi-time sections are selected as input features. The spatiotemporal dynamic characteristics of the system are extracted through the heterogeneous graph transformation architecture HGT and the time attention gated recursive unit TGRU. The constructed dynamic heterogeneous graph can effectively take into account the heterogeneity and topological changes of the nodes; the details are as follows:
[0021] S2-1: Model construction:
[0022] The dynamic heterogeneous graph model includes two components: heterogeneous graph conversion architecture and temporal attention gated recursive unit. The heterogeneous graph conversion architecture takes into account the differences in the input feature dimensions of nodes in the power grid and retains important node information. The core idea is to update the target node information by aggregating the neighbor node information of the target node. It is divided into three modules: heterogeneous mutual attention, heterogeneous message passing and aggregation. The temporal attention gated recursive unit captures the evolution pattern of the topology through the changes in the node embedding of multiple time sections, and introduces the temporal attention mechanism to learn the embedding representation of different nodes at different times.
[0023] S2-1-1: Heterogeneous graph conversion architecture: Given a graph G(V,E,O,R), V and E are the node and edge sets respectively, O and R are the node and edge types respectively, and there are type mapping functions: τ(v):V→O,φ(e):E→R; the attention weight operation for the target node n and its neighbor node m∈M(t) is given by formula (3):
[0024]
[0025] in, Nodes m and n are linearly mapped and The vector of the i-th attention head obtained; is the edge type maintenance matrix, μ is the attention scaling factor; M(t) is the set of neighbor nodes; h is the number of attention heads; X (l-1) [*] is the output vector of the l-1th layer. If l=1, then X is the original input feature X0 of the model;
[0026] The message passing operation of multi-head attention is given by formula (4):
[0027]
[0028] in, is the vector of the i-th attention head obtained after linear mapping of the source node m, The maintenance matrix for edge types;
[0029] Aggregation operation aggregates the information of all neighbor nodes to the target node based on the attention weight. Finally, the target node response of the lth layer is mapped to the distribution of the corresponding node category, and the residual connection is introduced to obtain the final representation of the target node n as shown in formula (5):
[0030]
[0031] in, For direct and operation, A-Linear τ(n) is a linear mapping operation, δ is an activation function;
[0032] S2-1-2: Temporal Attention Gated Recurrent Unit: The structure is a cascade structure of multiple layers of gated recurrent units and a temporal attention mechanism layer. The embedded representation of each node at time t is obtained by the heterogeneous graph conversion architecture. The iterative formula of the gated recursive unit at time t is shown in formula (6):
[0033]
[0034] Among them, z t ,r tare the update gate and forget gate at time t respectively; W z ,W r ,W s ,b z ,b r ,b s is a trainable parameter; σ is an activation function; x t is the input feature at time t, obtained from the output of the heterogeneous graph conversion architecture; s t and t-1 are the outputs of the hidden layer at time t and the previous time respectively;
[0035] The hidden layer output of the gated recurrent unit at node i at different times As input, the temporal attention performs a scaled dot product attention mechanism, and its output Given by formula (7):
[0036]
[0037] in is a trainable parameter; is the mask matrix;
[0038] After the temporal attention mechanism is used in this step, different node embedding representations at each moment can be obtained, and the node information at the final moment is used as the input of the downstream transient stability assessment task.
[0039] S2-2: Model performance evaluation indicator construction:
[0040] The transient power angle and voltage multi-task evaluation to be solved by the present invention is a multi-label binary classification problem. A 4×4 multidimensional confusion matrix is constructed to visualize the model classification effect, where each column represents the predicted category of the sample and each row represents the true category of the sample; each sample label is described by a 1×2 matrix, and the first element of the matrix represents the stable state of the power angle, where the stable label is δ s , the instability label is δ us ; The second element represents the voltage stable state, the stable label is V s , the instability label is V us , so there are four types of probabilistic label combinations, namely [δ s ,V s ]、[δ s ,V us ]、[δ us ,V s ]、[δ us ,V us]; Each element value in the confusion matrix is the count value of different combinations of true labels and predicted labels, corresponding to the above four situations. The sample prediction category is consistent with the true category, and the sample prediction category is inconsistent with the true category, and is represented by F. The subscript values represent the number of rows and columns, corresponding to different stable scenarios. Furthermore, constructing corresponding evaluation indicators based on the confusion matrix is a key step to ensure the effectiveness of the proposed model. The evaluation indicators are mainly performance indicators extended based on the confusion matrix. Different evaluation indicators can show the performance of different aspects of the model.
[0041] Phase 3: Based on the Shapley method, the interpretability of the proposed algorithm is studied, the decision rules of the model are quantified, and the attribution and source analysis of the prediction results are completed to provide a basis for transient preventive measures in the online application phase. The details are as follows:
[0042] S3-1: Transient stability assessment process:
[0043] S3-1-1: Sample generation and stability annotation: The time-domain simulation operation scenario generates an initial sample set by taking into account factors such as different power flow operation modes, different fault information and clearing methods. Different input electrical quantities that cover all transient characteristics as much as possible are selected for each type of node, and the transient response trajectory is further extracted. The transient power angle and voltage stability state of the sample set are annotated based on the maximum Lyapunov index criterion;
[0044] S3-1-2: Offline training: The sample set is randomly divided into a training set and a test set according to a ratio of 7:3. The training set is used to train the model. During the training process, the model parameters are updated by backpropagation with the goal of minimizing the loss function until the loss function value meets the threshold condition. The model performance is evaluated on the test set based on the trained model. The prediction contribution of each feature in a specific scenario is quantitatively analyzed based on the Shapley method, and the extracted decision rules are further demonstrated in combination with electrical field knowledge.
[0045] S3-1-3: Online application: In this stage, the PMU is used to monitor the grid measurement information in real time, and the information is input into the trained model to quickly give the grid transient stability status judgment results under different operation scenarios, so that the control personnel can take preventive control measures. At the same time, the confidence level can be set and human assistance can be intervened to ensure accurate assessment of the system stability situation;
[0046] S3-2: Single sample rule extraction for model prediction mechanism;
[0047] At present, the focus of data-driven method research is still on the improvement of model algorithms and the applicability of models in different operating scenarios, ignoring the closed-loop analysis of the "data-model-data" analysis framework, that is, returning from the model prediction results to the input data level. The complexity and multidimensionality of the model algorithm itself bring difficulties to the interpretability of the decision-making process. The extraction of decision rules in transient assessment tasks can not only enhance the credibility of the results, but also provide a basis for decision-making in the real-time assessment stage. To a certain extent, it can overcome the limitations of the current data-driven method in practical application and improve the application ability of online stability assessment. The Shapley method is used to quantify the contribution of input features to the model output. The model's prediction value for one of the samples is equal to the sum of the baseline prediction value and the characteristic Shapley value contained in the sample. Its calculation method is shown in formula (8):
[0048]
[0049] Among them, N is the total number of input features; F is the model's predicted value for a certain sample; F base is the model's predicted benchmark value (expected value) for the total number of test set samples; φ i is the Shapley value of feature i of the sample; S is the feature subset that does not contain feature i; φ i The absolute value of represents the influence of the feature.
[0050] The present invention uses the above method to analyze the selected specific fault scenario, solves the Shapley value of the important feature of a single sample, explores the correlation trend between the feature itself and the transient power angle and voltage stability, quantitatively extracts the model decision rules, and combines the stability mechanism and expert experience analysis to demonstrate the credibility of the interpretable results.
[0051] The data-driven method for multi-task transient stability assessment of power systems provided by the present invention has the following beneficial effects compared with the prior art: (1) At the mechanism criterion level, the physical characteristics of the v'-Δv phase plane are analyzed based on the measurement of the terminal voltage and the load bus voltage, the mapping relationship between the transient stability state and the maximum Lyapunov exponent is established, and the power angle and voltage criterion containing more transient mechanism information is constructed. The criterion proposed by the present invention performs better in the assessment task than the conventional criterion. (2) At the model algorithm level, a dynamic heterogeneous graph multi-task transient stability assessment model taking into account the node heterogeneity and topological dynamics is proposed, and the performance of the commonly used graph deep learning model in various performance indicators is compared. It can be seen that the model proposed by the present invention achieves the best in terms of indicators. (3) At the model interpretation level, the model decision rules are quantitatively extracted with the help of the model-independent Shapley interpretability analysis algorithm, and the differences in the impact of different input features on multi-task transient stability prediction under specific fault scenarios are compared, and the transient mechanism and expert experience analysis are used to assist in verifying the credibility of the model in transient stability assessment. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 The topology diagram of the IEEE39 node test system constructed for the specific embodiment of the present invention.
[0053] Figure 2 It is a time-domain simulation curve of transient voltage in the scenario analyzed in the first stage of the specific embodiment of the present invention.
[0054] Figure 3 It is a two-dimensional v'-Δv phase trajectory curve in the scenario analyzed in stage 1 of the specific embodiment of the present invention.
[0055] Figure 4 This is the Lyapunov exponent change curve in the scenario analyzed in stage 1 of the specific embodiment of the present invention.
[0056] Figure 5 This is a flow chart of the technical solution for constructing the maximum Lyapunov exponent criterion in stage one of the specific implementation example of the present invention.
[0057] Figure 6 This is a basic principle diagram of the heterogeneous graph conversion architecture constructed in stage two of the specific embodiment of the present invention.
[0058] Figure 7 This is a basic architecture diagram of the dynamic heterogeneous graph model constructed in stage two of the specific embodiment of the present invention.
[0059] Figure 8 This is a flow chart of the technical solution for transient stability assessment of a power system in stage three of a specific embodiment of the present invention.
[0060] Fig. 9It is a radar chart comparing model performance indicators under different criterion frameworks in a specific embodiment of the present invention.
[0061] Fig.10 This is a graph showing the contribution of the single power angle evaluation task characteristics in a specific embodiment of the present invention.
[0062] Fig.11 Graph showing the contribution of voltage single assessment task characteristics in a specific embodiment of the present invention. DETAILED DESCRIPTION
[0063] The present invention is further described below in conjunction with specific embodiments.
[0064] This embodiment is based on Figure 1 The IEEE39 node test system shown in the figure is built in the power system transient security assessment software (TSAT) and the power system real-time data simulation tool (RTDS) as a hybrid electromechanical-electromagnetic simulation model. In order to consider the impact of new energy access on the transient stability of the system, the synchronous generators connected to nodes 35 and 37 are replaced with double-fed wind turbines according to the principle of single-machine multiplication.
[0065] A data-driven method for multi-task transient stability assessment of a power system provided by the present invention comprises the following steps:
[0066] Phase 1: First, verify the feasibility of the “voltage response information-phase trajectory concave-convex characteristics-maximum Lyapunov exponent” criterion framework, and use the maximum Lyapunov exponents of the terminal voltage and load bus voltage trajectory as the power angle and voltage transient stability criteria. The details are as follows:
[0067] S1-1: Phase trajectory physical characteristics analysis:
[0068] Take the given time domain simulation fault scenario conditions as an example for analysis: at 95% benchmark load, the most serious three-phase short circuit fault is applied at a location 40% away from the head end of the line (node 28-node 29) when the simulation duration is 1s. The following are obtained when the fault duration is 80ms and 100ms: Figure 2 The transient response trajectory of the generator terminal voltage amplitude and the load bus voltage amplitude is shown in FIG. Further, according to formula (1), the following are respectively drawn: Figure 3 The v'-Δv two-dimensional phase trajectory is shown. Figure 2 The middle dashed line shows the voltage response curve when the fault removal time is 80ms. The voltage amplitude at the generator end and the voltage amplitude at the load bus quickly recover to a stable value after being disturbed, corresponding to Figure 3 The middle dashed line shows that the horizontal and vertical coordinates of the phase trajectory in this scenario converge to a certain fixed value, which appears as a convergence trend on the two-dimensional phase plane; Figure 2The solid line in the middle shows the voltage response curve when the fault removal time is 100ms. The voltage amplitude at the generator end and the voltage amplitude at the load bus first drop after being disturbed and then continue to oscillate. They cannot tend to a stable value within the given time domain simulation time, corresponding to Figure 3 The solid line in the middle shows that the horizontal and vertical coordinates of the two-dimensional phase trajectory in this scenario cannot tend to a fixed value, and show a divergent trend on the two-dimensional phase plane. It can be seen that the power angle stability represented by the machine-end voltage amplitude and the voltage stability represented by the load bus amplitude can be analyzed from the contraction trend of the v'-Δv phase trajectory, thereby completing the mapping from the transient response information of different stable states to the concave and convex characteristics of the phase trajectory.
[0069] S1-2: Construction of the maximum Lyapunov exponent criterion:
[0070] The maximum Lyapunov exponent of time series data is usually obtained using the Wolf phase space reconstruction algorithm. The basic idea is to monitor the Euclidean distance between the initial point of the reference trajectory and the initial point of the adjacent trajectory. When the distance exceeds a certain threshold, the adjacent reference point is replaced by a nearby point of the current reference point until the end of the time series is reached. The calculation steps are roughly divided into:
[0071] (1) Phase space reconstruction: N-step observation reconstruction of the voltage time series state quantity of one dimension (single node) is selected to characterize the overall characteristics of the system.
[0072] (2) Parameter selection principle: The reference trajectory and the adjacent trajectory come from the same time series trajectory and differ in time by one sampling period. The starting point of the maximum Lyapunov exponent calculation begins when the ordinate of the phase trajectory reaches the maximum positive value for the first time.
[0073] (3) Calculate the maximum Lyapunov exponent according to formula (2).
[0074] According to the above solution method, the fault scenario set in S1 is analyzed, and the trajectory from the time of fault occurrence to the end of simulation is selected. Figure 4 The curve of the Lyapunov exponent changing with the number of iterations is shown. It can be found that, whether in the power angle or voltage stability scenario, the Lyapunov exponent tends to be stable with the number of iterations, and the maximum Lyapunov exponent is less than 0; on the contrary, in the instability scenario, the Lyapunov exponent continues to oscillate, and the maximum value is greater than 0. The judgment results are consistent with the voltage transient response curve and the phase trajectory concave-convex characteristic analysis results. This verifies the effectiveness of the proposed criterion analysis framework.
[0075] Figure 5 A flow chart of the technical solution for constructing the maximum Lyapunov exponent criterion proposed in this embodiment is shown.
[0076] Phase II: Considering three types of nodes, namely synchronous generator nodes, doubly-fed wind turbine nodes and busbar nodes, the time series electrical quantities and topological information of multi-time sections are selected as input features. The spatiotemporal dynamic characteristics of the system are extracted through the heterogeneous graph transformation architecture HGT and the time attention gated recursive unit TGRU. The constructed dynamic heterogeneous graph can effectively take into account the heterogeneity and topological changes of the nodes; the details are as follows:
[0077] S2-1: Model construction:
[0078] The dynamic heterogeneous graph model constructed in this embodiment takes the time series electrical quantity and the adjacency matrix of multiple time sections as input. For the selection of node input features, the steady-state time t0 before the fault occurs, the fault occurrence time t f and the fault removal phase t c -t c+3 There are a total of six node electrical quantities in time sections. The node types considered are synchronous generator nodes, double-fed wind turbine nodes and other bus nodes. Multi-dimensional input feature quantities at the system level and device level are selected respectively. Table 1 lists the specific features and dimensions selected in this embodiment.
[0079] Table 1
[0080]
[0081] Formula (9) gives the final input feature representation of different types of nodes:
[0082]
[0083] Among them, I, M, and N are three types of node sets respectively. and A t are the initial input electrical quantities and adjacency matrix at the corresponding moments respectively.
[0084] Figure 6 The schematic diagram of the heterogeneous graph conversion architecture constructed in this embodiment is shown; Figure 7 The basic cascade architecture of the dynamic heterogeneous graph model constructed in this embodiment is demonstrated, including a three-layer heterogeneous graph conversion architecture to realize feature extraction of heterogeneous neighbor nodes and information aggregation of multi-neighbor nodes, and the loss function is ReLu; a layer of time attention gated recursive unit to realize information expression of multi-section dynamic topology, and the activation function is Softmax; the downstream task outputs the instability probability of transient power angle and voltage evaluation tasks through a fully connected layer and a Softmax layer, and the instability threshold is set to 0.5. If the output value is greater than 0.5, it is judged to be in an unstable state. The number of heads of the multi-head attention mechanism involved in the algorithm is 4.
[0085] S2-2: Model performance evaluation indicator construction:
[0086] To evaluate the model performance, a performance evaluation index as shown in formula (10) is constructed based on the confusion matrix. Accuracy is one of the most commonly used indicators in classification problems, which represents the proportion of samples correctly classified by the model to the total samples. This embodiment constructs a global accuracy rate A cc , single task power angle and voltage accuracy A cc-δ and A cc-V Three accuracy indicators; at the same time, the false positive rate and false negative rate indicators F are constructed a and M r , which respectively represent the proportion of unstable samples predicted as stable samples to the total number of unstable samples and the proportion of stable samples that are not correctly predicted to the total number of stable samples.
[0087]
[0088] Where N c is the total number of samples in the test set.
[0089] Phase 3: Based on the Shapley method, the interpretability of the proposed algorithm is studied, the decision rules of the model are quantified, and the attribution and source analysis of the prediction results are completed to provide a basis for transient preventive measures in the online application phase. The details are as follows:
[0090] S3-1: Transient stability assessment process: Figure 8 The flowchart of the technical solution for transient stability assessment of the power system involved in this embodiment is shown.
[0091] S3-1-1: Sample generation and stability marking: The load level of the flow is set in steps of 5%, and 5 load levels are set from 90% to 110%; at each load level, a three-phase short-circuit fault is imposed on each line without a transformer; the fault is set at 10%, 40%, and 90% from the head end of each line; the fault access start time is 1s at the beginning of the simulation, and the fault duration is set to 2 to 8 cycles; the way to consider topological changes in the data set is to cut off the fault line (excluding islands) after the fault is cleared. Except for samples where the hybrid simulation failed and the flow did not converge, 2804 initial samples were finally obtained. Based on the maximum Lyapunov exponent criterion of this embodiment, the commonly used power angle transient stability index criterion, and the single binary meter voltage criterion (the time when the load bus voltage continues to drop below 0.75 (per unit value) during the transient process does not exceed 1s) is used as the standard), the stability state of each sample is marked. Table 2 lists the stability marking of various criteria.
[0092] Table 2
[0093]
[0094] S3-1-2: Offline training: The Adam optimization algorithm and Hamming loss function are used in this process; the batch size is 256; the iteration round is 150 rounds; the learning rate is 0.001; the deep learning model involved is implemented based on the Pytorch framework, and the computer configuration is: Inter Core i9-12900H CPU / 12.00GB RAM GPU / NVIDIA GeForce RTX3060. In order to comprehensively verify the performance of different criteria and model algorithms in multi-task transient stability assessment tasks, this embodiment selects conventional criteria and maximum Lyapunov criteria to train on the proposed model respectively. Fig. 9 The performance comparison of the models under different criteria is demonstrated. The two types of criteria have little difference in overall accuracy and accuracy of a single task. The false positive rate and false negative rate using the conventional criterion are 2.26% and 1.90% respectively, while the false positive rate and false negative rate using the maximum Lyapunov criterion are reduced to 0.58% and 0.62%, which verifies the effectiveness of the proposed maximum Lyapunov index criterion in transient stability assessment tasks.
[0095] S3-1-3: Online application: In this stage, the PMU is used to monitor the grid measurement information in real time, and the information is input into the trained model to quickly give the grid transient stability status judgment results under different operation scenarios, so that the control personnel can take preventive control measures. At the same time, the confidence level can be set and human assistance can be intervened to ensure accurate assessment of the system stability situation;
[0096] This embodiment selects common graph deep learning methods for comparison: graph convolutional neural network, graph attention network, heterogeneous graph attention network, heterogeneous graph conversion architecture, and dynamic heterogeneous graph model combining heterogeneous graph conversion architecture and time attention gated recursive unit. The input data and structural parameters of the comparison model are as follows: graph convolutional neural network and graph attention network cannot directly process input node information of different dimensions. First, the original high-dimensional time series electrical quantity needs to be reduced to the same dimension; heterogeneous graph attention network requires artificial design of heterogeneous graph element path, while heterogeneous graph conversion architecture can implicitly and automatically extract the element path with connecting edges. The number of intermediate layers in the above models is 3 layers, and the number of hidden layer units in each layer is consistent. Table 3 shows the analysis results. The overall accuracy of the proposed dynamic heterogeneous graph model is better than other models, and the accuracy of a single evaluation task is more than 99%. Heterogeneous graphs take into account the differences in nodes in the process of executing attention, retain the transient information contained in different nodes themselves at the input data level, and are better than convolutional neural networks, graph attention networks and other isomorphic graph models in terms of various performance indicators such as misjudgment rate and missed rate. Compared with the static graph model, the dynamic heterogeneous graph model designed in this embodiment executes the temporal self-attention mechanism for the changing topology of multiple time sections, which can effectively capture the temporal topological evolution pattern, achieve the best in various performance indicators, and can adaptively learn the input features of important moments.
[0097] Table 3
[0098] Model Algorithm A / % <![CDATA[A cc-δ / %]]> <![CDATA[A cc-V / %]]> F / % M / % Graph Convolutional Neural Networks 96.32 97.72 98.01 1.27 2.82 Graph Attention Network 96.79 97.97 98.31 0.95 2.26 Different Graph Attention Network 97.86 98.72 98.89 0.68 1.24 Heterogeneous Graph Conversion Architecture 98.28 99.03 99.14 0.77 0.90 The dynamic heterogeneous graph model proposed in this embodiment 98.68 99.32 99.25 0.58 0.62
[0099] S3-2: Single sample rule extraction for model prediction mechanism;
[0100] This embodiment takes the given time-domain simulation fault scenario conditions as an example for analysis: the load level is 100%, a three-phase short circuit fault occurs in branch 5-8 at 40% of the distance from the line head end, and the fault duration is 5 cycles; Table 4 shows the stability prediction results and their judgment basis in this scenario. It can be found that the mechanism analysis results on the power angle and voltage evaluation tasks are consistent with the model prediction results.
[0101] Table 4
[0102] Criteria Type in accordance with Prediction results Transient stability index power angle criterion TSI=65.67>0 Stablize Maximum Lyapunov exponent power angle criterion MLE=-1.1208<0 Stablize Voltage criterion for single binary meter Load busbars are within the limit Stablize Maximum Lyapunov exponent voltage criterion MLE=-1.2547<0 Stablize
[0103] Fig.10 and Fig.11The Shapley value attribution analysis of the single transient power angle assessment and transient voltage assessment tasks under the above fault scenarios is shown respectively. The model's expectation for the transient power angle stability task is 0.4344, and the expectation for the transient voltage power angle task is 0.5846. The Shapley values corresponding to the six features that have the greatest impact on the evaluation results are marked in the figure. Positive values represent a positive effect on the prediction result of instability (left part), and negative values represent a positive effect on the prediction result of stability (right part). The output values of the model proposed in this embodiment for the transient power angle assessment task and the transient voltage assessment task are 0.19 and 0.18 respectively, both of which are less than the set instability probability threshold of 0.5, so they are all predicted to be stable, indicating that the overall contribution of each feature to the prediction of stability is greater than the overall contribution to the prediction of instability.
[0104] Feature 30 (the terminal voltage of the generator at node 39) plays a positive role in predicting transient power angle and voltage stability. Features 35-37 (the power angles of the synchronous generators connected to nodes 34, 36, and 38, respectively) play a positive role in predicting transient power angle stability. Feature 73 (the voltage amplitude of the doubly fed wind turbine node 35) and feature 31 (the power angle of the generator at node 32) are the two features that contribute most to the prediction of power angle instability. Feature 69 (the voltage amplitude of node 31), feature 21 (the terminal voltage of the synchronous generator node 30) and feature 18 (the output reactive power of the doubly fed wind turbine at node 37) are the three features that contribute most to the prediction of voltage instability. The reasons are analyzed as follows:
[0105] (1) Node 39 is both a synchronous machine node and a load node. The generator and load power can be balanced, so it has a positive effect on the system transient power angle and voltage stability;
[0106] (2) Nodes 31 and 32 are the two synchronous generators closest to the fault line. After time domain curve analysis, the busbar amplitude of load node 31 drops the most when the fault begins. The power angle of the generator connected to synchronous generator node 32 swings away first and then recovers. The power angle stability state analyzed by using the generator terminal voltage is consistent with the conclusion of the generator power angle state judgment in the conventional sense.
[0107] (3) The output reactive power of the doubly-fed wind turbine connected to node 37 is large, indicating that the system has a large reactive power shortage and a high risk of transient voltage instability.
[0108] (4) The features related to transient power angle instability are concentrated on the generator side variables, while the features related to transient voltage instability are concentrated on the load side variables, which is consistent with the conventional mechanism cognition of the two types of stability categories.
[0109] The scope of protection claimed by the present invention is not limited to the above specific embodiments. For those skilled in the art, the present invention may have various deformations and changes. Any modifications, improvements and equivalent substitutions made within the concept and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A data-driven method for multi-task transient stability assessment of power systems, characterized by: The steps include: Phase 1: First, the feasibility of the "voltage response information-phase trajectory concave-convex characteristics-maximum Lyapunov exponent" criterion framework was verified, and the maximum Lyapunov exponents of the generator terminal voltage and load bus voltage trajectory were used as the power angle and voltage transient stability criteria respectively; Phase II: Considering three types of nodes, namely synchronous generator nodes, doubly-fed wind turbine nodes and busbar nodes, the time series electrical quantities and topological information of multi-time sections are selected as input features. The spatiotemporal dynamic characteristics of the system are extracted through the heterogeneous graph transformation architecture HGT and the time attention gated recursive unit TGRU. The constructed dynamic heterogeneous graph can effectively take into account the heterogeneity and topological changes of nodes. Phase 3: Based on the Shapley method, the interpretability of the proposed algorithm is studied, the decision rules of the model are quantified, and the attribution and traceability analysis of the prediction results is completed to provide a basis for transient preventive measures in the online application phase.
2. A data-driven method for multi-task transient stability assessment of power systems according to claim 1, characterized in that: Phase 1 is as follows: S1-1: Phase trajectory physical characteristics analysis: The v'-Δv phase trajectory is plotted with the generator terminal voltage and the load bus voltage deviation as the horizontal coordinates and the voltage amplitude change rate as the vertical coordinate, as shown in formula (1): Among them, v kΔt and v (k-1)Δt are the voltage amplitude measurements at time kΔt and time (k-1)Δt, Δv kΔt is the deviation between the voltage and the reference voltage at time kΔt, v' kΔt is the voltage amplitude change rate at the current moment relative to the previous moment, f s is the sampling frequency; By analyzing the phase plane composed of different transient voltage response information and observing the physical characteristics of the phase trajectory in the two-dimensional plane, and based on this characteristic, a detailed analysis is made of the system during transient disturbance. The stable and unstable scenarios show different divergent or contraction trends on the phase plane, so as to characterize different transient stability types and states. S1-2: Construction of the maximum Lyapunov exponent criterion: If the value of the maximum Lyapunov exponent is greater than zero, it is considered that the distance between adjacent trajectories in the phase space of the system increases exponentially with time, and the system is considered to be chaotic. If it is less than zero, it is considered that the system tends to be stable. The calculation method is given by formula (2): The subscript values represent the sampling moments corresponding to the voltage values.
3. A data-driven method for multi-task transient stability assessment of power systems according to claim 2, characterized in that: Phase 2 is as follows: S2-1: Model construction: The dynamic heterogeneous graph model includes two components: heterogeneous graph conversion architecture and temporal attention gated recursive unit. The heterogeneous graph conversion architecture takes into account the differences in the input feature dimensions of nodes in the power grid and retains important node information. The core idea is to update the target node information by aggregating the neighbor node information of the target node. It is divided into three modules: heterogeneous mutual attention, heterogeneous message passing and aggregation. The temporal attention gated recursive unit captures the evolution pattern of the topology through the changes in the node embedding of multiple time sections, and introduces the temporal attention mechanism to learn the embedding representation of different nodes at different times. S2-1-1: Heterogeneous graph conversion architecture: Given a graph G(V,E,O,R), V and E are the node and edge sets respectively, O and R are the node and edge types respectively, and there are type mapping functions: τ(v):V→O,φ(e):E→R; the attention weight operation for the target node n and its neighbor node m∈M(t) is given by formula (3): in, Nodes m and n are linearly mapped and The vector of the i-th attention head obtained; is the edge type maintenance matrix, μ is the attention scaling factor; M(t) is the set of neighbor nodes; h is the number of attention heads; X (l-1) [*] is the output vector of the l-1th layer. If l=1, then X is the original input feature X0 of the model; The message passing operation of multi-head attention is given by formula (4): in, is the vector of the i-th attention head obtained after linear mapping of the source node m, The maintenance matrix for edge types; Aggregation operation aggregates the information of all neighbor nodes to the target node based on the attention weight. Finally, the target node response of the lth layer is mapped to the distribution of the corresponding node category, and the residual connection is introduced to obtain the final representation of the target node n as shown in formula (5): in, For direct and operation, A-Linear τ(n) is a linear mapping operation, δ is an activation function; S2-1-2: Temporal Attention Gated Recurrent Unit: The structure is a cascade structure of multiple layers of gated recurrent units and a temporal attention mechanism layer. The embedded representation of each node at time t is obtained by the heterogeneous graph conversion architecture. The iterative formula of the gated recursive unit at time t is shown in formula (6): Among them, z t ,r t are the update gate and forget gate at time t respectively; W z ,W r ,W s ,b z ,b r ,b s is a trainable parameter; σ is an activation function; x t is the input feature at time t, obtained from the output of the heterogeneous graph conversion architecture; s t and t-1 are the outputs of the hidden layer at time t and the previous time respectively; The hidden layer output of the gated recurrent unit at node i at different times As input, the temporal attention performs a scaled dot product attention mechanism, and its output Given by formula (7): in is a trainable parameter; is the mask matrix; S2-2: Model performance evaluation indicator construction: A 4×4 multidimensional confusion matrix is constructed to visualize the model classification effect, where each column represents the predicted category of the sample and each row represents the true category of the sample; each sample label is described by a 1×2 matrix, and the first element of the matrix represents the stable state of the power angle, where the stable label is δ s , the instability label is δ us ; The second element represents the voltage stable state, the stable label is V s , the instability label is V us , so there are four types of probabilistic label combinations, namely [δ s ,V s ]、[δ s ,V us ]、[δ us ,V s ]、[δ us ,V us ]; Each element in the confusion matrix is the count value of different combinations of true labels and predicted labels, corresponding to the above four situations. The sample prediction category is consistent with the true belonging category, which is represented by T, and the sample prediction category is inconsistent with the true belonging category, which is represented by F. The subscript values represent the row number and column number, corresponding to different stable scenarios.
4. A data-driven method for multi-task transient stability assessment of power systems according to claim 3, characterized in that: Phase 3 is as follows: S3-1: Transient stability assessment process: S3-1-1: Sample generation and stability annotation: The time-domain simulation operation scenario generates an initial sample set by taking into account factors such as different power flow operation modes, different fault information and clearing methods. Different input electrical quantities covering all transient characteristics are selected for each type of node, and transient response trajectories are extracted. The transient power angle and voltage stability state of the sample set are annotated based on the maximum Lyapunov index criterion; S3-1-2: Offline training: The sample set is randomly divided into a training set and a test set according to a ratio of 7:
3. The training set is used to train the model. During the training process, the model parameters are updated by backpropagation with the goal of minimizing the loss function until the loss function value meets the threshold condition. The model performance is evaluated on the test set based on the trained model. The prediction contribution of each feature in a specific scenario is quantitatively analyzed based on the Shapley method, and the extracted decision rules are further demonstrated in combination with electrical field knowledge. S3-1-3: Online application: In this stage, the PMU is used to monitor the grid measurement information in real time, and the information is input into the trained model to quickly give the grid transient stability status judgment results under different operation scenarios, so that the control personnel can take preventive control measures. At the same time, the confidence level can be set and human assistance can be intervened to ensure accurate assessment of the system stability situation; S3-2: Single sample rule extraction for model prediction mechanism; The Shapley method is used to quantify the contribution of input features to the model output. The model's prediction value for one of the samples is equal to the sum of the baseline prediction value and the characteristic Shapley value contained in the sample. The calculation method is shown in formula (8): Among them, N is the total number of input features; F is the model's predicted value for a certain sample; F base is the model's predicted benchmark value (expected value) for the total number of test set samples; φ i is the Shapley value of feature i of the sample; S is the feature subset that does not contain feature i; φ i The absolute value of represents the influence of the feature.
Citation Information
Cited By
Method, system and device for evaluating transient power angle stability of power system and medium
CN119294271A
Power system transient power angle stability assessment method, system, device and medium
CN119294271B
Method for building power grid transient voltage severity index calculation model
CN119692182A
Scene reduction-based power angle stability evaluation method and system for power system
CN119808538A
A method and system for evaluating power system power angle stability based on scenario reduction
CN119808538B
