Iterative optimization learning-based interpretable frequency stability prediction method

Abstract

The invention discloses an interpretable frequency stability prediction method based on iterative optimization learning. The method comprises the steps of performing N times of time domain simulation for N fault types of a target power system to obtain N groups of long-scale time domain simulation data; constructing time series data input by the model; constructing model input spatial data; recording a frequency minimum value f in each group of time domain simulation data, judging the frequency stability of the power system according to the f, and generating a label for the time domain simulation data according to a judgment result; on the basis of the time sequence data and the corresponding labels and spatial data, deep learning is carried out by adopting a residual image neural network, and system frequency stability prediction information is output; debugging and optimizing hyper-parameters of the residual image neural network model to obtain a pre-trained residual image neural network model; counting the prediction accuracy of the prediction information corresponding to the test data; evaluating and optimizing the prediction accuracy of the preset model; and carrying out interpretability analysis on a prediction result of the residual image neural network model.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to an interpretable frequency stability prediction method based on iterative optimization learning. Background Art

[0002] The application of new energy power generation and UHV transmission technology not only alleviates the problem of power surplus in the power grid and promotes the sustainable development of the power industry, but also improves the efficiency of resource allocation and helps to cope with the growing load demand of the power system. However, with the increase in the penetration rate of new energy and the increase in UHV cross-regional transmission power, the power system faces increasingly severe risks and challenges in transient frequency stability. On the one hand, the rotational inertia of the power system in the asynchronous operation state is significantly lower than that of the traditional synchronous large power grid, resulting in the system showing low inertia characteristics, and the ability to resist disturbances such as power mutation is greatly reduced. The inertia of UHV fed into the power grid is reduced, which further weakens the system frequency support capacity, making the system more prone to transient frequency instability after a large disturbance. On the other hand, with the continuous increase in the capacity of new energy grid connection, the uncertainty and randomness of system operation are further enhanced, and the power imbalance phenomenon in the system is more significant, which leads to an increase in the risk of transient frequency stability of the system. When the system suffers from UHV AC and DC faults, the transient frequency stability of the sending and receiving power grids will face serious threats, and the transient frequency stability of the system will face severe challenges. In this context, there is an urgent need for a power system transient frequency stability prediction method to predict the evolution of the power system transient frequency, so as to make timely intervention before the frequency instability event occurs, effectively curb the occurrence of frequency instability events, and prevent the power system frequency from being in an unstable state for a long time.

[0003] In response to the challenges of transient frequency stability prediction of power systems, the main research strategies currently include time domain simulation method, equivalent model method and data-driven method. The first two methods are difficult to strike a balance between computational efficiency and prediction accuracy; while the data-driven method shows advantages in efficiency and accuracy. However, the current data-driven method mainly relies on the historical operation data of the target power system or the transient time series data generated during the simulation process, which to a certain extent ignores the complex internal connections between nodes in the power grid and pays insufficient attention to the spatial correlation between nodes in the actual power grid, resulting in insufficient feature learning. At the same time, the quality and quantity of training data of the data-driven model also affect the performance of stability assessment. In addition, the prediction process of the data-driven method has a significant black box characteristic, that is, it is difficult for researchers to intuitively understand the working principle and specific prediction process of the method. This characteristic makes the promotion and implementation of this method in practical applications face great obstacles. Therefore, it is urgent to further develop the frequency stability prediction method of power systems based on spatiotemporal information, while improving the timeliness and accuracy of the prediction, to conduct interpretable analysis of the model prediction results, clarify the key inputs that affect the prediction accuracy of the model, and provide guiding information for subsequent frequency stability control. Summary of the invention

[0004] In view of this, the present invention provides an interpretable frequency stability prediction method based on iterative optimization learning, which is used to at least solve the problems of the power system transient frequency stability prediction technology in the prior art.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: An interpretable frequency-stable prediction method based on iterative optimization learning, comprising the following steps: S1. Perform N time-domain simulations for N fault types of the target power system to obtain N groups of long-scale time-domain simulation data; set the fault type at the beginning of each time-domain simulation, and obtain N groups of long-scale time-domain simulation data within a long time-scale window after the set fault occurs. Each group of long-scale time-domain simulation data includes the active power data P and frequency data F of each monitoring node in the current long time-scale window, and obtain the minimum system frequency value from the frequency data F. ; S2. Construct the time series data of the model input; in each set of long-scale time domain simulation data, select a short observation time window after the fault occurs , the short-term observation time window Active power data within and frequency data As time series data, the time series data is integrated into a time series observation matrix; S3. Constructing the model input spatial data; obtaining the impedance of each monitoring node in the target power system, obtaining the electrical connection matrix according to the relationship between the impedances of each monitoring node, and using the electrical connection matrix as spatial data; S4. Record the lowest frequency value in each set of long-scale time domain simulation data ,according to The frequency stability of the power system is judged by the size, and labels are generated for the time domain simulation data according to the judgment results. If the transient frequency minimum value Greater than the preset system frequency stability threshold Then it is judged that the power system frequency is stable; S5. Based on the time series data and its corresponding labels and spatial data, the residual graph neural network is used for in-depth learning to output the system frequency stability prediction information; S6. Debug and optimize the hyperparameters of the residual graph neural network model to obtain the hyperparameter combination with the highest prediction accuracy, and obtain the pre-trained residual graph neural network model; S7. Obtain test data, obtain prediction information corresponding to the test data through the pre-trained residual graph neural network model, and calculate the prediction accuracy of the prediction information corresponding to the test data, wherein the test data includes test time series data, test space data, and test label data; S8. According to the preset model prediction accuracy evaluation standard, the pre-trained residual graph neural network model that does not meet the evaluation standard is optimized again. If it meets the evaluation standard, it is used as the final residual graph neural network model and enters S9; S9. Perform interpretability analysis on the prediction results of the residual graph neural network model.

[0006] Preferably, the specific content of the time series observation matrix in S2 is: Each row represents a short observation time window The active power data P and frequency data F of each monitoring node at any time within the short observation time window. Each column represents the active power data P and frequency data F of each monitoring node at any time within the short observation time window. The change track of active power data P or frequency data F within; Before integrating the time series observation matrix, the active power data P and frequency data F are standardized to ensure that they have a uniform dimension and a mean of zero.

[0007] Preferably, the specific content of obtaining the electrical connection matrix in S3 includes: Obtain the impedance matrix of key bus nodes in the target power system: , in , and Represent the number of rows and columns of the impedance matrix, Represents the total number of critical bus nodes in the system; Calculate the electrical connection matrix: , In order to characterize the relative distance between each node, a normalization operation is performed: , in, is the normalized electrical connection matrix.

[0008] Preferably, the specific content of S4 includes: During the time domain simulation of each long time scale window, the minimum value of the system transient frequency is obtained , preset system frequency stability threshold ; when When , it is judged that there is a risk of frequency instability in the power system, and the predicted output system frequency stability result S=0; when When , the power system frequency is judged to be stable, and the predicted output system frequency stability result S=1.

[0009] Preferably, the specific structure of the residual graph neural network in S5 includes: The residual graph neural network adopts a dual-channel input architecture, which is used to input time series data and spatial data respectively; The two-dimensional convolution module includes a convolution kernels, which are used to extract features from the input time series data and spatial data respectively to obtain a feature map; The residual module is used to mine the key features of the data according to a feature map through the initial residual and identity mapping. The key features form a one-dimensional vector and are connected to the fully connected layer to output the feature vector; The feature vector is activated by the softmax function to predict the frequency stability result of the output system.

[0010] Preferably, the specific content of S7 includes: Obtain test time series data, test space data and test label data, input the test time series data and test space data into the pre-trained residual graph neural network model to obtain the corresponding prediction information, compare the prediction information with the test label data, and count the ratio of the number of correct prediction information Nr to the total number N, and then obtain the model prediction accuracy R=Nr / N×100%.

[0011] Preferably, the specific content of S8 includes: If the accuracy R reaches or exceeds the preset accuracy threshold, it is determined that the current pre-trained residual graph neural network model is reliable in predicting power system frequency stability, and then enters S9; If the accuracy rate does not meet the standard, the pre-trained residual graph neural network model that does not meet the evaluation standard is optimized again until the accuracy rate R meets the preset accuracy rate threshold, and then enter S9; The steps of further optimization include: The test time series data, test space data and test label data obtained in S7 are combined with the time series data, space data and label data as new training data for the residual graph neural network model; Input the new training data into the residual graph neural network model, and gradually reduce the learning rate to perform optimization training again; After optimizing the training again, continue to judge the accuracy. If the accuracy R still does not reach 98.5%, repeat step S8. Otherwise, go to S9.

[0012] Preferably, the specific contents of the interpretability analysis of the prediction results of the residual graph convolutional network model in S9 include: Assume that the input time series observation matrix is , among which The input is , the prediction value of the residual graph convolutional network model for the system frequency stability state is S, and the mean contribution of all inputs of the residual graph convolutional network model is , then the contribution value follows: , in represent The contribution value of is used to reflect the contribution of the first The contribution of the first input to the final prediction value. The input is Active power data and frequency data at each moment; like , it means that in the time series data matrix The represented feature quantity improves the prediction value and has a positive effect; like , it means that in the time series data matrix The represented feature quantity leads to a decrease in the predicted value and has an adverse or no effect; By summing up the input contribution corresponding to each node, the total contribution of each monitoring node in the system is obtained, and the monitoring nodes are sorted according to the size of the total contribution. The bus nodes ranked before the preset ranking are the key bus nodes that have a significant impact on the system frequency evolution.

[0013] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses an interpretable frequency stability prediction method based on iterative optimization learning, which has the following beneficial effects: The present invention uses residual graph neural network technology to accurately predict the transient frequency stability of the power system under long time scales, and analyzes and identifies the key busbars that affect the system frequency stability results, thereby providing technical support for the safe and stable operation and control of the frequency of low inertia power systems. This technology aims at the defects of existing data-driven methods that mainly rely on time series information and ignore other important information. The spatial information of the power system nodes is integrated into the residual graph neural network as a supplement to promote the residual graph convolution network to more effectively learn and reveal the complex relationship between the spatiotemporal information of the power system and the transient frequency change, and make more full use of the known spatiotemporal information of the power system, so as to accurately predict the system frequency stability state. On this basis, in order to further optimize the prediction accuracy of the residual graph neural network, this technology adopts an iterative optimization learning method to optimize and supplement the residual graph neural network training data to improve its prediction accuracy. In addition, the present invention adopts the interpretability method SHAP to identify the key nodes that have a decisive influence on the prediction results of the residual graph convolution network, and verifies the reliability of the residual graph neural network combined with spatiotemporal information in the prediction results by explaining the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 A flowchart of an interpretable frequency stability prediction method based on iterative optimization learning provided by the present invention; Figure 2 A schematic diagram of an execution framework provided for an embodiment of the present invention; Figure 3 A diagram showing the structure of a matrix of input spatiotemporal data provided by an embodiment of the present invention; Figure 3 (a) is the time series observation matrix, Figure 3 (b) is the electrical connection matrix; Figure 4 A schematic diagram of a residual graph neural network structure provided by an embodiment of the present invention; Figure 5 A schematic diagram of the process structure of iterative optimization learning provided by an embodiment of the present invention; Figure 6 A schematic diagram of simulation results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] The present invention provides an interpretable frequency stability prediction method based on iterative optimization learning, such as Figure 1-2 As shown, the following steps are included: S1. Perform N time-domain simulations for N fault types of the target power system to obtain N groups of long-scale time-domain simulation data; set the fault type at the beginning of each time-domain simulation, and obtain N groups of long-scale time-domain simulation data within a long time-scale window after the set fault occurs. Each group of long-scale time-domain simulation data includes the active power data P and frequency data F of each monitoring node in the current long time-scale window, and obtain the minimum system frequency value from the frequency data F. ; S2. Construct the time series data of the model input; in each set of long-scale time domain simulation data, select a short observation time window after the fault occurs , the short-term observation time window Active power data within and frequency data As time series data, the time series data is integrated into a time series observation matrix. The time series observation matrix is ​​as follows Figure 3 (a) as shown; Figure 3 (a), The first At each moment, f and p represent active power and frequency respectively. Label the monitoring node; S3. Construct the model input spatial data; obtain the impedance of each monitoring node in the target power system, obtain the electrical connection matrix according to the relationship between the impedances of each monitoring node, and use the electrical connection matrix as spatial data. The electrical connection matrix is ​​as follows: Figure 3 (b) as shown; Figure 3 (b), Representative monitoring nodes, the space matrix is ​​a symmetric matrix, so the horizontal and vertical coordinates are repeated; S4. Record the lowest frequency value in each set of long-scale time domain simulation data ,according to The frequency stability of the power system is judged by the size, and labels are generated for the time domain simulation data according to the judgment results. If the transient frequency minimum value Greater than the preset system frequency stability threshold Then it is judged that the power system frequency is stable; S5. Based on the time series data and its corresponding labels and spatial data, the residual graph neural network is used for in-depth learning to output the system frequency stability prediction information; S6. Debug and optimize the hyperparameters of the residual graph neural network model to obtain the hyperparameter combination with the highest prediction accuracy, and obtain the pre-trained residual graph neural network model; S7. Obtain test data, obtain prediction information corresponding to the test data through the pre-trained residual graph neural network model, and calculate the prediction accuracy of the prediction information corresponding to the test data, wherein the test data includes test time series data, test space data, and test label data; S8. According to the preset model prediction accuracy evaluation standard, the pre-trained residual graph neural network model that does not meet the evaluation standard is optimized again. If it meets the evaluation standard, it is used as the final residual graph neural network model and enters S9; S9. Perform interpretability analysis on the prediction results of the residual graph neural network model.

[0018] It should be noted that: Usually, the system frequency will reach the lowest value within a long period of time from tens of seconds to several minutes after the fault occurs. Therefore, the long time scale window is usually set to 10s~60s; the short time observation time window t is usually set to a time period less than 0.5s, which is much smaller than the time domain simulation duration; In this embodiment, the Optuna algorithm is used to automatically debug and optimize the hyperparameters of the residual graph neural network model (including Layers, LR, Batch Size, with initial values ​​of 50, 0.001, and 64 respectively); The specific method for obtaining test data in S7 is: generate a new batch of time domain simulation data using the S1 method, and obtain new time series data, spatial data, and label data through steps S2, S3, and S4.

[0019] In addition, in this embodiment, each node refers to a bus node in the power system equipped with a measuring device such as a PMU.

[0020] In this embodiment, for an actual receiving-end power grid in a certain area, four typical operating modes are set, namely, large flood, small flood, large flood, and large drought. At the same time, a UHV AC N-2 line break fault and a UHV DC single-pole and double-pole locking fault are set. The simulation time of each time is set to 60s, and 2196 time domain simulation data are obtained through batch simulation; the fault start time is set to the 3rd second after the simulation, and the time period of 3s-3.3s after the start of the simulation is set as the short-term observation time window, and the active power data and frequency data in the time window are recorded as the input time series data; these data are organized into a two-dimensional matrix, in which each row represents the characteristic information of each node at a certain moment, and each column represents the dynamic response trajectory of a specific electrical quantity of a certain node in the observation time window; for the active power data P and frequency data F in the matrix, standardization processing is also required to ensure that they have a unified dimension and a mean of zero, so as to eliminate the possible influence of different features due to differences in magnitude or distribution.

[0021] S6 carefully debugs the hyperparameters of the residual graph neural network, optimizes the initial learning rate, tries different types of loss functions, and observes the impact of the time window length. At the same time, it adjusts the normalization method of the input spatiotemporal data in order to obtain the residual graph convolutional network with the highest prediction accuracy for subsequent use.

[0022] In order to further implement the above technical solution, the specific content of the time series observation matrix in S2 is: Each row represents a short observation time window The active power data P and frequency data F of each monitoring node at any time within the short observation time window. Each column represents the active power data P and frequency data F of each monitoring node at any time within the short observation time window. The change track of active power data P or frequency data F within; Before integrating the time series observation matrix, the active power data P and frequency data F are standardized to ensure that they have a uniform dimension and a mean of zero.

[0023] In order to further implement the above technical solution, the specific contents of obtaining the electrical connection matrix in S3 include: Obtain the impedance matrix of key bus nodes in the target power system: , in , and Represent the number of rows and columns of the impedance matrix, Represents the total number of critical bus nodes in the system; Calculate the electrical connection matrix: , In order to characterize the relative distance between each node, a normalization operation is performed: , in, is the normalized electrical connection matrix.

[0024] It should be noted that: The present invention uses the electrical connection matrix between key nodes in the target system as spatial data. In view of the fixedness of the system grid structure and the fact that failures will not affect the network architecture, the electrical connection matrix is ​​constant and unique, and is easy to obtain; the electrical connection matrix can effectively reflect the close relationship between nodes, even when there is no direct AC line connection between some nodes; In this embodiment, the value of G is 21.

[0025] In order to further implement the above technical solution, the specific contents of S4 include: During the time domain simulation of each long time scale window, the minimum value of the system transient frequency is obtained , preset system frequency stability threshold ; when When , it is judged that there is a risk of frequency instability in the power system, and the predicted output system frequency stability result S=0; when When , the power system frequency is judged to be stable, and the predicted output system frequency stability result S=1.

[0026] It should be noted that: In each time domain simulation process, in the 30s-50s period after the short-term observation time window, the system transient frequency will evolve to a minimum frequency value. , the lowest frequency It is a key indicator for judging whether the system frequency is unstable. This key information is accurately screened out as sample label data for training and verification of residual graph convolutional network. According to the guidelines for safe operation of power grid, the system frequency stability threshold is set in this embodiment. is 49.5Hz.

[0027] In order to further implement the above technical solution, Figure 4 As shown, the specific structure of the residual graph neural network in S5 includes: The residual graph neural network adopts a dual-channel input architecture, which is used to input time series data and spatial data respectively; The two-dimensional convolution module includes a convolution kernels, which are used to extract features from the input time series data and spatial data respectively to obtain a feature map; The residual module is used to mine the key features of the data according to a feature map through the initial residual and identity mapping. The key features form a one-dimensional vector and are connected to the fully connected layer to output the feature vector; The feature vector is activated by the softmax function to predict the frequency stability result of the output system.

[0028] It should be noted that: This embodiment is based on the Python language and the Pytorch framework to construct a residual graph neural network with a dual-channel input architecture, which is used to input the timing information and spatial information of the nodes respectively, and output the system frequency stability prediction information, thereby revealing the mapping relationship between the input spatiotemporal information and the system frequency evolution results. The input spatiotemporal information x passes through the residual block structure, which can solve the problem of over-smoothing of traditional deep neural networks and deeply explore the intrinsic connection of training data. In the residual block structure, the residual mapping F(x) is fitted through stacked weight layers, F(x)=H(x)-x, where H(x) represents the original mapping. Even if a large number of layer structures are stacked in the network, the residual mapping relationship can ensure that the output of each layer structure retains at least part of the characteristics of the input spatiotemporal information, which alleviates the problem of gradient disappearance to a certain extent and ensures the accuracy of the prediction results; In order to further implement the above technical solution, the specific contents of S7 include: Obtain test time series data, test space data and test label data, input the test time series data and test space data into the pre-trained residual graph neural network model to obtain the corresponding prediction information, compare the prediction information with the test label data, and count the ratio of the number of correct prediction information Nr to the total number N, and then obtain the model prediction accuracy R=Nr / N×100%.

[0029] In order to further implement the above technical solution, the specific contents of S8 include: If the accuracy R reaches or exceeds the preset accuracy threshold, it is determined that the current pre-trained residual graph neural network model is reliable in predicting power system frequency stability, and then enters S9; If the accuracy rate does not meet the standard, the pre-trained residual graph neural network model that does not meet the evaluation standard is optimized again until the accuracy rate R meets the preset accuracy rate threshold, and then enter S9; The steps of further optimization include: The test time series data, test space data and test label data obtained in S7 are combined with the time series data, space data and label data as new training data for the residual graph neural network model; Input the new training data into the residual graph neural network model, and gradually reduce the learning rate to perform optimization training again; After optimizing the training again, continue to judge the accuracy. If the accuracy R still does not reach 98.5%, repeat step S8. Otherwise, go to S9.

[0030] It should be noted that: In this embodiment, the preset accuracy threshold is set to 98.5%. Figure 5 As shown, the new time series data, spatial data and label data obtained in S7 are merged into the time series data, spatial data and label data of S2-S4 as new training data for the residual graph neural network; in the optimization training process of the residual graph neural network, other parameters are not changed, and the learning rate is set to a smaller value for fine-tuning to ensure that the residual graph convolutional network does not significantly change the original parameters when learning new data, thereby ensuring that the residual graph convolutional network is further improved on the basis of the existing accuracy; after the residual graph convolutional network is optimized, continue with S7. If the accuracy R still does not reach 98.5%, repeat the current step, otherwise, enter S9.

[0031] In order to further implement the above technical solution, the specific contents of the interpretability analysis of the prediction results of the residual graph convolutional network model in S9 include: Assume that the input time series observation matrix is , among which The input is , the prediction value of the residual graph convolutional network model for the system frequency stability state is S, and the mean contribution of all inputs of the residual graph convolutional network model is , then the contribution value follows: , in represent The contribution value of is used to reflect the contribution of the first The contribution of the first input to the final prediction value. The input is Active power data and frequency data at each moment; like , it means that in the time series data matrix The represented feature quantity improves the prediction value and has a positive effect; like , it means that in the time series data matrix The represented feature quantity leads to a decrease in the predicted value and has an adverse or no effect; By summing up the input contribution corresponding to each node, the total contribution of each monitoring node in the system is obtained, and the monitoring nodes are sorted according to the size of the total contribution. The bus nodes ranked before the preset ranking are the key bus nodes that have a significant impact on the system frequency evolution.

[0032] It should be noted that: This embodiment adopts an additive interpretation algorithm SHAP, which can generate corresponding contribution values ​​for each time series data input. Through the size of these contribution values, the key bus nodes that affect the stability of the system frequency can be determined. Assume that the input time series data matrix is , in the matrix The input is , the prediction value of the residual graph convolutional network for the system frequency steady state is S, and the benchmark of the entire network (i.e. the mean of all input contributions) is , then the contribution value follows the following equation: ,in represent Contribution value. Intuitively speaking, It reflects the contribution of the first input in the current input time series matrix to the final prediction value. , it indicates that the feature improves the prediction value and has a positive effect; on the contrary, if , it indicates that this feature leads to a decrease in the predicted value and has a counter-effect. The input time series data matrix records the frequency and active power of 21 key bus nodes at different times. By summing the input contribution corresponding to each node, the total contribution value of the key bus nodes in the system can be obtained, and the nodes are sorted according to the size of their contribution. The top-ranked bus nodes are the key bus nodes that have a significant impact on the frequency evolution of the system. The size of the contribution value and its sorting can effectively explain the prediction logic of the residual graph neural network.

[0033] In this embodiment, common classification prediction networks CNN, ResNet, and LSTM are used to perform interpretable frequency stability prediction and compared with the above-mentioned interpretable frequency stability prediction method disclosed in the present invention. The simulation comparison experimental results are as follows: Figure 6 As shown, GCNII in the figure is the abbreviation of the residual graph neural network provided by the present invention.

[0034] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. An interpretable frequency-stable prediction method based on iterative optimization learning, characterized in that: The following steps are involved: S1. Perform N time domain simulations for N fault types of the target power system to obtain N groups of long-scale time domain simulation data; At the beginning of each time domain simulation, the fault type is set. In a long time scale window after the set fault occurs, N groups of long-scale time domain simulation data are obtained. Each group of long-scale time domain simulation data includes the active power data P and frequency data F of each monitoring node in the current long time scale window, and the minimum system frequency value is obtained from the frequency data F. ; S2. Construct the time series data of the model input; in each set of long-scale time domain simulation data, select a short observation time window after the fault occurs , the short-term observation time window Active power data within and frequency data As time series data, the time series data is integrated into a time series observation matrix; S3. Build model input spatial data; The impedance of each monitoring node in the target power system is obtained, and an electrical connection matrix is ​​obtained according to the relationship between the impedances of each monitoring node, and the electrical connection matrix is ​​used as spatial data; S4. Record the lowest frequency value in each set of long-scale time domain simulation data ,according to The frequency stability of the power system is judged by the size, and labels are generated for the time domain simulation data according to the judgment results. If the transient frequency minimum value Greater than the preset system frequency stability threshold Then it is judged that the power system frequency is stable; S5. Based on the time series data and its corresponding labels and spatial data, the residual graph neural network is used for in-depth learning to output the system frequency stability prediction information; S6. Debug and optimize the hyperparameters of the residual graph neural network model to obtain the hyperparameter combination with the highest prediction accuracy, and obtain the pre-trained residual graph neural network model; S7. Obtain test data, obtain prediction information corresponding to the test data through the pre-trained residual graph neural network model, and calculate the prediction accuracy of the prediction information corresponding to the test data, wherein the test data includes test time series data, test space data, and test label data; S8. According to the preset model prediction accuracy evaluation standard, the pre-trained residual graph neural network model that does not meet the evaluation standard is optimized again. If it meets the evaluation standard, it is used as the final residual graph neural network model and enters S9; S9. Perform interpretability analysis on the prediction results of the residual graph neural network model.

2. The interpretable frequency stability prediction method based on iterative optimization learning according to claim 1 is characterized in that: The specific content of the time series observation matrix in S2 is: Each row represents a short observation time window The active power data P and frequency data F of each monitoring node at any time within the short observation time window. Each column represents the active power data P and frequency data F of each monitoring node at any time within the short observation time window. The change track of active power data P or frequency data F within; Before integrating the time series observation matrix, the active power data P and frequency data F are standardized to ensure that they have a uniform dimension and a mean of zero.

3. The interpretable frequency stability prediction method based on iterative optimization learning according to claim 1 is characterized in that: The specific contents of obtaining the electrical contact matrix in S3 include: Obtain the impedance matrix of key bus nodes in the target power system: ， in , and Represent the number of rows and columns of the impedance matrix, Represents the total number of critical bus nodes in the system; Calculate the electrical connection matrix: ， In order to characterize the relative distance between each node, a normalization operation is performed: ， in, is the normalized electrical connection matrix.

4. The interpretable frequency stability prediction method based on iterative optimization learning according to claim 1 is characterized in that: The specific contents of S4 include: During the time domain simulation of each long time scale window, the minimum value of the system transient frequency is obtained , preset system frequency stability threshold ; when When , it is judged that there is a risk of frequency instability in the power system, and the predicted output system frequency stability result S=0; when When , the power system frequency is judged to be stable, and the predicted output system frequency stability result S=1.

5. The interpretable frequency stability prediction method based on iterative optimization learning according to claim 1 is characterized in that: The specific structure of the residual graph neural network in S5 includes: The residual graph neural network adopts a dual-channel input architecture, which is used to input time series data and spatial data respectively; The two-dimensional convolution module includes a convolution kernels, which are used to extract features from the input time series data and spatial data respectively to obtain a feature map; The residual module is used to mine the key features of the data according to a feature map through the initial residual and identity mapping. The key features form a one-dimensional vector and are connected to the fully connected layer to output the feature vector; The feature vector is activated by the softmax function to predict the frequency stability result of the output system.

6. The interpretable frequency stability prediction method based on iterative optimization learning according to claim 1 is characterized in that: The specific contents of S7 include: Obtain test time series data, test space data and test label data, input the test time series data and test space data into the pre-trained residual graph neural network model to obtain the corresponding prediction information, compare the prediction information with the test label data, and count the ratio of the number of correct prediction information Nr to the total number N, and then obtain the model prediction accuracy R=Nr / N×100%.

7. The interpretable frequency stability prediction method based on iterative optimization learning according to claim 6 is characterized in that: The specific contents of S8 include: If the accuracy R reaches or exceeds the preset accuracy threshold, it is determined that the current pre-trained residual graph neural network model is reliable in predicting power system frequency stability, and then enters S9; If the accuracy rate does not meet the standard, the pre-trained residual graph neural network model that does not meet the evaluation standard is optimized again until the accuracy rate R meets the preset accuracy rate threshold, and then enter S9; The steps of further optimization include: The test time series data, test space data and test label data obtained in S7 are combined with the time series data, space data and label data as new training data for the residual graph neural network model; Input the new training data into the residual graph neural network model, and gradually reduce the learning rate to perform optimization training again; After optimizing the training again, continue to judge the accuracy. If the accuracy R still does not reach 98.5%, repeat step S8. Otherwise, go to S9.

8. The interpretable frequency stability prediction method based on iterative optimization learning according to claim 1 is characterized in that: The specific contents of the interpretability analysis of the prediction results of the residual graph convolutional network model in S9 include: Assume that the input time series observation matrix is , among which The input is , the prediction value of the residual graph convolutional network model for the system frequency stability state is S, and the mean contribution of all inputs of the residual graph convolutional network model is , then the contribution value follows: ， in represent The contribution value of is used to reflect the contribution of the first The contribution of the first input to the final prediction value. The input is Active power data and frequency data at each moment; like , it means that in the time series data matrix The represented feature quantity improves the prediction value and has a positive effect; like , it means that in the time series data matrix The represented feature quantity leads to a decrease in the predicted value and has an adverse or no effect; By summing up the input contribution corresponding to each node, the total contribution of each monitoring node in the system is obtained, and the monitoring nodes are sorted according to the size of the total contribution. The bus nodes ranked before the preset ranking are the key bus nodes that have a significant impact on the system frequency evolution.

Images