Power grid transient voltage stability evaluation and explainability feature selection method and system

Abstract

The application relates to the technical field of power systems, and discloses a power grid transient voltage stability evaluation and explainable feature selection method and system, which comprises the following steps: acquiring key electrical quantity response data, performing feature selection and sample construction, obtaining a training sample set and a test sample set; training a prediction model, mining a mapping relationship between input features and voltage stability margin through the prediction model; evaluating the performance of the prediction model by using preset evaluation indexes; introducing an S-LIME method to construct an explainability analysis framework on the basis of the evaluated prediction model, combining explainability evaluation indexes, performing aggregated analysis on the prediction result of the prediction model from the feature and time double dimensions, identifying key features and main action time periods; optimizing the input feature set, retraining the model, forming a closed-loop optimization, and completing power grid transient voltage stability evaluation and explainable feature selection. The application guarantees the safe and stable operation of a power system.

Description

Technical Field

[0001] This invention relates to the field of power system technology, specifically to a method and system for assessing transient voltage stability of a power grid and selecting interpretable characteristics. Background Technology

[0002] Renewable energy and high-voltage direct current (HVDC) transmission technologies are rapidly developing and widely used in modern power systems. However, as synchronous generators are gradually replaced by power electronic interface power supplies, the reactive power regulation capability and dynamic voltage support capability of the system are constantly weakening, and voltage stability issues are becoming increasingly prominent. In AC / DC hybrid power grids, when the system suffers severe disturbances or is under critical operating conditions, transient voltage instability is easily triggered, and even voltage collapse may occur, posing a serious challenge to the safe and stable operation of large-scale power systems.

[0003] Currently, transient voltage stability analysis mainly employs time-domain simulation and direct methods. Time-domain simulation can accurately depict the dynamic evolution of power systems, and its evaluation results have high accuracy and reliability. However, its computational complexity is high, and simulation time is long, making it difficult to meet the needs of online analysis and real-time decision-making. Direct methods determine system stability by constructing transient energy functions, exhibiting high computational efficiency. However, in large-scale complex power systems, the nonlinear characteristics are significant, and constructing energy functions that accurately reflect actual dynamic behavior remains challenging. With the development of wide-area measurement technology and intelligent computing methods, data-driven transient voltage stability assessment methods have gradually become a research hotspot. These methods do not rely on precise physical modeling but instead achieve rapid system stability assessment by mining the mapping relationship between power system measurement data and steady-state conditions.

[0004] Existing transient voltage stability assessment methods based on response signals typically extract features from time-series response signals (such as the dynamic trajectory of bus voltage) during disturbances and employ shallow machine learning models such as decision trees and backpropagation neural networks for stability discrimination. Compared to traditional analytical methods, these methods have certain advantages in terms of computational efficiency and assessment accuracy. However, limited by the depth of the model structure and the ability to represent features, they are unable to fully characterize the complex nonlinear dynamic characteristics of high-dimensional power systems. To improve the model's ability to learn the spatiotemporal features of transient processes, deep learning models such as convolutional neural networks, residual networks, and long short-term memory networks have been introduced in recent years, which have improved the predictive performance of transient voltage stability assessment to some extent. However, with the continuous expansion of the power grid and the increasing number of devices, the feature dimensions involved in the dynamic process are rapidly increasing. Traditional deep learning models face problems such as severe feature redundancy and limited ability to extract effective discriminative information in high-dimensional data environments, making it difficult for them to maintain stable predictive performance under complex operating conditions.

[0005] In data-driven transient voltage stability assessment methods, feature selection and feature extraction are widely considered crucial to model performance, with their effectiveness directly impacting the quality of model input and the accuracy of system state representation. Existing feature selection methods mainly include wrapper methods, embedded methods, and filtering methods. Wrapper methods evaluate feature subsets through an iterative search process based on model performance, achieving optimal feature combinations. However, they incur significant computational overhead in high-dimensional power system applications and are prone to getting trapped in local optima. Embedded methods perform feature selection synchronously during model training, resulting in relatively high computational efficiency. However, their selection results are highly dependent on the model structure, and they are prone to overfitting when the feature dimension is large, thus weakening the model's generalization ability. Filtering methods select features based on the correlation between them and the target variable, offering high computational efficiency. However, they often ignore the correlation and coupling relationships between features, making it difficult to effectively eliminate redundant information. Therefore, existing feature selection methods still have shortcomings in fully integrating prior knowledge of the power system while simultaneously identifying key features and removing redundant information, hindering the efficient and accurate representation of system state. Summary of the Invention

[0006] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for evaluating transient voltage stability of power grids and selecting interpretable features. This solves the technical problem that the existing technology cannot achieve accurate identification of key features and effective removal of redundant information in the high-dimensional data environment of large-scale complex AC / DC hybrid power grids, resulting in the evaluation accuracy, efficiency and stability being unable to meet the needs of online analysis and real-time decision-making, and thus failing to fully guarantee the safe and stable operation of the power system.

[0007] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for assessing transient voltage stability of a power grid and selecting interpretable features, including: Acquire key electrical quantity response data after power grid disturbance, perform feature selection and sample construction on the key electrical quantity response data, and obtain a training sample set and a test sample set containing input feature set and output label; Based on the training sample set, a ResNet-BiLSTM-Attention voltage stability margin prediction model is trained, and the mapping relationship between input features and voltage stability margin is mined through the ResNet-BiLSTM-Attention voltage stability margin prediction model. The performance of the ResNet-BiLSTM-Attention voltage stability margin prediction model was evaluated using preset evaluation metrics. Based on the evaluated ResNet-BiLSTM-Attention voltage stability margin prediction model, the S-LIME method is introduced to construct an interpretability analysis framework. Combined with interpretability evaluation indicators, the prediction results of the ResNet-BiLSTM-Attention voltage stability margin prediction model are aggregated and analyzed from both feature and time dimensions to identify key features and main periods of action. Based on the identified key features and main periods of action, the input feature set is optimized, and the model is retrained to form a closed-loop optimization, thus completing the assessment of power grid transient voltage stability and the selection of interpretable features.

[0008] Preferably, the key electrical quantity response data includes the voltage amplitude, active power, and reactive power of the key bus after disturbance clearance; the input feature set is a 126-dimensional original feature set, constructed by selecting response quantities related to grid load level, induction motor load ratio, and new energy access ratio.

[0009] Preferably, the output label is the system voltage stability level ξmin, where ξmin is the minimum value of the large disturbance voltage stability level ξ of each monitored bus; the calculation formula for ξ is as follows:

[0010] In the formula: Indicates the fault clearing time; Indicates the simulation termination time; u(s) is the instantaneous value of the bus voltage, u N This is the rated voltage of the busbar; and This is the threshold value.

[0011] Preferably, the ResNet-BiLSTM-Attention voltage stability margin prediction model includes an input layer, a ResNet feature extraction layer, a pooling layer, a BiLSTM layer, an Attention layer, and an output layer; the ResNet feature extraction layer adopts a residual block structure and selects ELU as the activation function, and its output expression is as follows:

[0012] in, and These represent the output feature map and the input feature map, respectively; symbols This represents a one-dimensional convolution operation; and These are the kernel weights and bias terms, respectively. The element-wise ELU activation function is defined as follows: .

[0013] Furthermore, the BiLSTM layer captures the bidirectional dependency characteristics of time series through two propagation paths, forward and backward. Its hidden state expression is as follows:

[0014]

[0015]

[0016] in, For forward LSTM output, For output to the backward LSTM; The Attention layer mechanism borrows from how human attention works, assigning higher weight to key content and lower weight to less important content when processing information. The expression for the Attention layer mechanism is as follows:

[0017]

[0018]

[0019] In the formula, for The hidden unit at time step t, The vector represents the preceding and following time periods. For attention vectors, This is the output vector calculated using the attention mechanism; The model is randomly initialized at the beginning and continuously optimized and learned through backpropagation during training.

[0020] Preferably, the preset evaluation indicators include root mean square error, mean absolute error, mean absolute percentage error, and prediction accuracy. The expression for prediction accuracy is as follows:

[0021] In the formula, This is the result of normalized root mean square error.

[0022] Preferably, the implementation process of the S-LIME method includes data generation, data re-standardization and selection, CLT-based path selection, and construction of the final local proxy model; The data restandardization is achieved through Transform weighted regression into an ordinary least squares problem.

[0023] Preferably, the interpretability evaluation metrics include fidelity, comprehensibility, and robustness; fidelity is composed of comprehensiveness and sufficiency, comprehensibility is composed of sparsity and local fit, and robustness is used to reflect the stability of the interpretation under input perturbations.

[0024] Preferably, the aggregation analysis of the prediction results of the ResNet-BiLSTM-Attention voltage stability margin prediction model from both feature and time dimensions specifically includes, in the feature dimension, summarizing the feature contribution values ​​of different samples and time steps and calculating the global importance index; in the time dimension, accumulating statistical interpretation results according to time intervals and constructing a feature-time two-dimensional importance distribution.

[0025] Secondly, the present invention also provides a system for assessing power grid transient voltage stability and selecting interpretable features, comprising: The sample construction module is used to acquire key electrical quantity response data after power grid disturbance, perform feature selection and sample construction on the key electrical quantity response data, and obtain a training sample set and a test sample set containing input feature set and output label; The model training module is used to train the ResNet-BiLSTM-Attention voltage stability margin prediction model based on the training sample set, and to mine the mapping relationship between input features and voltage stability margin through the ResNet-BiLSTM-Attention voltage stability margin prediction model. The performance evaluation module is used to evaluate the performance of the ResNet-BiLSTM-Attention voltage stability margin prediction model using preset evaluation metrics. The interpretability analysis module is used to construct an interpretability analysis framework based on the evaluated ResNet-BiLSTM-Attention voltage stability margin prediction model by introducing the S-LIME method. Combined with interpretability evaluation indicators, it aggregates and analyzes the prediction results of the ResNet-BiLSTM-Attention voltage stability margin prediction model from both feature and time dimensions to identify key features and main periods of action. The closed-loop optimization module is used to optimize the input feature set based on the identified key features and main periods of action, and retrain the model to form a closed-loop optimization, thereby completing the assessment of power grid transient voltage stability and the selection of interpretable features.

[0026] Compared with the prior art, the present invention has the following beneficial technical effects: A method for assessing transient voltage stability in power grids and selecting interpretable features is proposed. This method initially reduces data redundancy by constructing a high-quality sample set using key electrical quantity response data. A ResNet-BiLSTM-Attention model is employed to accurately mine the mapping relationship between features and voltage stability margin, improving prediction accuracy and efficiency under high-dimensional data. Model parameters are optimized through multi-index evaluation to enhance prediction stability. An S-LIME framework is introduced to construct an interpretability framework, identifying key features and time periods from two dimensions, breaking the "black box" limitation of the model. Closed-loop optimization continuously refines the feature set, further improving model performance. This invention achieves accurate and efficient assessment of transient voltage stability in power grids and selects interpretable features, providing reliable support for the safe and stable operation of power grids and possessing significant engineering application value. Attached Figure Description

[0027] Figure 1 This is a flowchart of the method for evaluating transient voltage stability of the power grid and selecting interpretable features in an embodiment of the present invention; Figure 2 This is a schematic diagram of the ResNet-BiLSTM-Attention network structure in an embodiment of the present invention; Figure 3 This is a schematic diagram of the LSTM structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the S-LIME method structure in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the power grid transient stability assessment and interpretation in an embodiment of the present invention; Figure 6 This is a schematic diagram of the voltage collapse calculation example in an embodiment of the present invention; Figure 7 This is a comparison chart of prediction results from different models in this embodiment of the invention; Figure 8 A radar chart comparing the interpretability methods in this embodiment of the invention; Figure 9 This is a graph showing the result of ranking the importance of features in an embodiment of the present invention; Figure 10 This is the S-LIME weight matrix diagram in an embodiment of the present invention; Figure 11 This is a temporal clustering diagram of S-LIME feature weights in an embodiment of the present invention; Figure 12 This is a schematic diagram of the power grid transient voltage stability assessment and interpretability feature selection system in an embodiment of the present invention; In the diagram: 1. Sample construction module; 2. Model training module; 3. Performance evaluation module; 4. Interpretive analysis module; 5. Closed-loop optimization module. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] The purpose of this invention is to provide a method and system for evaluating transient voltage stability of power grids and selecting interpretable features, in order to solve the technical problem that existing technologies cannot achieve accurate identification of key features and effective removal of redundant information in the high-dimensional data environment of large-scale complex AC / DC hybrid power grids, resulting in evaluation accuracy, efficiency and stability that cannot meet the needs of online analysis and real-time decision-making, and thus cannot fully guarantee the safe and stable operation of the power system.

[0031] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1 This invention provides a method for assessing transient voltage stability in a power grid and selecting interpretability features, including: Step 1: Obtain key electrical quantity response data after power grid disturbance; perform feature selection and sample construction on the key electrical quantity response data to obtain a training sample set and a test sample set containing input feature set and output label; Specifically, the key electrical quantity response data includes the voltage amplitude, active power, and reactive power of the key bus after disturbance clearance; the input feature set is a 126-dimensional original feature set, which is constructed by selecting response quantities related to grid load level, induction motor load ratio, and new energy access ratio.

[0032] In this embodiment, factors such as grid load level, induction motor load ratio, and renewable energy integration ratio have a significant impact on grid voltage stability. When the grid suffers severe disturbances that lead to transient voltage instability in local areas, the dynamic response characteristics of various devices in the system can reflect the voltage stability state to a certain extent. Based on a comprehensive consideration of key influencing factors and data acquisition feasibility, this invention extracts response quantities such as voltage amplitude, active power, and reactive power that characterize the output characteristics of the devices after disturbance removal, and constructs a 126-dimensional feature set for the original input. In this embodiment, the output label is the system voltage stability level ξmin, where ξmin is the minimum value of the large disturbance voltage stability level ξ for each monitored bus. The criterion for large disturbance voltage stability is set as follows: during the transient period, the duration of the voltage drop at the system's central point being less than 0.75 pu is no more than 1 second, and the bus voltage should recover to 0.9 pu or above after the transient phase ends. Based on the voltage binary table (0.75 pu, 1s), this paper constructs an evaluation index that can quantify the system's large disturbance voltage stability level. as follows:

[0033] In the formula: Indicates the fault clearing time; Indicates the simulation termination time; u(s) is the instantaneous value of the bus voltage, u N This is the rated voltage of the busbar; and This is the threshold value. and The values ​​are set to 0.75 pu and 1 s. The minimum value of the large disturbance voltage stability level of each monitoring bus is taken as the system voltage stability level, and this value is used as the label for the sample.

[0034] Step 2: Based on the training sample set, train the ResNet-BiLSTM-Attention voltage stability margin prediction model, and use the ResNet-BiLSTM-Attention voltage stability margin prediction model to mine the mapping relationship between input features and voltage stability margin. Specifically, according to Figure 2 As shown, the ResNet-BiLSTM-Attention voltage stability margin prediction model includes an input layer, a ResNet feature extraction layer, a pooling layer, a BiLSTM layer, an Attention layer, and an output layer. In this embodiment, to accurately assess the grid voltage stability level under large disturbances, the input layer comprehensively considers the system operation mode and fault information, selecting key response feature time series after a certain period of time after fault clearance as the model input. The input can be represented as a time series feature matrix. ,in The length of the time series. The input feature dimension for a single time segment.

[0035] The control and operating variables of the critical bus within a certain time period after fault clearance are selected as model inputs, denoted as . .

[0036] The ResNet layer employs a residual block structure and uses the Exponential Linear Unit (ELU) as the activation function to enhance the model's training stability. Compared to the standard ReLU function, ELU exhibits a continuous and non-zero response in the negative input region, which helps alleviate neuron inactivation problems and feature distribution shifts, thereby improving the ability to model the system's inherent nonlinear dynamic characteristics. Therefore, the output of a ResNet block can be expressed as:

[0037] in, and These represent the output feature map and the input feature map, respectively; symbols This represents a one-dimensional convolution operation; and These are the kernel weights and bias terms, respectively. The element-wise ELU activation function is defined as follows: .

[0038] according to Figure 3 As shown, the LSTM units in the BiLSTM layer regulate the information flow through forget gates, input gates, and output gates, effectively alleviating the gradient vanishing problem. However, standard LSTM can only model time series in one direction, and its ability to characterize the dependencies between sequences is limited. To capture the bidirectional dependencies of sequences, the Bidirectional Long Short-Term Memory (BiLSTM) network introduces two propagation paths, forward and backward, and concatenates their hidden states to form the final feature representation. The mathematical model of the BiLSTM layer can be expressed as follows.

[0039]

[0040]

[0041] .

[0042] The attention mechanism borrows from how human attention works, assigning higher weight to key content and lower weight to less important content when processing information.

[0043] In the Attention mechanism:

[0044]

[0045]

[0046] In the formula, for The hidden unit at time step t, The vector represents the preceding and following time periods. For attention vectors, This is the output vector calculated using the attention mechanism. The model is randomly initialized at the beginning and continuously optimized and learned through backpropagation during training.

[0047] The output layer consists of fully connected layers, using the output vector of the Attention mechanism as input to calculate the predicted voltage stability level of each load bus. The model output... As shown below:

[0048] In the formula: and These are the output layer weight matrix and bias.

[0049] Step 3: Evaluate the performance of the ResNet-BiLSTM-Attention voltage stability margin prediction model using preset evaluation metrics; Specifically, the preset evaluation indicators include root mean square error, mean absolute error, mean absolute percentage error, and prediction accuracy. In this embodiment, the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are selected as model performance evaluation indicators. The calculation formulas for each indicator are as follows:

[0050]

[0051]

[0052]

[0053] In the formula: Indicates the total number of samples in the test set; subscript Indicates the first One test sample; Indicates the predicted value; This represents the true value. Each evaluation index reflects the overall error level of the regression prediction; the lower the index value, the better the model's predictive performance.

[0054] To further evaluate the accuracy of the prediction network in predicting steady-state levels, based on Performance evaluation metrics for constructing prediction networks: prediction accuracy The indicator is expressed in the following form:

[0055] Step 4: Based on the evaluated ResNet-BiLSTM-Attention voltage stability margin prediction model, the S-LIME method is introduced to construct an interpretability analysis framework. Combined with interpretability evaluation indicators, the prediction results of the ResNet-BiLSTM-Attention voltage stability margin prediction model are aggregated and analyzed from both feature and time dimensions to identify key features and main periods of action. In this embodiment, to reveal the decision-making basis of the deep learning model, an interpretability analysis framework is constructed by introducing S-LIME on the ResNet-BiLSTM-Attention model. Subsequently, interpretability evaluation metrics are established, and the interpretation results are aggregated and analyzed from two dimensions: features and time, to identify key features and the main periods of influence.

[0056] This invention employs the S-LIME method to improve the stability of neighborhood weighting, variable selection, and random perturbation processes, thereby obtaining a more statistically reliable local interpretation. For example... Figure 4 As shown, the overall framework of S-LIME can be decomposed into four steps: data generation, restandardization processing, CLT-based path selection, and construction of local proxy models. The specific process is as follows.

[0057] Data generation: Around the target sample Perturbation samples are randomly generated within its neighborhood, and the corresponding results are obtained from the original model. This forms the pseudo-sample set required for fitting the local proxy model. .

[0058] Data restandardization and selection: First, the perturbed weighted samples and the predicted results are recalibrated to convert the weighted least squares equivalent into an unweighted form:

[0059] in These are pseudo-samples after restandardization, obtained by multiplying the sample by... This can transform weighted regression into an ordinary least squares problem, thereby reducing the ill-conditioned effects caused by kernel weights. Based on restandardized data, S-LIME performs LASSO through multiple subsample extractions:

[0060] in, For the first The index set for secondary subsample extraction. Then, the selection probability of each feature is estimated through stable selection:

[0061] Based on this, a stable feature set is obtained:

[0062] in, A threshold is selected for stability, used to remove unstable noise features.

[0063] CLT path selection: Although stable selection reduces the impact of random disturbances, the order in which variables enter the LASSO path still needs to be statistically distinguishable. Therefore, S-LIME constructs a path significance test based on the multivariate central limit theorem. Let the current residual be... Candidate features The empirical correlation quantity with the residual is:

[0064] According to the central limit theorem, we have:

[0065] in Let be the covariance matrix of the residuals and features. When the confidence level is... To ensure that the order in which variables enter the path is statistically distinguishable, the following conditions must be met:

[0066] in For the standard normal distribution Quantiles. If this condition is not met, it indicates that the perturbation sample size is insufficient. S-LIME then adaptively updates the sample size accordingly:

[0067] in For the present Standardized statistics. Final local proxy model: Finally, after satisfying the stability and significance criteria, S-LIME is found in the stable feature subset. Fit the final local proxy model:

[0068] in, This represents the projection of the samples onto the stable feature dimension. The model provides a highly consistent, statistically guaranteed local interpretation under repeated perturbations.

[0069] In summary, S-LIME, based on data restandardization, introduces a path selection mechanism that follows the central limit theorem, thereby systematically correcting the shortcomings of traditional LIME in terms of kernel weight sensitivity, variable selection inconsistency, and random perturbation instability, and constructing a more stable and statistically reliable local interpretability framework.

[0070] This embodiment measures the quality of the explanation method from three perspectives: Fidelity, Understandability, and Robustness. The definitions of each indicator are as follows: Loyalty (Fidelity) Fidelity is used to characterize the consistency between the explained results and the model's actual decision-making behavior. It consists of comprehensiveness and sufficiency. The metric is defined as follows:

[0071]

[0072]

[0073] in, This is the set of key features identified by the interpretation method. This is the model's raw output for the samples. To remove feature set The output of the post-model, This means retaining only the feature set. The output of the post-model. In loyalty index evaluation, it is often taken as... The higher the fidelity, the more accurately the interpretation results reflect the model's internal decision-making process.

[0074] Understandability Understandability measures the simplicity and readability of the explanation results, and it is composed of sparsity and local accuracy. This metric is defined as follows:

[0075]

[0076] in, This indicates the number of features selected in the interpretation. The total number of features input to the model; LocalAccuracy This indicates that the model interprets the input samples. The degree of local fit to the original model in the vicinity. A commonly used metric for interpretability is... The higher the comprehensibility, the more concise and clear the explanation, and the easier it is to understand and use.

[0077] Robustness Robustness reflects the stability of the interpretation under input disturbances. A higher robustness index indicates greater stability of the interpretation under these disturbances. The index is defined as follows:

[0078] in, Indicates the application applied to the input. The disturbance, the amplitude of the disturbance satisfies , The upper bound of acceptable perturbation. This indicates the explanation result. It serves as a distance metric between interpretations, reflecting the amount of change in interpretations due to disturbances.

[0079] The three metrics mentioned above are used to quantitatively evaluate the interpretability and transparency of interpretable methods for black-box functions. Better performance on these metrics indicates that the results of the interpretable method are more accurate, easier to understand, and more robust.

[0080] In this embodiment, a feature-time two-layer aggregation framework was constructed. In terms of features, the contribution values ​​of different samples and time steps are summarized according to feature names, and the global importance index is calculated by mean or weighted summation to identify the key features that have the most significant impact on the model output, providing a quantitative basis for feature selection and model simplification.

[0081] In the time dimension, the interpretation results are cumulatively statistically analyzed according to time intervals to construct a two-dimensional feature-time importance distribution, which is used to locate the main role stages of each feature in the transient process. Through the above two-dimensional aggregation analysis, this paper achieves an interpretable characterization of the prediction mechanism of deep models, and provides a unified quantitative analysis framework for key feature extraction and key period identification.

[0082] Step 5: Optimize the input feature set based on the identified key features and main periods of action, and retrain the model to form a closed-loop optimization, thus completing the power grid transient voltage stability assessment and interpretability feature selection.

[0083] In this embodiment, the response-driven grid transient voltage stability assessment and interpretability analysis process is as follows: Figure 5 As shown, it includes three stages: offline modeling, online evaluation, and interpretation and optimization.

[0084] In the offline modeling phase, voltage stability margin is calculated based on electrical response data from multiple scenarios. A large-disturbance voltage stability assessment sample set containing input features and labels is constructed and divided into training and test sets. By extracting transient voltage response features, a ResNet-BiLSTM-Attention network model is designed and trained. The network structure and parameters are optimized based on prediction performance to obtain a voltage stability assessment model with good predictive capabilities.

[0085] During the online evaluation phase, when a disturbance occurs in the system, WAMS collects key measurement sequences in real time and inputs them into the trained model to achieve rapid prediction of voltage stability margin, providing a basis for operational evaluation and control decisions.

[0086] In the interpretation and optimization phase, the S-LIME method is introduced to perform feature-time dual-dimensional analysis on the model prediction results, identify the key response features that play a dominant role in stability margin prediction and their duration of action, and use this information to filter and optimize the input feature set. The model is then retrained in the offline phase, and finally a voltage stability assessment model with both prediction accuracy and physical interpretability is constructed.

[0087] Example 1 This embodiment uses the voltage collapse engineering case study released by the China Electric Power Research Institute as the simulation system for analysis. The system structure is as follows: Figure 6 As shown, this example is a DC receiving-end system with renewable energy sources. When an N-2 fault occurs on the AC line of the system, voltage instability may occur. The system used in this example is based on a 500kV grid architecture, containing 100 nodes, with a base capacity of 100MW. As shown in the figure, Bus represents a 500kV bus, Gen-T represents conventional generator units, Gen-P and Gen-W represent photovoltaic power plants and wind farms, respectively, where " " is the component number. In this system, the installed capacity of renewable energy units is 2400MW, the installed capacity of conventional units is 6300MW, and the system is equipped with a DC transmission line with a rated input power of 800MW. Buses Bus1 to Bus21 are set as load buses, and the load model adopts a comprehensive load model composed of static loads and induction motors in a certain proportion.

[0088] This invention comprehensively considers the influencing factors of voltage stability level and the principle of convenient online data acquisition. From multiple perspectives such as conventional units, new energy units, load and line power flow, it selects relevant response information that can characterize the power output characteristics of the equipment after disturbance removal, and constructs an original input feature set containing 126 features. To fully cover the response characteristics of the actual power grid under different operating scenarios, this invention utilizes BPA simulation software to generate the required dataset. Under different load levels, various basic operating modes are constructed by adjusting the output ratio of conventional and renewable energy units, the transmission power of the DC system, and the proportion of induction motors in the overall load. Based on these operating modes, multiple fault scenarios are further designed. N-2 type faults are set on four 500 kV lines including double-circuit lines, with fault locations at 2%, 50%, and 98% of the AC line length, respectively. The fault clearing time is set from 0.1 s to 0.3 s after the disturbance occurs, with a step size of 0.05 s, for a total of five clearing durations. Finally, a total of 5400 sample data points are generated.

[0089] Using 5400 generated data samples, 80% were allocated to the training set and the remaining 20% ​​to the test set. Taking the key response features after fault clearance as input and the system voltage stability level as output, a ResNet-BiLSTM-Attention network was trained using the training set to evaluate the voltage stability level of the power grid under large disturbances.

[0090] To compare the prediction performance of the proposed ResNet-BiLSTM-Attention model with classic network models such as BP, BiLSTM, ResNet, and ResNet-BiLSTM, the prediction results of each model were evaluated, and the calculated metrics are shown in Table 1. Figure 7 Further observation reveals the distribution of absolute prediction errors for different models on the samples. Among them, the ResNet-BiLSTM-Attention model exhibits the most concentrated and smallest overall error, further validating its advantage in prediction accuracy.

[0091] Table 1. Prediction performance of different network models.

[0092] It can be seen that the proposed network has good predictive performance, and the regression prediction results are generally more consistent with the true values, indicating that the model has strong generalization performance.

[0093] To accurately identify the dominant factors affecting the model's prediction results, this embodiment selects the interpretability model with the best overall performance based on multidimensional evaluation indicators, and uses this model as a basis to deeply extract and quantify the key response characteristics of the system's transient voltage stability.

[0094] First, the prediction results of typical fault cases are selected and compared and analyzed using various interpretability methods, such as... Figure 8 As shown in Table 2, there are significant differences in the performance of each method in terms of fidelity, understandability, and robustness. The S-LIME method exhibits high stability in both fidelity and robustness, generating more consistent and reliable feature contribution evaluation results compared to LIME and A-LIME. Furthermore, although its computation time is slightly longer than LIME, it is still far shorter than SHAP, balancing computational efficiency with interpretation quality. Normalization of each indicator and the creation of a radar chart are shown in Fig. 8. It can be seen that S-LIME performs relatively well across all dimensions, demonstrating superior overall performance.

[0095] Table 2. Effects of different interpretation methods

[0096] Based on the S-LIME method, the local explanatory weights of all samples are standardized and their absolute value statistics are calculated according to the feature dimension, thereby obtaining the global feature importance ranking. Figure 9 The S-LIME weight distribution of the top 10 features is presented, and the statistical characteristics of the explanatory power of each feature are shown in the form of a cloud and rain plot.

[0097] The ranking of feature importance in the figure shows that the model is most sensitive to reactive power-related features when determining the system's transient voltage stability. Among them, Q-B19, Q-P6, and Q-T6 are the most important, indicating that the reactive power consumption of the overall load, the reactive power regulation capability of the photovoltaic power station, and the reactive power support of conventional units have a significant impact on the model's prediction results. The model is highly sensitive to the system's reactive power supply and demand balance, paying particular attention to the reactive power demand on the load side and the reactive power support capability on the generation side.

[0098] Overall, the distribution of feature importance exhibits a clear spatial topological correlation: load nodes and generator variables closer to weak points in the grid are significantly more important than those in other areas. This phenomenon indicates that the model can not only identify the key role of reactive power in voltage support, but also effectively learn the coupling relationship between the system's electrical topology and dynamic response, thus demonstrating good interpretability and physical consistency.

[0099] After obtaining the global feature importance distribution of the model, this paper conducts a local interpretability analysis of the time-series features to further reveal the model's discrimination criteria at the individual sample level and the response differences under different operating conditions. Based on the S-LIME method, the sensitivity distribution of each time-series feature at different times after disturbance removal is calculated. To demonstrate the model's response characteristics under typical operating conditions, this paper randomly selects a representative scenario from both unstable and stable samples. The specific operating scenarios are as follows.

[0100] Table 3 Operating Mode

[0101] The time sensitivity analysis results of the two typical samples selected above are as follows: Figure 10 As shown, the color and size of the dots represent the direction and magnitude of the local slope, respectively, thus reflecting the importance contribution of each feature to the model output at different times.

[0102] As shown in the figure, the feature sensitivity of both types of samples is mainly concentrated in the very early stage after fault clearance (approximately 0–0.03 s). In both samples, Q-B19 exhibits a significant negative slope, indicating that the load at node 19 has a prominent impact on system voltage stability and is a relatively weak receiving-end node in the system. Furthermore, features such as Q-P6 and Q-T6 show high-amplitude positive slopes, indicating that reactive power support in this region plays a positive role in restoring the stability margin after fault clearance. This result is consistent with the analysis conclusions, further illustrating the important role of key nodes and regional reactive power support in the transient voltage recovery process. As time progresses, the slope amplitudes of each feature in both types of samples rapidly decrease and approach zero, indicating that the model's dependence on time-series features decreases significantly after 0.03 s, and its judgment is mainly determined by the instantaneous changes in key features after the fault.

[0103] To further analyze the overall response pattern of the model over time, this invention aggregated the S-LIME feature contribution values ​​of 200 samples in the test set over time and calculated the average contribution amplitude at each fault clearing time. To remove noisy data with minimal impact on model prediction, all contribution values ​​with an absolute value less than 0.005 were filtered out. The calculation results are as follows: Figure 11 As shown in the figure, the horizontal axis represents the fault clearing time, and the vertical axis represents the average S-LIME weight, which reflects the sensitivity of the model to the input features at different time points.

[0104] As shown in the figure, the voltage stability margin prediction model exhibits typical pre-centralization characteristics in its time-dimensional response. The very early period after fault clearance shows a significantly higher average contribution value, indicating that the model is most sensitive to system dynamic changes during this stage and primarily relies on transient information at this time to determine the stability margin. As time progresses, the average contribution amplitude rapidly decays and stabilizes near zero after approximately 0.03 s, suggesting that the model has limited focus on the subsequent longer steady-state process. Overall, the discrimination logic of the voltage stability assessment model is clearly biased towards capturing the system response during the short-term transition phase after fault clearance.

[0105] Based on the generated data samples, several input methods were constructed, including a subset of key features, short-time series input, and a combination of key features and short-time series input. These were then used as inputs for model training and validation. Experimental results are shown in the table. Different feature extraction methods effectively compressed the input dimensionality while also improving model prediction performance. Feature selection effectively removed redundant information, while time step truncation further highlighted the contribution of early dynamic features.

[0106] From an overall performance perspective, the combined approach of key feature extraction and time step truncation demonstrates the most significant advantages. It significantly shortens prediction time while maintaining or even improving prediction accuracy, indicating that the original input has high redundancy in both feature and temporal dimensions. In particular, the preservation of the early post-fault time period allows the model to focus more intently on the most discriminative temporal information, thereby improving the response efficiency and generalization ability of the ResNet-BiLSTM-Attention model.

[0107] Table 4. Impact of different extraction methods on prediction results

[0108] In summary, the ResNet-BiLSTM-Attention model constructed in this invention fully integrates the spatial feature extraction capabilities of ResNet, the temporal modeling advantages of BiLSTM, and the key feature focusing characteristics of the attention mechanism. It can effectively characterize the mapping relationship between voltage transient stability level and system operating state, achieving accurate assessment of transient voltage stability. The feature selection method based on S-LIME proposed in this invention can quantify the importance and influence direction of each input feature in an interpretable manner, effectively eliminating redundant information and retaining key features, providing a reliable basis for constructing a lighter and more generalizable voltage stability assessment model. Based on S-LIME, this invention conducts an in-depth analysis of the variation law of feature contribution at different time points, revealing the role characteristics of key features in the temporal dimension from the perspective of interpretability, clarifying the specific dependence of the model on temporal information in voltage stability assessment, and clearly characterizing the role mechanism of key features from the temporal dimension, thus improving the interpretability and credibility of the model.

[0109] Example 2 according to Figure 12 As shown, this embodiment also provides a power grid transient voltage stability assessment and interpretability feature selection system, including: Sample construction module 1 is used to acquire key electrical quantity response data after power grid disturbance, perform feature selection and sample construction on the key electrical quantity response data, and obtain a training sample set and a test sample set containing input feature set and output label; Model training module 2 is used to train a ResNet-BiLSTM-Attention voltage stability margin prediction model based on the training sample set, and to mine the mapping relationship between input features and voltage stability margin through the ResNet-BiLSTM-Attention voltage stability margin prediction model. Performance evaluation module 3 is used to evaluate the performance of the ResNet-BiLSTM-Attention voltage stability margin prediction model using preset evaluation metrics. The interpretability analysis module 4 is used to construct an interpretability analysis framework by introducing the S-LIME method on the basis of the evaluated ResNet-BiLSTM-Attention voltage stability margin prediction model. Combined with interpretability evaluation indicators, it aggregates and analyzes the prediction results of the ResNet-BiLSTM-Attention voltage stability margin prediction model from the dual dimensions of features and time, and identifies key features and main periods of action. The closed-loop optimization module 5 is used to optimize the input feature set based on the identified key features and main periods of action, and retrain the model to form a closed-loop optimization, thereby completing the assessment of power grid transient voltage stability and the selection of interpretable features.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for assessing transient voltage stability and selecting interpretability features in a power grid, characterized in that, include: Acquire key electrical quantity response data after power grid disturbance, perform feature selection and sample construction on the key electrical quantity response data, and obtain a training sample set and a test sample set containing input feature set and output label; Based on the training sample set, a ResNet-BiLSTM-Attention voltage stability margin prediction model is trained, and the mapping relationship between input features and voltage stability margin is mined through the ResNet-BiLSTM-Attention voltage stability margin prediction model. The performance of the ResNet-BiLSTM-Attention voltage stability margin prediction model was evaluated using preset evaluation metrics. Based on the evaluated ResNet-BiLSTM-Attention voltage stability margin prediction model, the S-LIME method is introduced to construct an interpretability analysis framework. Combined with interpretability evaluation indicators, the prediction results of the ResNet-BiLSTM-Attention voltage stability margin prediction model are aggregated and analyzed from both feature and time dimensions to identify key features and main periods of action. Based on the identified key features and main periods of action, the input feature set is optimized, and the model is retrained to form a closed-loop optimization, thus completing the assessment of power grid transient voltage stability and the selection of interpretable features.

2. The method for evaluating transient voltage stability and selecting interpretability features of a power grid according to claim 1, characterized in that, The key electrical quantity response data includes the voltage amplitude, active power, and reactive power of the key bus after disturbance clearance; the input feature set is a 126-dimensional original feature set, constructed by selecting response quantities related to grid load level, induction motor load ratio, and new energy access ratio.

3. The method for evaluating transient voltage stability and selecting interpretability features of a power grid according to claim 1, characterized in that, The output label represents the system voltage stability level ξmin, where ξmin is the minimum value of the large disturbance voltage stability level ξ for each monitored bus; the formula for calculating ξ is as follows: In the formula: Indicates the fault clearing time; Indicates the simulation termination time; u(s) is the instantaneous value of the bus voltage, u N This is the rated voltage of the busbar; and This is the threshold value.

4. The method for evaluating transient voltage stability and selecting interpretability features of a power grid according to claim 1, characterized in that, The ResNet-BiLSTM-Attention voltage stability margin prediction model includes an input layer, a ResNet feature extraction layer, a pooling layer, a BiLSTM layer, an attention layer, and an output layer. The ResNet feature extraction layer adopts a residual block structure and uses ELU as the activation function. Its output expression is as follows: in, and These represent the output feature map and the input feature map, respectively; symbols This represents a one-dimensional convolution operation; and These are the kernel weights and bias terms, respectively. The element-wise ELU activation function is defined as follows: 。 5. The method for evaluating transient voltage stability and selecting interpretability features of a power grid according to claim 4, characterized in that, The BiLSTM layer captures the bidirectional dependency characteristics of time series through two propagation paths, forward and backward. Its hidden state expression is as follows: in, For forward LSTM output, For output to the backward LSTM; The Attention layer mechanism borrows from how human attention works, assigning higher weight to key content and lower weight to less important content when processing information. The expression for the Attention layer mechanism is as follows: In the formula, for The hidden unit at time step t, The vector represents the preceding and following time periods. For attention vectors, This is the output vector calculated using the attention mechanism; The model is randomly initialized at the beginning and continuously optimized and learned through backpropagation during training.

6. The method for evaluating transient voltage stability and selecting interpretability features of a power grid according to claim 1, characterized in that, The preset evaluation indicators include root mean square error, mean absolute error, mean absolute percentage error, and prediction accuracy. The expression for prediction accuracy is as follows: In the formula, This is the result of normalized root mean square error.

7. The method for evaluating transient voltage stability and selecting interpretability features of a power grid according to claim 1, characterized in that, The implementation process of the S-LIME method includes data generation, data re-standardization and selection, CLT-based path selection, and construction of the final local proxy model. The data restandardization is achieved through Transform weighted regression into an ordinary least squares problem.

8. The method for evaluating transient voltage stability and selecting interpretability features of a power grid according to claim 1, characterized in that, The interpretability evaluation metrics include fidelity, comprehensibility, and robustness; fidelity is composed of comprehensiveness and sufficiency, comprehensibility is composed of sparsity and local fit, and robustness is used to reflect the stability of the interpretation under input perturbations.

9. The method for evaluating transient voltage stability and selecting interpretability features of a power grid according to claim 1, characterized in that, The aggregation analysis of the prediction results of the ResNet-BiLSTM-Attention voltage stability margin prediction model from both feature and time dimensions includes, in the feature dimension, summarizing the feature contribution values ​​of different samples and time steps and calculating the global importance index; in the time dimension, accumulating statistical interpretation of the results according to time intervals and constructing a feature-time two-dimensional importance distribution.

10. A system for evaluating transient voltage stability and selecting interpretability features in a power grid, characterized in that, include: The sample construction module is used to acquire key electrical quantity response data after power grid disturbance, perform feature selection and sample construction on the key electrical quantity response data, and obtain a training sample set and a test sample set containing input feature set and output label; The model training module is used to train the ResNet-BiLSTM-Attention voltage stability margin prediction model based on the training sample set, and to mine the mapping relationship between input features and voltage stability margin through the ResNet-BiLSTM-Attention voltage stability margin prediction model. The performance evaluation module is used to evaluate the performance of the ResNet-BiLSTM-Attention voltage stability margin prediction model using preset evaluation metrics. The interpretability analysis module is used to construct an interpretability analysis framework based on the evaluated ResNet-BiLSTM-Attention voltage stability margin prediction model by introducing the S-LIME method. Combined with interpretability evaluation indicators, it aggregates and analyzes the prediction results of the ResNet-BiLSTM-Attention voltage stability margin prediction model from both feature and time dimensions to identify key features and main periods of action. The closed-loop optimization module is used to optimize the input feature set based on the identified key features and main periods of action, and retrain the model to form a closed-loop optimization, thereby completing the assessment of power grid transient voltage stability and the selection of interpretable features.

Images