Model building method, evaluation method and system for power grid timing stability evaluation with high precision and strong interpretability

Abstract

This application provides a model building method, evaluation method, and system for power grid time-series stability assessment with high accuracy and strong interpretability. The model building method includes: initializing an Informer encoder; initializing a multivariate decision tree improved model; acquiring multidimensional time-series data after power grid disturbances and simultaneous system static feature data, processing them separately to obtain fixed-dimensional transient time-series feature vectors and static feature vectors; concatenating and fusing the fixed-dimensional transient time-series feature vectors and static feature vectors using a dimension superposition method to form a fused feature vector; inputting the fused feature vector into the initialized multivariate decision tree improved model for iterative training to obtain the power grid time-series stability assessment results and decision path results. This application generates a highly accurate power grid time-series stability assessment tool with a trustworthy decision-making process, providing efficient and accurate guidance for power grid time-series stability assessment. It is applicable to the technical field of transient stability assessment.

Description

Technical Field

[0001] This application relates to the technical field of transient stability assessment, specifically to a model establishment method, assessment method, and system for power grid time-series stability assessment with high accuracy and strong interpretability. Background Technology

[0002] The rapid development of energy transition has driven the continuous expansion of the power grid. With the integration of a high proportion of power electronic equipment and the gradual increase in the penetration rate of new energy sources, large-scale grid connection of new energy power generation, represented by wind and solar power, has become the norm. However, the large-scale integration of power electronic equipment and renewable energy sources such as wind and solar power has profoundly changed the operating characteristics of the power grid, making it more volatile, random, and less inertial. If appropriate decisions cannot be made in a timely manner to address transient faults and effectively intervene in disturbances after they occur in the power grid, the transient stability of the system may be disrupted, even developing into a subsequent cascading fault, leading to large-scale power outages and causing serious consequences. Therefore, transient voltage stability, power angle stability, and other time-series stability issues have become increasingly prominent, posing a major threat to the safe operation of the power system.

[0003] In existing technologies, traditional TSA methods mainly include time-domain simulation and transient energy function methods. Time-domain simulation establishes a detailed mathematical model of the power system and judges stability based on the generator power angle change trajectory. While this method offers high accuracy, it involves enormous computational load and long response time, making it difficult to meet the needs of online real-time assessment. The transient energy function method analyzes system stability based on energy functions and assesses stability by constructing transient energy indices. While this method is computationally fast, its overly simplified model leads to insufficient accuracy and poor adaptability to large-scale power grids. Consequently, traditional TSA methods often struggle to balance accuracy and speed when dealing with complex operating conditions due to their limited feature dimensions and insufficient dynamic adaptability, making them unsuitable for the complex operating conditions of modern power grids, such as high proportions of renewable energy integration and frequent topology changes.

[0004] Existing machine learning evaluation models, on the one hand, include deep learning models such as graph convolutional networks and long short-term memory networks. While these models offer high accuracy, their end-to-end "black box" operation makes their decision-making logic obscure and difficult to understand, making it hard to gain the complete trust of dispatching and operation personnel. On the other hand, while traditional decision trees offer strong interpretability, their feature representation capabilities are limited when dealing with high-dimensional, highly time-series power grid dynamic data, often failing to achieve ideal evaluation accuracy. Existing hybrid models are often just simple integrations of model results or only provide a global transient voltage stability situation, resulting in relatively coarse and limited information. The results obtained are relatively conservative and inaccurate, requiring more refined transient stability evaluation models and criteria.

[0005] Therefore, developing an accurate, efficient, and reliable power grid timing stability assessment tool is a key line of defense and an urgent need to prevent cascading failures and large-scale power outages. Summary of the Invention

[0006] To address the aforementioned technical deficiencies, this application provides a method for establishing a model, an evaluation method, and a system for assessing power grid time-series stability with high accuracy and strong interpretability.

[0007] According to the first aspect of this application, a method for establishing a model for power grid time-series stability assessment with high accuracy and strong interpretability is provided, comprising:

[0008] The Informer encoder is pre-trained using the training set in the pre-trained total sample set. The classification head of the encoder after pre-training is removed, and the encoder weights are fixed to obtain the initialized Informer encoder.

[0009] The pre-trained total sample set is input into the initialized Informer encoder for processing to obtain a fixed-dimensional transient temporal feature vector training set;

[0010] By introducing a hierarchical structure and a cost-sensitive learning mechanism into the basic multivariate decision tree model, an improved multivariate decision tree model is obtained.

[0011] The improved multivariate decision tree model is trained using a fixed-dimensional transient time-series feature vector training set to obtain the initialized improved multivariate decision tree model.

[0012] After acquiring multidimensional time-series data following power grid disturbance and simultaneous system static feature data, the multidimensional time-series data is input into the initialized Informer encoder for processing to obtain a fixed-dimensional transient time-series feature vector. Rough set attribute reduction is performed on the system static feature data to obtain a static feature vector.

[0013] A dimension stacking method is used to concatenate and fuse the transient temporal feature vector with the static feature vector to form a fused feature vector;

[0014] The fused feature vectors are input into the initialized multivariate decision tree improvement model for iterative training. Based on the corresponding power grid time-series stability assessment results and decision path results, the optimal model is obtained and used as the model for power grid time-series stability assessment.

[0015] Preferably, the step of introducing a hierarchical structure and a cost-sensitive learning mechanism into the basic multivariate decision tree model to obtain an improved multivariate decision tree model; and training the improved multivariate decision tree model using a fixed-dimensional transient temporal feature vector training set to obtain an initialized improved multivariate decision tree model, specifically including:

[0016] A multivariate decision tree base model is constructed by replacing the internal nodes of the decision tree with support vector machines.

[0017] Furthermore, by introducing a hierarchical structure and a cost-sensitive learning mechanism into the basic multivariate decision tree model, an improved multivariate decision tree model based on an iterative weighted strategy is obtained.

[0018] Based on a hierarchical structure and cost-sensitive learning mechanism, a multivariate decision tree improvement model is trained serially using a training set of fixed-dimensional transient temporal feature vectors obtained by initializing the Informer encoder. During training, an iterative weighting strategy is applied to dynamically adjust sample weights, save relevant hyperparameters, and optimize and evaluate the model performance under different hyperparameter combinations. The iterative process is repeated until the model performance converges, and finally, the initialized multivariate decision tree improvement model is obtained.

[0019] Preferably, the step of using a fixed-dimensional transient temporal feature vector training set obtained by initializing the Informer encoder to perform serial stacking training on the multivariate decision tree improvement model based on hierarchical structure and cost-sensitive learning mechanism specifically includes:

[0020] By introducing multiple cost-sensitive decision trees as weak classifiers and concatenating them sequentially, a cost matrix is ​​defined, a cost ratio is set, and the weights of all samples are set to the initial weights.

[0021] Based on the current weight distribution, the first weak classifier is trained by calculating the classification error rate and weights, and the prediction result is output.

[0022] Adjust the sample weights so that the weight update result of the previous weak classifier after training is used as the basis for adjusting the input weights of the next weak classifier. Repeat the iterative process of superimposed error correction and focusing until the accuracy requirement is met.

[0023] The predictions of all weak classifiers are weighted and combined according to their classification accuracy to form a strong classifier.

[0024] Preferably, during the training process, an iterative weighting strategy is applied to dynamically adjust sample weights, save relevant hyperparameters, optimize and debug the model, evaluate the model performance under different hyperparameter combinations, and repeat the iterative process until the model performance converges. Specifically, this includes:

[0025] Assign the currently set weight value to each sample of the same category to form the current probability distribution;

[0026] Set weight update rules, dynamically update base classifier weights and adaptively perform iterative weighting through a loss function, minimize the cumulative cost of the classifier, and update sample weights.

[0027] Collect feedback data, analyze the direction of weight adjustment, save relevant hyperparameters, and check whether the performance of parameters and weights converges by debugging, optimizing and evaluating the model performance under different hyperparameter combinations. If converged, obtain the updated sample weights. If not converged, repeat the iterative process until the model performance converges.

[0028] Preferably, the step of setting weight update rules, which involves dynamically updating the base classifier weights and adaptively iteratively weighting through a loss function to minimize the cumulative cost of the classifier and update the sample weights, specifically includes: setting weight update rules, training with the currently set sample weights, identifying high-cost misclassified samples through a cost function, dynamically adjusting the sample weights based on the error cost, increasing the weights of misclassified samples, decreasing the weights of correctly classified samples, repeating the iterative process to minimize the cumulative cost of the classifier until the model weights converge;

[0029] The weight update rule is set as follows:

[0030] Let the k-th level classifier be ,sample The weight is ,but:

[0031]

[0032] In the formula: Represents the cost function, This represents the weight adjustment factor for level k;

[0033] The cumulative cost of minimizing the Kth-level classifier is:

[0034] .

[0035] Preferably, the step of performing rough set attribute reduction on the system's static feature data to obtain a static feature vector specifically includes:

[0036] Initialize and set the original decision table;

[0037] The importance of feature attributes is evaluated by calculating the positive domain and indistinguishable relations;

[0038] Construct an attribute reduction table, calculate the impact of attribute removal, and record the generation and simplification of the distinction matrix;

[0039] Based on the attribute reduction table, generate all possible attribute subsets, and select the smallest attribute subset and the corresponding decision rule;

[0040] Remove redundant features and irrelevant attributes whose importance is below a preset threshold;

[0041] The reduced static eigenvectors are obtained.

[0042] Preferably, the step of evaluating the importance of feature attributes by calculating positive regions and indistinguishable relations specifically involves:

[0043] Initialize the original decision table as follows: , where U is the sample set, A is the condition attribute set, and d is the decision attribute;

[0044] For attribute subsets The definition of the indistinguishable relation of attribute subset B is:

[0045]

[0046] In the formula: Let x and y be sample pairs, where both x and y belong to sample set U; The Cartesian product is the set of all possible pairs of samples in the sample set U. This means that for each attribute a in the attribute subset B, the values ​​of samples x and y in attribute a are equal.

[0047] property The importance of attribute a is:

[0048]

[0049] In the formula: Indicates the importance of attribute a. This represents the degree of dependence of attribute subset A on decision attribute d. This represents the subset after removing attribute a from the entire set of conditional attributes A;

[0050] attribute subset The dependency of attribute subset B on decision attribute d is:

[0051]

[0052] In the formula: This represents the degree of dependence of attribute subset B on decision attribute d. This represents the positive region of attribute subset B with respect to decision d.

[0053] Preferably, the classification head is set at the output of the Informer encoder, specifically including two fully connected layers and a Softmax activation function, used to perform preliminary transient and stable state binary classification prediction on the input time series data, so that the encoder learning focuses on distinguishing between transient and stable states;

[0054] The two fully connected layers contain 128 neurons and 2 neurons, respectively.

[0055] According to a second aspect of this application, a power grid time-series stability assessment method with high accuracy and strong interpretability is provided, comprising:

[0056] Real-time acquisition of multi-dimensional time-series data after power grid disturbances and simultaneous system static characteristic data;

[0057] The acquired multidimensional time-series data and system static characteristic data are input into the optimal model for processing to obtain the corresponding power grid time-series stability assessment results and decision path results;

[0058] The optimal model mentioned above is the optimal model constructed by the aforementioned model building method.

[0059] According to a third aspect of this application, a power grid timing stability assessment system with high accuracy and strong interpretability is provided, including a module for implementing the power grid timing stability assessment method with high accuracy and strong interpretability.

[0060] The model building method provided in this application introduces a hierarchical structure and a cost-sensitive learning mechanism into the basic multivariate decision tree model to obtain an improved multivariate decision tree model. The improved multivariate decision tree model is initialized by acquiring multidimensional time-series data after power grid disturbance and system static feature data at the same moment. After processing, fixed-dimensional transient time-series feature vectors and static feature vectors are obtained. The fixed-dimensional transient time-series feature vectors and static feature vectors are concatenated and fused using a dimension superposition method to form a fused feature vector. The fused feature vector is then input into the initialized improved multivariate decision tree model for iterative training. Based on the corresponding power grid time-series stability assessment results and decision path results, the optimal model is finally obtained.

[0061] This application aims to address the dilemma of balancing accuracy and interpretability in power grid time-series stability assessment methods under complex operating conditions such as the integration of high-proportion power electronic equipment and large-scale grid connection of new energy power generation. It proposes a model-building method for power grid time-series stability assessment that achieves both high accuracy and strong interpretability. This model integrates an improved multivariate decision tree model and the Informer model, innovatively combining the powerful long-term feature extraction capability of the Informer model with the transparent decision-making process of the multivariate decision tree model, thus achieving a balance between "high accuracy" and "strong interpretability." The Informer model is used... With its powerful temporal feature learning capability, the system "reduces" and "encodes" complex transient processes into highly informative feature vectors. By introducing a hierarchical structure and cost-sensitive learning mechanism into the multivariate decision tree base model, the improved model prioritizes the accuracy of identifying more dangerous "instability" states during decision-making, which better meets the actual needs of power grid safe operation. Based on the excellent rule mining and partitioning capabilities of the improved multivariate decision tree, a concise, efficient, and completely transparent set of decision rules is constructed, which is convenient for power grid operators to read and analyze. It can serve as a direct basis for dispatching operators to make rapid decisions in actual work, providing a highly reliable auxiliary decision-making tool for power grid operation.

[0062] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of what is pointed out in the written description and the accompanying drawings. Attached Figure Description

[0063] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0064] Figure 1 A flowchart of the power grid timing stability assessment method with high accuracy and strong interpretability provided in this application.

[0065] Figure 2 The ablation experiment results of the power grid time-series stability assessment model provided for this application, which has high accuracy and strong interpretability.

[0066] Figure 3 Performance analysis diagram of the model for power grid time-series stability assessment with high accuracy and strong interpretability provided for this application.

[0067] Figure 4 A comparison of the decision boundary and interpretability analysis of the high-precision and highly interpretable power grid time-series stability assessment model provided for this application and the traditional multivariate decision tree model. Detailed Implementation

[0068] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0069] To address some problems existing in the current technology:

[0070] In a first aspect, this application provides a model establishment method for power grid time-series stability assessment with high accuracy and strong interpretability.

[0071] For example, the model building method for power grid time-series stability assessment with high accuracy and strong interpretability, such as... Figure 2-4 As shown, the model establishment method for power grid time-series stability assessment includes:

[0072] The Informer encoder is pre-trained using the training set in the pre-trained total sample set. The classification head of the encoder after pre-training is removed, and the encoder weights are fixed to obtain the initialized Informer encoder.

[0073] The pre-trained total sample set is input into the initialized Informer encoder for processing to obtain a fixed-dimensional transient temporal feature vector training set;

[0074] By introducing a hierarchical structure and a cost-sensitive learning mechanism into the basic multivariate decision tree model, an improved multivariate decision tree model is obtained.

[0075] The improved multivariate decision tree model is trained using a fixed-dimensional transient time-series feature vector training set to obtain the initialized improved multivariate decision tree model.

[0076] After acquiring multidimensional time-series data following power grid disturbance and simultaneous system static feature data, the multidimensional time-series data is input into the initialized Informer encoder for processing to obtain a fixed-dimensional transient time-series feature vector. Rough set attribute reduction is performed on the system static feature data to obtain a static feature vector.

[0077] A dimension stacking method is used to concatenate and fuse the transient temporal feature vector with the static feature vector to form a fused feature vector;

[0078] The fused feature vectors are input into the initialized multivariate decision tree improvement model for iterative training. Based on the corresponding power grid time-series stability assessment results and decision path results, the optimal model is obtained and used as the model for power grid time-series stability assessment.

[0079] Based on the above scheme, the model constructed according to the model building method for power grid time series stability assessment is a power grid time series stability assessment model that integrates the improved multivariate decision tree (IMDT) and Informer. It innovatively combines the powerful long-term feature extraction capability of the Informer model with the transparent decision-making process of the multivariate decision tree, achieving a unity of "high accuracy" and "strong interpretability".

[0080] To verify the effectiveness of each component of the model, an ablation experiment was designed, and the results are as follows: Figure 2 As shown. The experimental setup is as follows:

[0081] Scenario 1: Multivariate decision tree MDT using only static features and a small number of handcrafted dynamic features (equivalent to the traditional multivariate decision tree MDT benchmark).

[0082] Scenario 2: Only the fixed-dimensional transient temporal feature vectors extracted by Informer are used as the input of IMDT, without using static features.

[0083] Scenario 3: Fusion of static features and fixed-dimensional transient temporal feature vectors, but without using cost-sensitive learning (i.e., C_FN:C_FP = 1:1).

[0084] Scenario 4: Full model (integrating all features and employing cost-sensitive learning).

[0085] Experimental results are from Figure 2 It is evident that using any single feature (Scenario 1, 2) alone does not perform as well as the fused version (Scenario 3, 4). Specifically, introducing a fixed-dimensional transient temporal feature vector significantly reduces the instability false alarm rate, demonstrating the crucial role of temporal features in capturing instability dynamics. Furthermore, the addition of cost-sensitive learning further reduces the instability false alarm rate from 2.5% to 0.9%. Although the stability false alarm rate increases slightly, this trade-off is entirely worthwhile for grid security. This ablation experiment fully demonstrates the effectiveness and necessity of the proposed model integrating the Improved Multivariate Decision Tree (IMDT) and Informer, along with the introduction of cost-sensitive learning.

[0086] To further demonstrate the superior performance of the evaluation model proposed in this application, a performance comparison of the power grid time-series stability evaluation model integrating the improved multivariate decision tree and Informer was conducted at different decision times. The results are as follows: Figure 3 As shown.

[0087] Figure 3 (a) shows that when the decision time reaches 0.6s, the model's recognition rate for both unstable and stable samples reaches 100%, indicating that the model can make a reliable judgment in a very short time after the failure occurs. Figure 3 (b) is an in-depth analysis of the model's misclassification rate. Combined with... Figure 2 and Figure 3 As can be seen from the figure, the stable false alarm rate is higher than the unstable false alarm rate under certain conditions. This is not due to the inadequacy of the model performance, but rather the expected result of actively balancing risks through cost-sensitive learning. It is the ultimate ability to avoid catastrophic risks of the system at the cost of acceptable false alarms.

[0088] To more clearly demonstrate the classification performance of the proposed model on the two types of samples, a pre-training total sample set was generated by simulating three-phase short-circuit faults under different fault lines on the IEEE 39-node standard test system. The total sample set was divided into an independent training set (70%), a validation set (15%), and a test set (15%). Standardized simulation tests were conducted using 1000 samples from the test set in the pre-training total sample set. The results are shown in Table 1, which is the confusion matrix of the IMDT-Informer model on the test set.

[0089] Table 1: Confusion Matrix of IMDT-Informer Model (Number of Test Samples: 1000)

[0090]

[0091] To further verify the overall performance of the IMDT-Informer model, the key performance of the IMDT-Informer model was compared with that of various benchmark models, and the results are shown in Table 2.

[0092] Table 2: Comparison of overall performance of different models

[0093]

[0094] As can be seen from Table 2:

[0095] 1. Accuracy Advantage: The IMDT-Informer model proposed in this application performs well on all key metrics: its overall accuracy is as high as 99.0%, which is on par with the state-of-the-art black box model Standalone Informer (99.3%), and significantly outperforms the Long Short-Term Memory Network model LSTM (97.8%) and the traditional multivariate decision tree model MDT (93.5%).

[0096] 2. Safety: The IMDT-Informer model has a false negative rate of only 0.9%, which is the lowest among all models. This is due to the cost-sensitive learning mechanism introduced in the IMDT-Informer model, which makes the IMDT-Informer model prioritize the accuracy of identifying more dangerous unstable states when making decisions. This also makes it the strongest in capturing real unstable risks and the safest.

[0097] 3. Balance between efficiency and interpretability: Although the Standalone Informer model has slightly higher overall accuracy, it is a complete "black box". The IMDT-Informer model provides completely transparent decision rules while achieving the same accuracy. Moreover, its inference time (0.3ms) is much faster than deep learning models and only slightly slower than the simplest traditional multivariate decision tree model MDT, which fully meets the timeliness requirements of online applications.

[0098] To further illustrate the differences in decision boundaries between the traditional multivariate decision tree model and the IMDT-Informer model, a comparative analysis of the decision boundaries and interpretability of the two models was conducted. The results are as follows: Figure 4 As shown. Figure 4 (a) shows that the traditional multivariate decision tree model (MDT) requires four rules to roughly divide the stable and unstable regions, and there are obvious overlaps and ambiguities between the boundaries, which stems from its insufficient understanding of dynamic features. Figure 4 (b) shows that IMDT-Informer achieved a clearer and more robust partitioning using only 3 rules. According to Figure 4 (b) In rule a, the blue dots (samples of stable power grid operating conditions) are more densely covered and more concentrated in scope. According to statistical calculations, their support increased from 70.13% to 80.25%, with a confidence level as high as 99.9%. This indicates that the fixed-dimensional transient time-series features provided by Informer greatly enhance the model's ability to aggregate and identify stable operating conditions. The reduction in decision complexity directly proves that high-quality feature fusion can lead to simpler and more efficient decision logic.

[0099] As can be seen from the above, the model proposed in this application effectively solves the inherent problems of traditional single models, such as high accuracy but huge computational load, long response time and poor interpretability, or although fast and interpretable, insufficient accuracy and poor adaptability, and quickly and efficiently achieves the unity of "high accuracy" and "strong interpretability".

[0100] In some possible implementations of the first aspect, a hierarchical structure and a cost-sensitive learning mechanism are introduced into the basic multivariate decision tree model to obtain an improved multivariate decision tree model; the improved multivariate decision tree model is trained using a fixed-dimensional transient temporal feature vector training set to obtain an initialized improved multivariate decision tree model, specifically including:

[0101] A multivariate decision tree base model is constructed by replacing the internal nodes of the decision tree with support vector machines.

[0102] Furthermore, by introducing a hierarchical structure and a cost-sensitive learning mechanism into the basic multivariate decision tree model, an improved multivariate decision tree model based on an iterative weighted strategy is obtained.

[0103] Based on the aforementioned hierarchical structure and cost-sensitive learning mechanism, the improved multivariate decision tree model is trained sequentially using a training set of fixed-dimensional transient temporal feature vectors obtained by the initialized Informer encoder. During the training process, an iterative weighting strategy is applied to dynamically adjust the sample weights, save the relevant hyperparameters, and optimize and evaluate the model performance under different hyperparameter combinations. The iterative process is repeated until the model performance converges, and finally the initialized improved multivariate decision tree model is obtained.

[0104] Based on the above scheme, a cost matrix is ​​introduced to adjust the Gini index calculation in the CART algorithm. The cost matrix is ​​defined as: C(i, j) = { C_FN, if i=unstable, j=stable; C_FP, if i=stable, j=unstable; 0, if i=j}. Considering the extreme importance of power grid safety, this scheme sets the cost ratio C_FN : C_FP to 50:1. This ratio is determined based on a detailed sensitivity analysis. Grid search and cross-validation were performed on the validation set for several representative ratios (e.g., 10:1, 20:1, 100:1). When the ratio is 50:1, the model can minimize the false alarm rate (MAR) while keeping the false alarm rate (FAR) within an engineering-acceptable range (<1.5%), achieving the best balance between risk avoidance and model generalization performance.

[0105] In some possible implementations of the first aspect, the step of sequentially stacking the improved multivariate decision tree model using a training set of fixed-dimensional transient temporal feature vectors obtained after processing the initialized Informer encoder, based on a hierarchical structure and cost-sensitive learning mechanism, specifically includes:

[0106] By introducing multiple cost-sensitive decision trees as weak classifiers and concatenating them sequentially, a cost matrix is ​​defined, a cost ratio is set, and the weights of all samples are set to the initial weights.

[0107] Based on the current weight distribution, the first weak classifier is trained by calculating the classification error rate and weights, and the prediction result is output.

[0108] Adjust the sample weights so that the weight update result of the previous weak classifier after training is used as the basis for adjusting the input weights of the next weak classifier. Repeat the iterative process of superimposed error correction and focusing until the accuracy requirement is met.

[0109] The predictions of all weak classifiers are weighted and combined according to their classification accuracy to form a strong classifier.

[0110] Based on the above scheme, the powerful temporal feature learning capability of the Informer model is utilized to "reduce" and "encode" complex transient processes into highly informative feature vectors. By introducing a hierarchical structure and cost-sensitive learning mechanism into the multivariate decision tree base model, the traditional decision tree hierarchical nodes and single complex decision tree form are abandoned. Multiple sets of shallow cost-sensitive multivariate decision trees are trained serially as weak classifiers, so that each level of weak classifier focuses on correcting the classification error of the previous level. Finally, a strong classifier is formed by weighted combination.

[0111] In some possible implementations of the first aspect, the sample weights are dynamically adjusted during training, relevant hyperparameters are saved, and after debugging and optimization, the model performance under different hyperparameter combinations is evaluated. The iterative process is repeated until the model performance converges, specifically including:

[0112] Assign the currently set weight value to each sample of the same category to form the current probability distribution;

[0113] Set weight update rules, dynamically update base classifier weights and adaptively perform iterative weighting through a loss function, minimize the cumulative cost of the classifier, and update sample weights.

[0114] Collect feedback data, analyze the direction of weight adjustment, save relevant hyperparameters, and check whether the performance of parameters and weights converges by debugging, optimizing and evaluating the model performance under different hyperparameter combinations. If converged, obtain the updated sample weights. If not converged, repeat the iterative process until the model performance converges.

[0115] In some possible implementations of the first aspect, the setting of weight update rules, dynamically updating the base classifier weights through a loss function and adaptively performing iterative weighting to minimize the cumulative cost of the classifier, and updating the sample weights specifically includes: setting weight update rules, training using the currently set sample weights, calculating through a cost function, identifying high-cost misclassified samples, dynamically adjusting the sample weights according to the error cost, increasing the weights of misclassified samples, decreasing the weights of correctly classified samples, repeating the iterative process, minimizing the cumulative cost of the classifier, until the model weights converge;

[0116] The weight update rule is set as follows:

[0117] Let the k-th level classifier be ,sample The weight is ,but:

[0118]

[0119] In the formula: Represents the cost function, This represents the weight adjustment factor for level k;

[0120] The cumulative cost of minimizing the Kth-level classifier is:

[0121] .

[0122] Based on the above scheme, during the hierarchical model construction process, the model dynamically updates the weights of the base classifiers according to the loss function, adaptively increasing the weight of samples misclassified by the previous classifier, especially the weight of missed samples that would incur high error costs, while decreasing the weight of correctly classified samples. This forces subsequent classifiers to continuously monitor these misclassified samples, ensuring the overall model's ability to identify high-risk events and the accuracy of its judgments through weighted iterations.

[0123] In some possible implementations of the first aspect, the rough set attribute reduction operation on the system's static feature data to obtain a static feature vector specifically includes:

[0124] Initialize and set the original decision table;

[0125] The importance of feature attributes is evaluated by calculating the positive domain and indistinguishable relations;

[0126] Construct an attribute reduction table, calculate the impact of attribute removal, and record the generation and simplification of the distinction matrix;

[0127] Based on the attribute reduction table, generate all possible attribute subsets, and select the smallest attribute subset and the corresponding decision rule;

[0128] Remove redundant features and irrelevant attributes whose importance is below a preset threshold;

[0129] The reduced static eigenvectors are obtained.

[0130] In some possible implementations of the first aspect, the assessment of feature attribute importance by calculating positive domains and indistinguishable relations specifically includes:

[0131] Initialize the original decision table as follows: , where U is the sample set, A is the condition attribute set, and d is the decision attribute;

[0132] For attribute subsets The definition of the indistinguishable relation of attribute subset B is:

[0133]

[0134] In the formula: Let x and y be sample pairs, where both x and y belong to sample set U; The Cartesian product is the set of all possible pairs of samples in the sample set U. This means that for each attribute a in the attribute subset B, the values ​​of samples x and y in attribute a are equal.

[0135] property The importance of attribute a is:

[0136]

[0137] In the formula: Indicates the importance of attribute a. This represents the degree of dependence of attribute subset A on decision attribute d. This represents the subset after removing attribute a from the entire set of conditional attributes A;

[0138] attribute subset The dependency of attribute subset B on decision attribute d is:

[0139]

[0140] In the formula: This represents the degree of dependence of attribute subset B on decision attribute d. This represents the positive region of attribute subset B with respect to decision d.

[0141] Based on the above scheme, attribute reduction is performed on the input static features according to rough set attribute reduction theory. By calculating the importance of each static attribute to the final decision (stable / unstable), redundant or irrelevant attributes with importance below a preset threshold are eliminated, reducing the interference of redundant static features on the model's decision-making and improving generalization ability. In this scheme, the preset threshold is determined by performing grid search and five-fold cross-validation on the model performance at different thresholds, and is finally set to 0.1 to achieve the best balance between simplifying the feature set and retaining key information.

[0142] In some possible implementations of the first aspect, the classification head is set at the output of the Informer encoder and specifically includes two fully connected layers and a Softmax activation function, used to perform preliminary transient and stable state binary classification prediction on the input time series data, so that the encoder learning focuses on distinguishing between transient and stable states.

[0143] The two fully connected layers contain 128 neurons and 2 neurons, respectively.

[0144] Based on the above scheme, the detachable classification head in this application does not bind the output layer to the prediction target. After pre-training, the classification head is removed, retaining only the trained encoder weights. This forces the encoder to focus on distinguishing transient stability. Through the binary classification task (stable / instable) and cross-entropy loss of the classification head, the encoder learns high-dimensional abstract features (fixed-dimensional transient time-series feature vectors) that can distinguish stable states of the power grid, rather than generalized time-series features, thus avoiding end-to-end task interference. The pre-training objective is not to build a classifier, but to calibrate the feature extractor, ensuring that the subsequently improved multivariate decision tree model can obtain high-quality input.

[0145] Secondly, this application provides a power grid timing stability assessment method with high accuracy and strong interpretability. This assessment method can be executed by a power grid timing stability assessment system with high accuracy and strong interpretability, or by components configured within the power grid timing stability assessment system, such as chips or chip systems, or by logic modules or software having some or all of the functions of the power grid timing stability assessment system. This application does not limit this aspect.

[0146] For example, this power grid time-series stability assessment method with high accuracy and strong interpretability, such as Figure 1 As shown, the power grid time-series stability assessment system includes:

[0147] Real-time acquisition of multi-dimensional time-series data after power grid disturbances and simultaneous system static characteristic data;

[0148] The acquired multidimensional time-series data and system static characteristic data are input into the optimal model for processing to obtain the corresponding power grid time-series stability assessment results and decision path results;

[0149] The optimal model mentioned above is the optimal model constructed by the aforementioned model building method.

[0150] Based on the above scheme, the improved multivariate decision tree model is trained using a fixed-dimensional transient time-series feature vector training set to obtain an initialized improved multivariate decision tree model. A PMU is used for real-time acquisition to obtain multidimensional time-series data after grid disturbances and simultaneous system static feature data. The multidimensional time-series data is input into the initialized Informer encoder for processing to obtain a fixed-dimensional transient time-series feature vector (TDFV). Rough set attribute reduction is performed on the system static feature data to obtain a static feature vector. The fixed-dimensional transient time-series feature vector and the static feature vector are concatenated and fused using a dimension stacking method to form a fused feature vector. The fused feature vector is input into the optimal model for processing to obtain the grid time-series stability assessment results and decision path results. This high-precision and highly interpretable grid time-series stability assessment method balances interpretability and efficiency. Compared with "black box" model assessment methods, this method achieves the same level of accuracy while generating clear and concise decision rules, and its decision complexity is reduced by 25% compared to traditional MDT, with a single-sample inference time of only 0.3ms, meeting the requirements of online applications.

[0151] To demonstrate the interpretability of the model, a high-confidence rule X, which plays a crucial role in 'instability' detection, is extracted from the optimal model of this application: IF (TDFV_feature2 < -0.85) AND (Static_feature5 > 1.1) THEN 'Unstable' (Support: 95%, Confidence: 99.5%).

[0152] Interpreting the physical meaning of rule X reveals that when the fixed-dimensional transient time-series feature (TDFV_feature2) extracted by Informer shows a rapid separation trend, and the system's static feature (Static_feature5, such as the power flow of a heavily loaded line) is also at a high load level, the system will be highly likely to experience power angle instability. To further verify this, the importance of the fixed-dimensional transient time-series feature (TDFV) in the improved multivariate decision tree (IMDT) model was ranked, revealing that TDFV_feature2 contributes the most among all instability discrimination rules. Furthermore, by calculating the correlation between this feature and the original time-series data, a strong correlation was found between it and the rate of change of the critical generator power angle within 0.2-0.5 seconds after the fault. The Pearson correlation coefficient between the value of TDFV_feature2 and the separation rate of the critical generator power angle in the original time-series data was calculated to be as high as 0.92, demonstrating a strong linear positive correlation and fully confirming that this feature vector dimension accurately captures the early dynamics of system separation. This further proves that such clear, concise rules with explicit physical meaning are the direct basis needed by dispatchers to make rapid decisions in their actual work, perfectly demonstrating the interpretability value of this model.

[0153] Thirdly, this application provides a power grid timing stability assessment system with high accuracy and strong interpretability, including a module for implementing the power grid timing stability assessment method with high accuracy and strong interpretability.

[0154] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as C, VHDL, Verilog, the object-oriented programming language Java, and the interpreted scripting language JavaScript.

[0155] This application is described with reference to flowchart illustrations of methods, apparatus, and computer program products according to embodiments of this application. It should be understood that each step in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart.

[0156] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0157] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 A process or multiple process diagrams specify the steps of a function.

[0158] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0159] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0160] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A model building method for power grid timing stability evaluation with high precision and strong interpretability, characterized in that, include: The Informer encoder is pre-trained using the training set in the pre-trained total sample set. The classification head of the encoder after pre-training is removed, and the encoder weights are fixed to obtain the initialized Informer encoder. The pre-trained total sample set is input into the initialized Informer encoder for processing to obtain a training set of transient temporal feature vectors with fixed dimensions; By introducing a hierarchical structure and a cost-sensitive learning mechanism into the basic multivariate decision tree model, an improved multivariate decision tree model is obtained. The improved multivariate decision tree model is trained using a fixed-dimensional transient time-series feature vector training set to obtain the initialized improved multivariate decision tree model. After acquiring multidimensional time-series data following power grid disturbance and simultaneous system static feature data, the multidimensional time-series data is input into the initialized Informer encoder for processing to obtain a fixed-dimensional transient time-series feature vector. Rough set attribute reduction is performed on the system static feature data to obtain a static feature vector. A dimension stacking method is used to concatenate and fuse the transient temporal feature vector with the static feature vector to form a fused feature vector; The fused feature vectors are input into the initialized multivariate decision tree improvement model for iterative training. Based on the corresponding power grid time-series stability assessment results and decision path results, the optimal model is obtained and used as the model for power grid time-series stability assessment.

2. The method for establishing a model for power grid time-series stability assessment with high accuracy and strong interpretability according to claim 1, characterized in that, The process involves introducing a hierarchical structure and a cost-sensitive learning mechanism into the basic multivariate decision tree model to obtain an improved multivariate decision tree model. The improved multivariate decision tree model is then trained using a fixed-dimensional transient temporal feature vector training set to obtain an initialized improved multivariate decision tree model, specifically including: A multivariate decision tree base model is constructed by replacing the internal nodes of the decision tree with support vector machines. Furthermore, by introducing a hierarchical structure and a cost-sensitive learning mechanism into the basic multivariate decision tree model, an improved multivariate decision tree model based on an iterative weighted strategy is obtained. Based on a hierarchical structure and cost-sensitive learning mechanism, a multivariate decision tree improvement model is trained serially using a training set of fixed-dimensional transient temporal feature vectors obtained by initializing the Informer encoder. During training, an iterative weighting strategy is applied to dynamically adjust sample weights, save relevant hyperparameters, and optimize and evaluate the model performance under different hyperparameter combinations. The iterative process is repeated until the model performance converges, and finally, the initialized multivariate decision tree improvement model is obtained.

3. The method for establishing a model for power grid time-series stability assessment with high accuracy and strong interpretability according to claim 2, characterized in that, The hierarchical structure and cost-sensitive learning mechanism, which uses a fixed-dimensional transient temporal feature vector training set obtained after initializing the Informer encoder to perform serial stacking training on the multivariate decision tree improvement model, specifically includes: By introducing multiple cost-sensitive decision trees as weak classifiers and concatenating them sequentially, a cost matrix is ​​defined, a cost ratio is set, and the weights of all samples are set to the initial weights. Based on the current weight distribution, the first weak classifier is trained by calculating the classification error rate and weights, and the prediction result is output. Adjust the sample weights so that the weight update result of the previous weak classifier after training is used as the basis for adjusting the input weights of the next weak classifier. Repeat the iterative process of superimposed error correction and focusing until the accuracy requirement is met. The predictions of all weak classifiers are weighted and combined according to their classification accuracy to form a strong classifier.

4. The method for establishing a model for power grid time-series stability assessment with high accuracy and strong interpretability according to claim 2, characterized in that, During training, an iterative weighting strategy is applied to dynamically adjust sample weights, save relevant hyperparameters, and optimize and optimize the model. The model performance under different hyperparameter combinations is evaluated, and the iterative process is repeated until the model performance converges. Specifically, this includes: Assign the currently set weight value to each sample of the same category to form the current probability distribution; Set weight update rules, dynamically update base classifier weights and adaptively perform iterative weighting through a loss function, minimize the cumulative cost of the classifier, and update sample weights. Collect feedback data, analyze the direction of weight adjustment, save relevant hyperparameters, and check whether the performance of parameters and weights converges by debugging, optimizing and evaluating the model performance under different hyperparameter combinations. If converged, obtain the updated sample weights. If not converged, repeat the iterative process until the model performance converges.

5. The method for establishing a model for power grid time-series stability assessment with high accuracy and strong interpretability according to claim 4, characterized in that, The setting of weight update rules, through a loss function, dynamically updates the base classifier weights and adaptively iterates the weights to minimize the cumulative cost of the classifier and update the sample weights. Specifically, this includes: setting weight update rules, training with the currently set sample weights, calculating the cost function, identifying high-cost misclassified samples, dynamically adjusting the sample weights based on the error cost, increasing the weights of misclassified samples, decreasing the weights of correctly classified samples, repeating the iterative process to minimize the cumulative cost of the classifier until the model weights converge. The weight update rule is set as follows: Let the k-th level classifier be ,sample The weight is ,but: In the formula: Represents the cost function, This represents the weight adjustment factor for level k; The cumulative cost of minimizing the Kth-level classifier is: 。 6. The method for establishing a model for power grid time-series stability assessment with high accuracy and strong interpretability according to claim 1, characterized in that, The rough set attribute reduction operation on the system's static feature data to obtain the static feature vector specifically includes: Initialize and set the original decision table; The importance of feature attributes is evaluated by calculating the positive domain and indistinguishable relations; Construct an attribute reduction table, calculate the impact of attribute removal, and record the generation and simplification of the distinction matrix; Based on the attribute reduction table, generate all possible attribute subsets, and select the smallest attribute subset and the corresponding decision rule; Remove redundant features and irrelevant attributes whose importance is below a preset threshold; The reduced static eigenvectors are obtained.

7. The method for establishing a model for power grid time-series stability assessment with high accuracy and strong interpretability according to claim 6, characterized in that, The evaluation of feature attribute importance by calculating positive regions and indistinguishable relations is specifically as follows: Initialize the original decision table as follows: , where U is the sample set, A is the condition attribute set, and d is the decision attribute; For attribute subsets The definition of the indistinguishable relation of attribute subset B is: In the formula: Let x and y be sample pairs, where both x and y belong to sample set U; The Cartesian product is the set of all possible pairs of samples in the sample set U. This means that for each attribute a in the attribute subset B, the values ​​of samples x and y in attribute a are equal. property The importance of attribute a is: In the formula: Indicates the importance of attribute a. This represents the degree of dependence of attribute subset A on decision attribute d. This represents the subset after removing attribute a from the entire set of conditional attributes A; attribute subset The dependency of attribute subset B on decision attribute d is: In the formula: This represents the degree of dependence of attribute subset B on decision attribute d. This represents the positive region of attribute subset B with respect to decision d.

8. The method for establishing a model for power grid time-series stability assessment with high accuracy and strong interpretability according to claim 1, characterized in that, The classification head is set at the output of the Informer encoder and specifically includes two fully connected layers and a Softmax activation function. It is used to perform preliminary transient and stable state binary classification prediction on the input time series data, so that the encoder learning focuses on distinguishing between transient and stable states. The two fully connected layers contain 128 neurons and 2 neurons, respectively.

9. A power grid time-series stability assessment method with high accuracy and strong interpretability, characterized in that, include: Real-time acquisition of multi-dimensional time-series data after power grid disturbances and simultaneous system static characteristic data; The acquired multidimensional time-series data and system static characteristic data are input into the optimal model for processing to obtain the corresponding power grid time-series stability assessment results and decision path results; The optimal model mentioned above is the optimal model according to any one of claims 1-8.

10. A power grid time-series stability assessment system with high accuracy and strong interpretability, characterized in that: It includes modules for implementing the power grid timing stability assessment method with high accuracy and strong interpretability as described in claim 9.

Images