Power grid dominant instability mode identification method considering feature engineering and image dimensionality

Abstract

The application discloses a power grid leading instability mode identification method considering feature engineering and image dimensionality, comprising the following steps: acquiring initial electrical quantity features and power grid fault sample input time series curves; constructing an operation state feature supplement sequence considering static features and dynamic features in the operation state; constructing a power system topology structure feature supplement sequence; constructing a fusion feature index; acquiring power grid fault sample input time series curves considering feature engineering; constructing a dynamic fluctuation weight index measuring the dynamic fluctuation degree of the curve; screening the power grid fault sample input time series curves considering feature engineering; encoding the screened input time series curves into two-dimensional images; inputting the two-dimensional images into a pre-trained power grid leading instability mode intelligent identification model to obtain leading instability mode identification results. The application further improves the application value of the power grid leading instability mode identification based on the data-driven method.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching and operation control technology, and in particular to a method for identifying the dominant instability mode of a power grid that considers feature engineering and image dimensionality enhancement. Background Technology

[0002] When a power system experiences a fault, its emergency dynamic response may exhibit complex instability modes. Therefore, rapidly and accurately identifying the dominant instability mode is crucial for maintaining the safety and stability of the power system. The identification results of the dominant instability mode after a power system fault can be used to formulate and implement subsequent emergency control decisions, which is of great significance for rapid decision-making during fault situations and for preventative decision-making in fault prediction scenarios.

[0003] Currently, the identification and analysis of instability modes relies heavily on expert experience and human judgment. However, with the continuous improvement of wide-area measurement system deployment and simulation analysis technology, the amount of data requiring analysis and identification is enormous. Yet, there is a severe shortage of personnel with the necessary expertise to perform this task. This leads to several deep-seated problems in the identification of instability modes: system modeling is often required, followed by the determination of corresponding criteria for individual data curves for targeted analysis, a complex and time-consuming process; manual simulation analysis is insufficient to meet the timeliness requirements for the large amount of instability data obtained from actual power grid simulations, necessitating significant manpower costs.

[0004] The existing data-driven dominant instability identification methods lack sufficient consideration of the characteristics of the input data, resulting in insufficient adaptability of the proposed methods. For the complex curve input screening work, the literature [3] attempts to set range indicators for linear screening, which may cause invalid curves to be included as selected curves, affecting the subsequent model identification; at present, the modeling method based on clustering algorithm and using Euclidean distance, slope of change and other data to evaluate clustering provides a new idea for the identification of dominant instability modes, but the defect of insufficient consideration of curves under different fluctuation conditions still needs to be focused on. In order to solve the problem of topological transferability, some researchers have mentioned that the spatiotemporal information in the time series data can be taken into consideration, and the topological features can be linked to the training process by using GCN, and the image topology can be linked by n-dimensional data for convolution operation, but it does not emphasize the identification of diverse spatiotemporal information carried in a single data, and lacks targeted features for all samples under the same current operation.

[0005] Overall, while existing data-driven instability identification methods can achieve accurate identification to a certain extent, they still have problems: the characteristics of the data used do not fully take into account the operating state and spatial topology information contained therein; the screening methods for the large number of power angle and voltage curves contained in a single operation sample are lacking; and for multi-dimensional time series characteristic curves, the models used have difficulty simultaneously and fully considering the global and local features of the sample data. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, the power grid dominant instability mode identification method considering feature engineering and image dimensionality enhancement provided by this invention solves the problems of existing instability mode identification methods failing to fully consider the operating state and spatial topology information of the power system, struggling to fully consider global and local features, and the curve screening method failing to fully consider fluctuations, resulting in insufficient identification accuracy.

[0007] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a method for identifying dominant instability modes of power grids considering feature engineering and image dimensionality enhancement, characterized in that it includes: Obtain initial electrical quantity characteristics and time-series curves of power grid fault sample inputs; Based on the initial electrical quantity characteristics and the input time series curves of power grid fault samples, an operation state feature supplementary sequence considering the static and dynamic characteristics in the operation state is constructed; Construct a power system topology feature supplementary sequence that considers the spatial adjacency matrix of the power system; A fusion feature index considering both the operating state information and topological structure information of power system instability samples is constructed based on the improved PageRank algorithm; The supplementary sequences of operating status features, supplementary sequences of power system topology features, and fusion feature indicators are combined and added to the initially acquired power grid fault sample input time series curve to obtain the power grid fault sample input time series curve considering feature engineering. Construct a dynamic fluctuation weight index to measure the degree of dynamic fluctuation of the curve; Based on the dynamic fluctuation weight index, the input time series curves of power grid fault samples with fluctuation levels higher than the preset value, different modes, and representing different fluctuation characteristics are selected to consider the feature engineering, and the selected input time series curves are obtained. The filtered input time series curves are encoded into two-dimensional images based on nonlinear image upscaling technology; Construct a pre-trained intelligent identification model for the dominant instability modes of the power grid; The two-dimensional image is input into the pre-trained intelligent identification model of the dominant instability mode of the power grid to obtain the identification result of the dominant instability mode.

[0008] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of a power grid dominant instability mode identification method that takes into account feature engineering and image dimensionality enhancement.

[0009] The beneficial effects of this invention are as follows: (1) The supplementary sequence of operating status features, the supplementary sequence of topology features, and the comprehensive index features that integrate operating status and topology constructed in this invention fully consider the multifaceted information when the power grid is transiently unstable. Through data preprocessing, the adaptability of the model under different power grid topologies and different power flow instability faults is guaranteed.

[0010] (2) The curve screening method based on the DTW distance improved clustering algorithm with dynamic fluctuation weight index correction proposed in this invention can quickly and accurately screen effective curves, avoiding excessive dimensionality caused by too many single sample input curves. Compared with the prior art, this scheme can effectively enhance the sample data, while ensuring that the dominant instability mode is accurately identified, minimizing the number of sample input curves, reducing the complexity of the judgment process, and can be applied to fault conditions under different power system conditions.

[0011] (3) The present invention, based on nonlinear image dimensionality enhancement technology and deep learning model, expands the detailed features of the data, while comprehensively considering the global and local features of the image to be identified. Compared with the prior art, the present invention better solves the shortcomings of traditional methods that are limited by expert experience and manual identification, and further enhances the application value of dominant instability mode identification based on data-driven methods. Attached Figure Description

[0012] Figure 1 A flowchart of a power grid dominant instability mode identification method considering feature engineering and image upscaling provided for an embodiment; Figure 2 The flowchart of the improved K-means clustering algorithm provided in the example is shown below; Figure 3 A schematic diagram illustrating the conversion of a one-dimensional image into a two-dimensional image for an embodiment; Figure 4 The standard example diagram of the CEPRI-36 node system used in the embodiment is provided; Figure 5 The following are schematic diagrams of the power angle sequence curves of each unstable sample obtained based on the example provided in the embodiment, wherein (a) is a schematic diagram of the power angle sequence curve in the stable state, (b) is a schematic diagram of the power angle sequence curve in the power angle-dominated unstable state, and (c) is a schematic diagram of the power angle sequence curve in the voltage-dominated unstable state. Figure 6The following are schematic diagrams of voltage sequence curves of each unstable sample obtained based on the example provided in the embodiment, wherein (a) is a schematic diagram of voltage sequence curve in the steady state, (b) is a schematic diagram of voltage sequence curve in the power angle-dominated unstable state, and (c) is a schematic diagram of voltage sequence curve in the voltage-dominated unstable state. Figure 7 The confusion matrix diagram obtained from the experimental verification provided in the embodiments; Figure 8 The comparison and verification index results of different comparison models provided in the embodiments are shown in the figure. Figure 9 The following is a graph showing the comparison and verification results of different ablation models provided in the examples; Figure 10 A schematic diagram of the added noise disturbance provided for the embodiment; Figure 11 The diagram illustrates the effects of different disturbances on the implementation examples. Detailed Implementation

[0013] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0014] like Figure 1 As shown, in one embodiment of the present invention, a method for identifying the dominant instability mode of a power grid considering feature engineering and image dimensionality enhancement includes the following steps: S1. Obtain the initial electrical quantity characteristics and the input time-series curve of the power grid fault sample.

[0015] The input timing curves for power grid fault samples include the fault voltage curve and the power angle curve.

[0016] S2. Based on the initial electrical quantity characteristics and the time-series curves of power grid fault sample input, construct an operation state feature supplementary sequence that considers the static and dynamic characteristics in the operation state.

[0017] After collecting the initial electrical quantity characteristics, the static and dynamic characteristics contained in the operating state are additionally considered to comprehensively construct the supplementary sequence of operating state characteristics. This enhances the grid fault sample input information architecture from the perspective of the transient operating state of the power system, and fully considers the diverse static and dynamic characteristics contained in the operating state in the sample input information.

[0018] For each power angle curve and voltage curve, the mean, variance, maximum, minimum, range, skewness, and kurtosis are selected as static characteristics in the operating state, and their expressions are as follows:

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025] In the formula, Represents the mean of the curve. This represents the total number of sampling times. Indicates the time index. Indicates time Electrical values ​​at that time; Represents the variance of the curve. Indicates the maximum value. Indicates the minimum value. Indicates time Electrical values ​​at that time; Indicates the range. Indicates skewness, Indicates the standard deviation of the curve. Indicates kurtosis.

[0026] The above features can more specifically and clearly show the degree of fluctuation of the time series curve and clarify the electrical changes at each moment.

[0027] Dynamic features can be used to demonstrate the dynamic patterns in the transient instability curve, including the mean of the first derivative of the input time series curve of the power grid fault sample, the number of inflection points, and the number of abrupt changes, the expressions of which are as follows:

[0028]

[0029]

[0030] In the formula, The first derivative means. express Electrical value at any given time Indicates the number of inflection points. Indicates the curve at The second difference at time, For indicator functions, Indicates the number of mutations. This represents the mutation threshold.

[0031] This step involves constructing supplementary features that characterize the operating state of each curve in all samples. Data preprocessing ensures the model's adaptability under different power flow instability fault conditions in different power systems.

[0032] S3. Construct a supplementary sequence of power system topology features that takes into account the spatial adjacency matrix of the power system.

[0033] By additionally considering the spatial adjacency characteristics of the power system and constructing a supplementary sequence of topological features, the architecture of the power grid fault sample input information is enhanced from the perspective of the power system's transient topology, fully considering the diverse features inherent in the topology of the sample input information. The power system topology features considering the spatial adjacency matrix of the power system include: The maximum difference between two adjacent buses at the same moment is expressed as:

[0034]

[0035] In the formula, Indicates busbar With busbar The difference at the same time, Indicates busbar At any moment The electrical quantity below, busbar At any moment The electrical quantities below; This represents the maximum difference between two adjacent buses at the same time. Indicates the number of adjacent buses; The voltage variation trend of a single node and its adjacent bus is expressed as follows:

[0036]

[0037] In the formula, Indicates adjacent busbars With adjacent busbar The Pearson correlation coefficient between them , Indicates adjacent busbars The mean of the corresponding input timing curve, Indicates adjacent busbars The mean of the corresponding input timing curve, This indicates the voltage variation trend between a single node and its adjacent bus. When... A value greater than 0.8 indicates a strong positive correlation. A correlation coefficient less than -0.5 indicates a strong negative correlation. Taking voltage nodes as an example, when the mean correlation coefficient is close to 1, it indicates that the voltage change trend of the node is highly consistent with that of the adjacent bus, and the network voltage regulation coordination is good. If the mean correlation coefficient suddenly drops, it may indicate a change in the topology.

[0038] This step involves constructing supplementary features for each curve in all samples to characterize the topological structure between the curves at each node. Data preprocessing ensures the model's adaptability under instability fault conditions in different topological power grids.

[0039] S4. Construct a fusion feature index based on the improved PageRank algorithm, which considers the operating state information and topology information of power system instability samples.

[0040] The specific method is as follows: The expression for correcting the element values ​​in the node link matrix of a power system is as follows:

[0041] In the formula, This indicates that the starting node in the link matrix is... The termination node is The element value, For nodes Side and node The absolute difference in active power on both sides For all nodes The sum of power changes in connected lines; Based on the corrected node link matrix, the fusion feature index is calculated, and its expression is as follows:

[0042] In the formula, Represents a node The corresponding fusion feature index, The damping factor, This represents the corrected node link matrix, where the superscript T indicates the transpose of the matrix. Represents nodes The total number of connected nodes. For nodes One of the connected nodes, It is the identity matrix. This represents the total number of nodes.

[0043] By constructing a fusion index feature that comprehensively considers the operating state information and topology information of power system instability samples, and then adding it to the original time series in the subsequent step S5, more comprehensive system information can be provided.

[0044] S5. Combine the supplementary sequence of operating status features, the supplementary sequence of power system topology features, and the fusion feature index and add them to the initially acquired power grid fault sample input time series curve to obtain the power grid fault sample input time series curve considering feature engineering.

[0045] In this embodiment, the obtained samples of the fault voltage curve and power angle curve for the initial transient instability are shown in the following formulas:

[0046]

[0047] in, , It is the first of the voltage sample and the power angle sample. i a curve, This indicates the total number of curves contained in a voltage or power angle sample.

[0048] Using one of the voltage curves For example, the initial sample collection time is t c Step size is l , No. i The voltage curve of the busbar is as follows: ; Then, based on feature engineering, each curve is supplemented with the following effective feature sequence: ; Obtain the first data after preprocessing. i The characteristics of the input time-series curves for power grid fault samples considering feature engineering are as follows: .

[0049] By combining the supplementary sequences of operating status features, the supplementary sequences of power system topology features, and the fusion feature indicators and adding them to the input time series curve of the initially acquired power grid fault samples, the data can be made more comprehensive, and the subsequent model identification can be more universal.

[0050] For a sample with N voltage or power angle curves, in order to select the required number of effective samples k, an improved K-means clustering algorithm considering DTW distance is designed, such as... Figure 2 As shown, the specific steps include S6 and S7.

[0051] S6. Construct a dynamic fluctuation weight index to measure the degree of dynamic fluctuation of the curve.

[0052] The specific method is as follows: The mean of the first derivative of the curve is defined as the drastic index; Define the curve deviation exponent, and its expression is:

[0053] In the formula, This indicates that the curve deviates from the exponent. Indicates the maximum value of the electrical quantity. Indicates the minimum value of an electrical quantity; A dynamic volatility weighted index is constructed based on the volatility index and the curve deviation index, and its expression is as follows:

[0054]

[0055] In the formula, This represents the initial dynamic fluctuation weighting index of the curve. Indicates the final dynamic fluctuation weight of the curve; index; Indicates the intensity index. This represents the total number of individual sample power angle curves or voltage curves. This represents the initial dynamic fluctuation weight index of the Nth curve.

[0056] Using both the volatility index and the deviation index to measure the degree of fluctuation of a curve can quickly and effectively assess the effectiveness of multidimensional time series information in an unstable sample, making it easier for subsequent curve selection and model identification.

[0057] S7. Based on the dynamic fluctuation weight index, select the power grid fault sample input time series curves that have fluctuation levels higher than the preset value, different modes and represent different fluctuation characteristics, and consider the characteristic engineering of the power grid fault.

[0058] The dynamic fluctuation weight index is calculated based on the input time series curve of the power grid fault sample considering characteristic engineering, according to the dynamic fluctuation weight index. The first k curves are selected as the fluctuation centers based on the dynamic fluctuation weight index; The Dynamic Time Warping (DTW) distance between the input time-series curves and the fluctuation centers of various power grid fault samples, considering characteristic engineering aspects, is calculated based on the Dynamic Time Warping (DTW) algorithm. The DTW distance measures the morphological similarity of the time-series curves. Classification based on similarity allows for the selection of the most representative and desired number of voltage and power angle curves from a large number of single-sample data. The specific method is as follows: The expression for calculating the distance matrix between the input time-series curve of a power grid fault sample considering characteristic engineering and the fundamental distance matrix serving as the fluctuation center curve is as follows:

[0059] In the formula, Based on the basic distance matrix, , This represents two curves.

[0060] Find an optimal alignment path based on the fundamental distance matrix. , K Let this be the length of the path. Calculate the optimal alignment path to reach the final alignment point. cumulative distance Its expression is: .

[0061] When calculating the cumulative distance to the final point of the path That is and The DTW distance between these two curves.

[0062] The cumulative distance is corrected using a dynamic fluctuation weighting exponent to obtain the final DTW distance, which is expressed as follows:

[0063] In the formula, This indicates the output DTW distance. This represents the cumulative distance from the optimal alignment path to the final alignment point.

[0064] Based on the obtained DTW distance, the curve sets are filtered and classified into different clusters; Select the cluster center curves of different curve sets and remove the remaining invalid curves to obtain the filtered input timing curves.

[0065] The above steps screen the multidimensional time-series data curves in the sample from multiple perspectives. They consider not only the essential characteristics of the curves, such as their intensity, deviation, and volatility, but also the need to include curves with different volatility patterns in the screening. This allows for the selection of curves with stronger representation and reduces their number to the required scale. While ensuring that the dominant instability mode is accurately identified, this minimizes the number of input curves and reduces the complexity of the judgment process. It is also applicable to fault conditions under different power system circumstances.

[0066] S8. The filtered input time series curves are encoded into two-dimensional images based on nonlinear image upscaling technology.

[0067] Based on nonlinear image upscaling techniques (Gram angle field) encoding, it is converted into a two-dimensional image (such as...). Figure 3As shown, the multidimensional power angle and voltage time series of a single sample are stitched together to form an image, which is then used as the input sample for the dominant instability mode identification model. By upscaling the one-dimensional curve and replacing the multidimensional time series with a two-dimensional image, the overfitting problem caused by multidimensional time series data can be directly reduced, while suppressing the possibility of dimensionality explosion in subsequent model training.

[0068] S9. Construct a pre-trained intelligent identification model for the dominant instability mode of the power grid; input the two-dimensional image into the pre-trained intelligent identification model for the dominant instability mode of the power grid to obtain the identification result of the dominant instability mode.

[0069] To achieve full learning of information from the input image, a dominant instability identification model is constructed, incorporating both spatial convolution and transform mechanisms. The convolution branch enables rapid extraction of local features from the input image, and a sliding window convolution operation maps image channels to achieve the desired input channel count C. in Change to the number of output channels C out The Transformer branch can capture and establish long-range dependencies between image pixels. It first performs convolution operations to compress local features and then converts the input channel C... in Compress to C mid Then, lightweight Transformer encoding is performed. This invention combines spatial convolution with the transform mechanism, fully considering both global and local features of the samples, which can effectively improve the recognition accuracy of deep learning models.

[0070] To verify the effectiveness and adaptability of the dominant instability mode identification method proposed in this invention, the 36-bus system of the China Electric Power Research Institute (CEPRI-36) was selected as a test case. Its specific architecture is as follows: Figure 4 As shown in the figure. In this embodiment, the load model of the CEPRI-36 node system is set as a combination model of constant impedance and induction motor, in which the proportion of induction motor has three different cases: 30%, 50%, and 70%; the main fault studied in the system is a three-phase short-circuit ground fault; there are three different fault locations: 2%, 50%, and 98%; there are 26 different fault lines, mainly the twenty-six AC transmission lines of the CEPRI-36 node system; and the faults will last for three different durations: 0.20s, 0.25s, and 0.30s, which will cause different degrees of instability consequences.

[0071] Following the above proportions, a total of 2106 samples were obtained, with each power angle sample containing 8 curves and each voltage sample containing 22 curves, totaling 63180 curves. Among these, the number of stable, power angle-dominated instability, and voltage-dominated instability samples were 938, 418, and 750, respectively. Schematic diagrams of the obtained instability sample curves for different types are shown below. Figure 5 , Figure 6 As shown in Table 1, the specific curve generation schemes are as follows.

[0072] Table 1

[0073] The obtained sample curves are processed sequentially according to the method steps provided in this invention, and then further trained according to the dominant instability mode identification model to obtain the following results. Figure 7 The confusion matrix shown in the figure has values ​​of 0, 1, and 2 for the stable class, power angle-dominated instability class, and voltage-dominated instability class, respectively.

[0074] From the confusion matrix results, the overall performance of identifying the three power system operating states—stable, power angle-dominated instability, and voltage-dominated instability—is particularly outstanding. For the true stable state, the correct identification rate reached 95.5%, with only 3.6% of samples misclassified as stable and 0.9% as power angle-dominated instability, indicating high accuracy in determining the system's stable state, with only a few marginal samples showing slight confusion. For the true power angle-dominated instability, the correct identification rate further improved to 97.6%, with only 2.4% of samples misclassified as voltage-dominated instability, demonstrating the invention's good feature capture capability for power angle instability. For the true voltage-dominated instability, 100% perfect identification was achieved, with no samples misclassified to other categories, indicating that the features of voltage-dominated instability are fully identified in the intelligent identification model for grid-dominated instability modes.

[0075] From the distribution of misclassifications, all misclassifications were concentrated within the same state, with no misclassification from unstable to stable states. The main misclassification was power angle instability being mistaken for voltage instability. This means that the present invention can maintain accurate judgment in the primary identification of stability and instability, and can meet the high requirements for state recognition in real-world scenarios.

[0076] To verify the effectiveness of the basic model used in this invention, different deep learning models were used to identify the dominant instability mode on the same input data. A time-series identification model and a two-dimensional image classification model were selected for thorough comparison and verification. The time-series identification models included LSTM and RF models; the two-dimensional image classification models included VGG and ConvNeXt models. The data used for the former was the multidimensional time-series data obtained in steps S1-S5, and the data used for the latter was the two-dimensional color GADF image data obtained in steps S1-S8. The specific results of the comparison and verification of different models are shown in Table 2. Figure 8 As shown, the performance metrics of each model are quantitatively evaluated using accuracy, precision, recall, and F1 score.

[0077] Table 2

[0078] In the comparison of time series identification models, the MobileViT model used in this invention demonstrates significant advantages. Compared to RF and LSTM network models in time series identification, the model used in this invention outperforms them in various metrics. Although the RF model shows some stability in small-sample time series data identification tasks due to the characteristics of traditional machine learning models, it still falls short of the overall performance of the model in this invention. Furthermore, the LSTM network model's various metrics are all below 92%, and its insufficient feature extraction ability and generalization performance further highlight the superiority of the model in this invention in time series data identification scenarios.

[0079] In the comparison of two-dimensional image classification models, the model used in this invention also holds a leading position. Compared to the VGG and ConvNeXt models, the core metrics of this invention's model are comprehensively superior. The VGG model improves its feature extraction capability by deepening the number of network layers, but its F1 score is only 96.24%. The ConvNeXt model, with its superior network structure design, improves its accuracy to 97.47% and its F1 score to 97.30%, which, although close to the performance of this invention's model, still lags behind. The various metrics of the MobileViT model fully demonstrate its advantages over new convolutional networks in image feature mining.

[0080] The above comparison results fully verify the effectiveness of the model of this invention. Compared with traditional time series identification models, the model of this invention can more accurately capture the feature information of power system instability processes, avoiding the problem of insufficient feature extraction in time series models. Compared with mainstream two-dimensional image models, the model of this invention combines convolutional branches and Transformer branches, which can fully focus on the global and local features of the data. Combining the advantages of various performance indicators, the model of this invention can efficiently and accurately complete the task of identifying the dominant instability mode, and has good practical application value.

[0081] To verify the improved K-means algorithm image encoding process through feature engineering and the optimized gain of the model used in this invention, four progressive models were constructed to conduct comparative experiments on dominant instability mode identification. M1: The original time-series data was directly input into the basic 1D-CNN model without any optimization strategy. M2: Based on M1, feature engineering and an improved K-means algorithm were incorporated to select samples, still using the 1D-CNN model. M3: Based on the optimized samples of M2, GADF image encoding technology was introduced to convert the time-series data into two-dimensional image data, and a 2D-CNN model was used. M4: The model proposed in this invention was used. The identification performance of each model was then quantified using four core indicators: accuracy, precision, recall, and F1 score, and the effectiveness of each model was systematically analyzed.

[0082] The reasons for setting each model are shown in Table 3.

[0083] Table 3

[0084] The specific comparison and validation index results of different models are shown in Table 4. Figure 9 As shown.

[0085] Table 4

[0086] Compared to M1, M2 shows an improvement of 0.34% in accuracy, 1.27% in precision, and 0.22% in F1 score. This result demonstrates that feature engineering effectively uncovers correlated features in the data, while the improved K-means algorithm reduces the interference of redundant data on the model. The combination of these two approaches achieves positive optimization of model performance, validating the effectiveness of this feature engineering strategy in improving identification accuracy.

[0087] Compared to M2, M4 introduces the GADF image encoding process, transforming time-series data into two-dimensional image data that is easier to extract deep features. Its accuracy jumps to 97.78%, with significant improvements across all four core metrics. This demonstrates that GADF image encoding can present the implicit features in time-series data in a visual image format, better aligning with the feature extraction mechanism of deep learning models. It significantly enhances the model's ability to identify instability patterns, fully validating the optimization value of this image encoding process.

[0088] Compared to M3: M3, which uses GADF image encoding combined with a 2D-CNN model, achieved an accuracy of 96.20% and an F1 score of 95.78%, demonstrating excellent recognition performance. M4, the model proposed in this invention, is a further optimization of M3, achieving significant improvements across all metrics, with all core indicators reaching the best levels observed in the experiments. This result demonstrates that the improvements in network structure design and feature fusion strategy in this invention can further extract key discriminative information from image data, effectively compensating for the performance shortcomings of traditional 2D-CNN models.

[0089] The results of the aforementioned progressive comparative experiments fully validate the effectiveness of the model in this invention. Feature engineering and the improved K-means algorithm fundamentally improved model performance, while the GADF image encoding process achieved a significant leap in model performance. Furthermore, the model in this invention, while inheriting the advantages of previous optimization strategies, significantly outperforms the comparative models in all core metrics. This result demonstrates that this invention can more accurately and efficiently complete the task of identifying the dominant instability modes of power systems, possessing high practical application value.

[0090] To verify the robustness of the model in noise-perturbed scenarios, three different perturbation ranges were designed: full-amplitude perturbation, center perturbation, and edge perturbation. For each type of perturbation, three different perturbation levels were set: α = 0.1, 0.2, and 0.3, respectively. The results are as follows: Figure 10 The diagram shows a noise perturbation. After processing the data obtained in the previous steps using the above perturbation method, the data is input into the dominant instability mode framework obtained in step 6. The training results are shown in Table 5. Figure 11 The diagram shows the effects of different disturbances. Then, using four numerical values ​​as quantitative indicators and setting a baseline accuracy of 95%, the model's performance in identifying dominant instability modes under complex noise environments was comprehensively evaluated. The experimental results fully reflect the model's adaptability to different noise interferences.

[0091] Table 5

[0092] In scenarios with full-amplitude perturbation, the average performance indicators of the model in this invention are significantly higher than the benchmark, fully demonstrating its good adaptability to full-amplitude noise, and its core identification ability is not affected by the strength of the perturbation. In scenarios with central perturbation, the average performance of the model is also consistently above the benchmark. Although the lowest accuracy of 96.52% occurs at α=0.1, it is still 1.52 percentage points higher than the benchmark and does not exceed the anti-interference threshold. As the degree of perturbation increases, the indicators gradually recover to 97.15%, reflecting the model's self-adjustment ability to noise in the central region, accurately focusing on core features and maintaining stable extraction performance when perturbation intensifies. In scenarios with edge perturbation, the average performance indicators of the model are close to the level of no perturbation, and the accuracy is stable above 97.15% in the range of α=0.1 to 0.3. It can be seen that noise perturbation does not interfere with the core features of the unstable mode, and the model continues to output reliable results.

[0093] Overall, the model's average performance across all perturbation scenarios significantly exceeds the 95th percentile benchmark, with metrics consistently above 96%, without substantial decline due to noise. Performance under full-amplitude and edge-based strong perturbations is closer to or surpasses the unperturbed state, validating the model's effectiveness and robustness, demonstrating its ability to effectively filter interference and accurately capture core features.

[0094] The method proposed in this invention can effectively improve the shortcomings of traditional methods in identifying dominant instability modes. It can quickly and accurately analyze and identify a large number of transient fault samples in power systems, which helps researchers formulate subsequent safety and stability control strategies to ensure the safe and stable operation of power systems.

Claims

1. A method for identifying the dominant instability mode of a power grid considering feature engineering and image upscaling, characterized in that, include: Obtain initial electrical quantity characteristics and time-series curves of power grid fault sample inputs; Based on the initial electrical quantity characteristics and the input time series curves of power grid fault samples, an operation state feature supplementary sequence considering the static and dynamic characteristics in the operation state is constructed; Construct a power system topology feature supplementary sequence that considers the spatial adjacency matrix of the power system; A fusion feature index considering both the operating state information and topological structure information of power system instability samples is constructed based on the improved PageRank algorithm; The supplementary sequences of operating status features, supplementary sequences of power system topology features, and fusion feature indicators are combined and added to the initially acquired power grid fault sample input time series curve to obtain the power grid fault sample input time series curve considering feature engineering. Construct a dynamic fluctuation weight index to measure the degree of dynamic fluctuation of the curve; Based on the dynamic fluctuation weight index, the input time series curves of power grid fault samples with fluctuation levels higher than the preset value, different modes, and representing different fluctuation characteristics are selected to consider the feature engineering, and the selected input time series curves are obtained. The filtered input time series curves are encoded into two-dimensional images based on nonlinear image upscaling technology; Construct a pre-trained intelligent identification model for the dominant instability modes of the power grid; The two-dimensional image is input into the pre-trained intelligent identification model of the dominant instability mode of the power grid to obtain the identification result of the dominant instability mode.

2. The method according to claim 1, characterized in that, The static characteristics in the operating state include: the mean, variance, maximum, minimum, range, skewness, and kurtosis of the input time-series curves of power grid fault samples, and their expressions are as follows: In the formula, Represents the mean of the curve. This represents the total number of sampling times. Indicates the time index. Indicates time Electrical values ​​at that time; Represents the variance of the curve. Indicates the maximum value. Indicates the minimum value. Indicates time Electrical values ​​at that time; Indicates the range. Indicates skewness, Indicates the standard deviation of the curve; Indicates kurtosis.

3. The method according to claim 1, characterized in that, Dynamic characteristics include the mean of the first derivative of the input time-series curve of power grid fault samples, the number of inflection points, and the number of abrupt changes, expressed as follows: In the formula, The first derivative means. express Electrical value at any given time Indicates the number of inflection points. Indicates the curve at The second difference at time, For indicator functions, Indicates the number of mutations. This represents the mutation threshold.

4. The method according to claim 1, characterized in that, The topological characteristics of a power system considering its spatial adjacency matrix include: The maximum difference between two adjacent buses at the same moment is expressed as: In the formula, Indicates busbar With busbar The difference at the same time, Indicates busbar At any moment The electrical quantity below, busbar At any moment The electrical quantities below; This represents the maximum difference between two adjacent buses at the same time. Indicates the number of adjacent buses; The voltage variation trend of a single node and its adjacent bus is expressed as follows: In the formula, Indicates adjacent busbars With adjacent busbar The Pearson correlation coefficient between them Indicates adjacent busbars The mean of the corresponding input timing curve, Indicates adjacent busbars The mean of the corresponding input timing curve, It indicates the voltage change trend of a single node and its adjacent bus.

5. The method according to claim 4, characterized in that, The specific method for constructing a fusion feature index based on the improved PageRank algorithm, which considers both the operating state information and topological structure information of power system instability samples, is as follows: The expression for correcting the element values ​​in the node link matrix of a power system is as follows: In the formula, This indicates that the starting node in the link matrix is... The termination node is The element value, For nodes Side and node The absolute difference in active power on both sides For all nodes The sum of power changes in connected lines; Based on the corrected node link matrix, the fusion feature index is calculated, and its expression is as follows: In the formula, Represents a node The corresponding fusion feature index, The damping factor, This represents the corrected node link matrix. Represents nodes The total number of connected nodes. For nodes One of the connected nodes, It is the identity matrix. This represents the total number of nodes.

6. The method according to claim 5, characterized in that, The specific method for constructing a dynamic volatility weighted index to measure the degree of dynamic volatility of a curve is as follows: The mean of the first derivative of the curve is defined as the drastic index; Define the curve deviation exponent, and its expression is: In the formula, This indicates that the curve deviates from the exponent. Indicates the maximum value of the electrical quantity. Indicates the minimum value of an electrical quantity; A dynamic volatility weighted index is constructed based on the volatility index and the curve deviation index, and its expression is as follows: In the formula, This represents the initial dynamic fluctuation weighting index of the curve. Indicates the final dynamic fluctuation weight of the curve; index; Indicates the intensity index. This represents the total number of individual sample power angle curves or voltage curves. This represents the initial dynamic fluctuation weight index of the Nth curve.

7. The method according to claim 6, characterized in that, Based on the dynamic volatility weighting index, curves with volatility levels higher than a preset value, different patterns, and representing different volatility characteristics are selected to obtain the filtered input time series curves. The specific method is as follows: The dynamic fluctuation weight index is calculated based on the input time series curve of the power grid fault sample considering characteristic engineering, according to the dynamic fluctuation weight index. The first k curves are selected as the fluctuation centers based on the dynamic fluctuation weight index; The DTW distance between the input time series curve of each power grid fault sample considering characteristic engineering and the fluctuation center is calculated based on the dynamic time warping algorithm. Based on the obtained DTW distance, the curve sets are filtered and classified into different clusters; Select the cluster center curves of different curve sets and remove the remaining invalid curves to obtain the filtered input timing curves.

8. The method according to claim 7, characterized in that, The specific method for calculating DTW distance is as follows: Calculate the input time-series curve of a power grid fault sample considering characteristic engineering and the fundamental distance matrix that serves as the fluctuation center curve; Find the optimal alignment path based on the fundamental distance matrix; Calculate the cumulative distance from the optimal alignment path to the final alignment point; The cumulative distance is corrected using a dynamic fluctuation weighting exponent to obtain the final DTW distance, which is expressed as follows: In the formula, This indicates the output DTW distance. This represents the cumulative distance from the optimal alignment path to the final alignment point.

9. The method according to claim 1, characterized in that, The intelligent identification model for the dominant instability mode of the power grid is the MobileViT model, which has spatial convolution and transformer mechanisms.

10. A computer-readable storage medium, characterized in that, The device stores a computer program, which, when executed by a processor, causes the processor to perform the steps of the power grid dominant instability mode identification method considering feature engineering and image dimensionality enhancement as described in any one of claims 1 to 9.

Images