An Online Update Method for a Parallel Local Model of a Power System

By using a matching degree index and Generative Adversarial Network (GAN) approach, target samples for parallel local models are generated efficiently, solving the problems of long processing time and insufficient accuracy in existing technologies, and achieving fast and accurate online updates and performance improvements.

CN115563866BActive Publication Date: 2026-11-17TSINGHUA UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211209816.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-09-30
Publication Date
2026-11-17
Estimated Expiration
2042-09-30

AI Technical Summary

Technical Problem

Existing online update methods for parallel local models are time-consuming and struggle to efficiently improve the performance of machine learning models in key areas, especially in transient voltage safety analysis where they are prone to insufficient accuracy.

Method used

By determining update requirements based on matching degree metrics, and combining softened samples with Generative Adversarial Networks (GANs) to efficiently generate target samples, the performance of machine learning models in key areas is improved.

Benefits of technology

It significantly reduces the time and computational complexity of updating parallel local models, improves the generalization error of machine learning models in key areas, and meets the real-time requirements of online transient voltage safety analysis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115563866B_ABST
    Figure CN115563866B_ABST
Patent Text Reader

Abstract

This invention provides an online update method for a parallel local model of a power system, comprising: (1) determining whether an update is needed at a certain time based on a matching degree index; (2) determining the update method of the parallel local model for a certain operating scenario to be evaluated; and (3) a key sampling method to improve the performance of machine learning models in key areas and an efficient sample supplementation method based on softened samples and Generative Adversarial Networks (GANs). When the machine learning evaluation model is applied online and needs to be updated, this method can improve the sample generation efficiency, thereby accelerating the online update efficiency of the data-driven model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of online updating of parallel local models, specifically a method for online updating of parallel local models of power systems. Background Technology

[0002] Existing data-driven models are typically updated periodically or through manual commands when used in online applications.

[0003] The update modes of the parallel local model include periodic updates and on-demand updates. Periodic updates are performed at fixed time intervals. To update the parallel local model, for example This update mode is a proactive update, enabling the parallel local model to adapt to changes in the online operation scenario of the power grid, avoiding the situation where the parallel local model becomes invalid due to the accumulation of small changes in the operation scenario. On-demand updates, on the other hand, update the parallel local model when its accuracy is insufficient, and are therefore a passive update.

[0004] There are two ways to update the parallel local model: (1) select a local model from the parallel local models for updating; (2) create a new local model and then add it to the parallel local model.

[0005] If the power grid operation scenario to be evaluated is within the effective range of the parallel local model, no update is needed. If the operation scenario to be evaluated is outside the effective range of the parallel local model but is close to the effective area, samples are generated based on the operation scenario according to the steps in Table 1, and a local model is selected from the parallel local models for updating. If the operation scenario to be evaluated is far from the effective area of ​​the parallel local model, samples are generated, a new local model is created, and then added to the parallel local model. In contrast, existing machine learning models always require updates based on the complete sample set, resulting in high computational complexity.

[0006] Table 1. Steps for generating samples based on numerical simulation

[0007] Numerous case studies show that over 95% of the time spent updating the parallel local model is used for generating samples based on numerical simulation.

[0008] With limited training samples, it's difficult to improve the overall performance of machine learning models. In transient voltage safety analysis, more attention is usually paid to a specific set of cases. For example, in transient voltage safety assessment, the accuracy of instability assessment is more important, as the harm of missing instability cases is far greater than the harm of misjudging them. In reactive power optimization with transient voltage safety constraints, the assessment accuracy near the transient voltage safety boundary is more important, because the assessment results directly affect the solution to the reactive power optimization problem.

[0009] Therefore, a key area can be designated in advance. In building and training machine learning models, the focus is on Internal performance.

[0010] To improve the performance of machine learning models To improve performance, a natural idea is to increase the number of training samples. However, due to... It only represents a portion of the input feature space. While random sampling within the input feature space can improve the overall performance of a machine learning model, it may not be suitable for all users. The efficiency of improving internal performance is relatively low.

[0011] In order to improve machine learning models in key areas Internal performance can be increased The samples within. However It is often difficult to express explicitly, making it challenging to supplement target samples. For example, if The TVSI index is greater than the threshold. In the region, The expression is: (9) in, Input features The corresponding TVSI metric.

[0012] Obviously, It is difficult to express visually. If we use random sampling within the input feature space and then only retain... Using in-sampling methods to generate target samples is inefficient. Therefore, a more efficient method for generating target samples is needed to fully realize the value of focusing sampling in improving the performance of machine learning models in key regions.

[0013] Currently, there is research on generating target samples based on GANs. This method believes that... The input features follow a certain distribution, and GANs learn this distribution. However, if the initial sample set has few target samples, GANs may not learn it well. Distribution of internal input features. Summary of the Invention

[0014] The purpose of this invention is to improve the efficiency of sample generation when machine learning evaluation models are applied online and need to be updated, thereby accelerating the efficiency of online updates of data-driven models.

[0015] An online update method for a parallel local model of a power system includes: 1) Determine whether an update is needed at a given time based on the matching degree metric; 2) For a given operational scenario to be evaluated, determine the update method of the parallel local model; 3) Key sampling methods to improve the performance of machine learning models in key areas, and efficient sample supplementation methods based on softened samples and Generative Adversarial Networks (GANs).

[0016] Wherein, 1) includes the following steps: Establish and train a parallel local model, and set a matching degree threshold. ; Read the power grid operation scenario to be evaluated; Based on equation (1), the matching degree index between each local model and the operating scenario to be evaluated is calculated. ; For local model Corresponding subsample set Assuming Input features of medium-sized samples Follows distribution To the PCA is performed on the matrix composed of input features from the middle samples to obtain the feature vector matrix. Assume the input characteristics of the power grid operation scenario to be evaluated are as follows: Then the operating scenario is the same as Matching degree between The calculation expression is: (1) in, It is a number Local model The front of the corresponding eigenvector matrix l List, The operating scenario to be evaluated and The degree of matching between them.

[0017] Based on equation (2), the local model with the highest matching degree to the operating scenario to be evaluated is obtained, and the matching degree index of this local model is denoted as follows: ; The local model number that best matches the operating scenario to be evaluated is: (2) in, It is the local model number that best matches the operating scenario to be evaluated.

[0018] serial number Local model This refers to the local model matched to the operating scenario to be evaluated.

[0019] if If so, the parallel local model needs to be updated; otherwise, no update is required.

[0020] Wherein, 2) includes the following steps: Define a subsample set Average similarity of internal running scenarios The expression is as follows: (3) in, , , (4) It is the TVSI index number The weight coefficients of each component are set by default. ; yes The Middle TVSI metric matrix for each operating scenario yes The center yes The number of samples in the middle; That is, the subset of samples The average similarity of internal running scenarios, its physical meaning is: Internal running scenarios and Average similarity between centers; Establish and train a parallel local model, and calculate the center of each sub-sample set based on equations (5) and (3) respectively. Average similarity of running scenarios within a subset of samples ; Define a subsample set center The expression is as follows: (5) in, It is the first The subset of samples, which is related to the first subset. Local Model correspond, yes The Middle TVSI metric matrix for each operating scenario yes The number of samples in the middle. That is, the subset of samples The center, its physical meaning is The average value of the TVSI index matrix for the internal operating scenario.

[0021] Read the power grid operation scenario to be evaluated; TVSI index matrix for the operational scenario to be evaluated is calculated based on numerical simulation. ; Calculate the center of the operating scenario to be evaluated and each subset of samples. similarity between The expression is as follows: (6) in, , , (7) It is the TVSI index number The weight coefficients of each component are set by default. ; The operational scenario and number to be evaluated The center of the subset The similarity between them; Identify the scenario that best matches the operational scenario to be evaluated. The number is calculated using the following expression: (8) in, It is a number The average similarity of the running scenarios within a subset of samples; It is the local model with the highest matching degree to the operating scenario to be evaluated. The number; if If so, create a new local model; otherwise, update the local model. ; The time required to update the parallel local model is similar to the time required to build and train a local model, typically requiring 5 to 15 minutes, which is longer than the time interval for online transient voltage safety analysis. Therefore, in practical applications, the accuracy of the parallel local model may be insufficient when performing transient voltage safety analysis, specifically manifested as follows: In this case, the parallel local model is still used to calculate the TVSI index, but a low-precision alarm is also given.

[0022] Specifically, the key sampling method for improving the performance of machine learning models in key areas mentioned in 3) includes: For local model Assuming its input feature space is The key areas are , If added in For samples with a high density and uniform distribution, then exist The generalization error within the range is reduced.

[0023] The efficient sample supplementation method based on softened samples and Generative Adversarial Networks (GANs) in step 3) includes the following steps: Obtain the initial sample set; Define key areas The mathematical expression is extracted from the initial sample set. The samples within are used to obtain the initial target sample set. ; Based on the initial target sample set Generate softened sample set ; based on Training a GAN; Use GANs to generate a sample set to be confirmed. ; Confirmation Sample Set Numerical simulations were performed on the samples to calculate the TVSI index of the samples, based on whether the samples were in the range. Inner handle Divided into true target sample set and fake target sample set ; Bundle Add to the initial sample set.

[0024] Technical effects of the present invention: The key sampling method proposed here can achieve the expected effect of reducing the generalization error of the local model in key regions of its input feature space.

[0025] Compared to methods based on random sampling, the advantage of the sample efficient supplementation method based on softened samples and Generative Adversarial Networks (GANs) lies in the unconfirmed sample set. The proportion of target samples is higher, requiring fewer numerical simulations to generate the same number of target samples, thus reducing the time consumed. This improves the efficiency of generating target samples, thereby enhancing the efficiency of updating the parallel local model. Attached Figure Description

[0026] The accompanying drawings illustrate various embodiments generally by way of example rather than limitation, and are used, together with the specification and claims, to explain embodiments of the invention. Where appropriate, the same reference numerals are used in all drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the apparatus or method.

[0027] Figure 1A schematic diagram of the present invention is shown, illustrating the efficient sample supplementation method based on softened samples and Generative Adversarial Network (GAN) for generating target samples. Figure 2 (a) shows an example 2 of the present invention. A schematic diagram of the sample distribution within each value interval. Figure 2 (b) illustrates Example 2 of the present invention. A schematic diagram of the mean square error within each value interval; Figure 3 Example 3 of the present invention is shown. A schematic diagram of the sample distribution for each value interval. Detailed Implementation

[0028] A schematic diagram illustrating the efficient sample completion method for generating target samples based on softened samples and Generative Adversarial Networks (GANs) is shown below. Figure 1 As shown in the figure, the initial sample set was obtained using a random sampling method, with target samples accounting for only 5%. After training the GAN with softened samples from the target samples in the initial sample set, the proportion of target samples in the samples to be confirmed generated by the GAN is approximately 50%, which is about 10 times the proportion of target samples in the samples obtained using the random sampling method. If 1000 target samples need to be generated, approximately 20,000 simulations are required using the random sampling method, while the method in this paper only requires approximately 2000 simulations, reducing the time spent on numerical simulation by 90%.

[0029] Case Analysis Example 1: Update of a parallel local model This example uses the IEEE 39-node model. The considered faults are shown in Table 2, the considered power grid topology parameters are shown in Table 3, and the considered typical power grid operating states are shown in Table 4. The initial number of operating scenarios is... The process of establishing and training the parallel local model will not be elaborated here. Assume there are 5 running scenarios to be evaluated, and their parameter configurations are shown in Table 5.

[0030] Table 2 Fault Information

[0031] Table 3 Power Grid Topology Parameter Information

[0032] Table 4 Typical Power Grid Operation Status Information

[0033] Table 5 Information on the operational scenarios to be evaluated

[0034] For operating scenarios 1, 2, and 3, the faults and grid topology parameters are in the initial operating scenario set, and the grid operating status is also covered by the sampling range. Therefore, it is preliminarily determined that updating the parallel local model is not required. For operating scenario 4, the faulty line is not in the initial operating scenario set; for operating scenario 5, the topology parameters are not in the initial operating scenario set. It is preliminarily determined that updating the parallel local model may be necessary for these two operating scenarios. The update determination results for the above five operating scenarios are shown in Table 6.

[0035] Table 6 Update determination results for each operating scenario

[0036] With the matching accuracy threshold set to 0.5, scenarios 1, 2, and 3 do not require updating the parallel local model, while scenarios 4 and 5 do, consistent with the initial assessment. Below, we will update the parallel local model for scenarios 4 and 5.

[0037] Calculate the average similarity of the running scenarios within each subset of samples. And the similarity between operating scenarios 4 and 5 and the centers of each subset. The results show that a new local model needs to be created for each of the two operating scenarios, and then added to the parallel local model.

[0038] Generating 5000 samples for each of scenarios 4 and 5 took 198.1 seconds. Subsequently, training two newly created local models took 6.03 seconds; the training process is not detailed here. Including other data processing time, updating the parallel local models took a total of 207.6 seconds. This demonstrates that updating the parallel local models is cost-effective, requires fewer training samples, and is fast.

[0039] Example 2: A key sampling method to improve the performance of parallel local models in key areas This example uses the IEEE 39-node model. It considers a three-phase N-1 fault on line "Bus15-Bus16", with a fault duration of 250ms. It takes into account the initial grid topology parameters and grid operating state, and initially sets the number of operating scenarios to 1. The sample input feature dimension is 131, where... The key areas are The corresponding region is expressed as: (10) in, This is a key area.

[0040] The total sample size is 5000, of which 2500 are randomly selected from the input feature space, and the other 2500 are selected from key regions. The samples are in... The distribution within each value interval is as follows Figure 2 As shown in (a), the key areas are... The sample density within is higher than Other value ranges. Output of the parallel local model after training. The mean square error within each value interval is as follows: Figure 2 As shown in (b), the parallel local model outputs... The mean square error is smaller in key areas, and The mean squared error within each value interval is roughly inversely proportional to the sample size within that interval. This indicates that the proposed key sampling method can achieve the expected results and is effective.

[0041] Example 3: An efficient sample supplementation method based on softened samples and GAN This example uses the Shaanxi power grid model. It considers a three-phase N-1 fault on line "3008 Yudang 1", with a fault duration of 150ms. Initial power grid topology parameters and operating states are considered. The initial number of operating scenarios is 1. The sample input feature dimension is 1520, where... The key areas are The corresponding region is expressed as: (11) in, This is a key area.

[0042] Assuming the initial sample set already contains 250 target samples, and 1000 more target samples need to be generated. The initial sample set contains... The distribution within each value interval is as follows Figure 3 As shown, the target sample accounts for only about 5% of the initial sample. If the target sample is generated based on a random sampling method, it would require approximately 20,000 simulations, making the generation of the target sample very inefficient.

[0043] Table 7 compares the time consumption and efficiency of generating 1000 target samples using random sampling, generative adversarial networks (GANs), and our proposed method. It is evident that compared to random sampling, our proposed method reduces time by 87.5% and increases the proportion of target samples by 8.1 times; compared to GANs, it reduces time by 30.7% and increases the proportion of target samples by 53.2%. Therefore, the proposed sample supplementation method based on softened samples and GANs is more efficient in supplementing target samples, achieving the expected results.

[0044] Table 7 Comparison of time and efficiency in generating target samples

[0045] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An online update method for a parallel local model of a power system, characterized in that, include: 1) Determine whether an update is needed at a given time based on the matching degree metric; 2) For a given operational scenario to be evaluated, determine the update method of the parallel local model; 3) Key sampling methods to improve the performance of machine learning models in key areas, and efficient sample supplementation methods based on softened samples and Generative Adversarial Networks (GANs); The first step includes the following steps: Establish and train a parallel local model, and set a matching degree threshold. ; Read the power grid operation scenario to be evaluated; Based on equation (1), the matching degree index between each local model and the operating scenario to be evaluated is calculated. ; For local model Corresponding subsample set Assuming Input features of medium-sized samples Follows distribution , for PCA is performed on the matrix composed of input features from the middle samples to obtain the feature vector matrix. Assuming the input characteristics of the power grid operation scenario to be evaluated are: Then the operating scenario is the same as Matching degree between The calculation expression is: (1) in, It is a number Local model The front of the corresponding eigenvector matrix l List, The operating scenario to be evaluated and The degree of matching between them; Based on equation (2), the local model with the highest matching degree to the operating scenario to be evaluated is obtained, and the matching degree index of this local model is denoted as follows: ; The local model number that best matches the operating scenario to be evaluated is: (2) in, It is the local model number that best matches the operating scenario to be evaluated; serial number Local model This refers to the local model matched to the operating scenario to be evaluated; if If so, update the parallel local model; otherwise, do not update. The second step includes the following steps: Define a subsample set Average similarity of internal running scenarios The expression is as follows: (3) in, , , (4) It is the TVSI index number The weight coefficients of each component are set by default. ; yes The Middle TVSI metric matrix for each operating scenario yes The center, its physical meaning is The average value of the TVSI index matrix for in-system operation scenarios. It is the first The subset of samples, which is related to the first subset. Local Model correspond, yes The number of samples in the middle; That is, the subset of samples The average similarity of internal running scenarios, its physical meaning is: Internal running scenarios and Average similarity between centers; Establish and train a parallel local model, and calculate the center of each sub-sample set based on equations (5) and (3) respectively. Average similarity of running scenarios within a subset of samples ; Define a subsample set center The expression is as follows: (5) Read the power grid operation scenario to be evaluated; TVSI index matrix of the operating scenario to be evaluated is calculated based on numerical simulation. ; Calculate the center of the operating scenario to be evaluated and each subset of samples. Similarity between The expression is as follows: (6) in, , , (7) It is the TVSI index number The weight coefficients of each component are set by default. ; The operational scenario and number to be evaluated The center of the subset The similarity between them; Identify the scenario that best matches the operational scenario to be evaluated. The number is calculated using the following expression: (8) in, It is a number The average similarity of the running scenarios within a subset of samples; It is the local model with the highest matching degree to the operating scenario to be evaluated. The number; if If so, create a new local model; otherwise, update the local model. ; The key sampling method for improving the performance of machine learning models in key areas mentioned in 3) specifically includes: For local model Assuming its input feature space is Key areas are , If added in For samples with a high density and uniform distribution, then exist The generalization error within the range is reduced; The efficient sample supplementation method based on softened samples and Generative Adversarial Networks (GANs) in section 3) includes the following steps: Obtain the initial sample set; Define key areas The mathematical expression: (9) in, Input features The corresponding TVSI index, The threshold for the TVSI index; Extract from the initial sample set The samples within are used to obtain the initial target sample set. ; Based on the initial target sample set Generate softened sample set ; based on Training a GAN; Use GANs to generate a sample set to be confirmed. ; Confirmation Sample Set Numerical simulations were performed on the samples to calculate the TVSI index of the samples, based on whether the samples were in the range. Inner handle Divided into true target sample set and fake target sample set ; Bundle Add to the initial sample set.