Multi-task transient voltage stability evaluation method based on TVSDTNet feature enhancement

Abstract

The invention discloses a multi-task transient voltage stability evaluation method based on TVSDTNet feature enhancement. The multi-task transient voltage stability evaluation method comprises the following steps: constructing time sequence electrical measurement data information; constructing a multi-task transient voltage stability evaluation model based on TVSDTNet feature enhancement; dividing the transient voltage stability evaluation problem into two mutually associated sub-tasks, namely a task I and a task II; defining a voltage instability node complete coverage rate fcr, a voltage stabilization node complete coverage rate tcr and a geometric mean value Gm of the voltage instability node complete coverage rate fcr and the voltage stabilization node complete coverage rate tcr so as to evaluate the capability of the multi-task transient voltage stability evaluation model for accurately identifying and dividing transient voltage instability nodes or regions; and based on the constructed multi-task transient voltage stability evaluation model, evaluation of the transient voltage stability state and accurate identification and division of voltage instability nodes or regions are realized. According to the method, expression of time sequence electrical measurement data information can be effectively enhanced, transient voltage stability evaluation of the system is realized, and voltage instability nodes or regions can be accurately divided.

Description

Technical Field

[0001] The present invention relates to the technical field of power system transient voltage stability assessment, and in particular to a multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement. Background Art

[0002] In the current power system, the penetration rate of new energy sources such as wind power and photovoltaics continues to rise, and the resulting uncertainty and randomness in system operation are also gradually increasing, which places higher demands on the assessment and protection of transient voltage stability. However, traditional transient voltage stability assessment methods usually rely on high-dimensional time-series electrical measurement data. Such data are inevitably subject to noise interference, causing some effective information to be obscured or distorted, thereby affecting the accuracy and stability of the assessment. Therefore, there is an urgent need for an assessment method that can effectively reduce the noise of high-dimensional measurement data and simultaneously extract and enhance potential features, so as to improve the overall transient voltage stability assessment performance under complex working conditions and further improve the ability to accurately identify unstable nodes or areas.

[0003] Currently, most research on transient voltage stability focuses on mining and utilizing time-series electrical measurement data, but lacks comprehensive noise reduction mechanisms. As noise levels increase, the model's assessment accuracy and stability inevitably suffer. While some studies have introduced attention mechanisms or diffusion models based on fixed convolution kernels to suppress noise, their adaptive control capabilities for multi-scale noise characteristics remain insufficient, making it difficult to achieve both stable and accurate assessment results in high-noise scenarios.

[0004] In addition, existing research lacks complete quantitative indicators to evaluate the accuracy and reliability of the division results when dividing transient voltage instability nodes or areas, resulting in much room for improvement in the accuracy and practicality of unstable area identification. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement, which can effectively enhance the expression of time-series electrical measurement data information, realize system transient voltage stability assessment, and accurately divide voltage instability nodes or areas.

[0006] The technical solution adopted by the present invention is:

[0007] The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement includes the following steps:

[0008] Step 1: Construct time-series electrical measurement data information;

[0009] Step 2: Construct a multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement;

[0010] Step 3: Split the transient voltage stability assessment problem into two interrelated subtasks: Task 1: Determine whether the power system has transient voltage instability under specific operating conditions; Task 2: If the system is determined to have transient voltage instability by Task 1, further identify the node or area where voltage instability has occurred.

[0011] Step 4: Define the voltage instability node complete coverage rate f cr , voltage stability node full coverage rate t cr and their geometric mean G m , in order to evaluate the ability of the multi-task transient voltage stability assessment model to accurately identify and divide transient voltage instability nodes or areas;

[0012] Step 5: Based on the multi-task transient voltage stability assessment model constructed in step 2, the transient voltage stability state is assessed, and the voltage instability nodes or areas are accurately identified and divided.

[0013] In step 1, a comprehensive study is conducted on the transient voltage stability problem, and the sampling characteristics of the time-series electrical measurement data information in actual engineering are considered to construct a comprehensive and integrated time-series electrical measurement data information. The time-series electrical measurement data information characteristics include the bus voltage amplitude and phase angle at the bus level; the AC branch active power and reactive power at the branch level; and the generator power angle, active power, reactive power, and angular velocity at the generator unit level.

[0014] In step 2, the multi-task transient voltage stability assessment model includes:

[0015] Single-step diffusion autoencoder OSD-AE is used to extract and enhance the time series electrical measurement data information and generate the latent feature vector z;

[0016] The single-step diffusion self-decoder OSD-Dec is used for the reconstruction of time-series electrical measurement data information, using a mean square error loss function to constrain the reconstructability of latent features;

[0017] Transformer attention network, used for classification tasks, uses a focal loss function to constrain the model's attention to difficult-to-classify samples;

[0018] Construct a weighted loss based on the mean square error loss function and the focal loss function as the comprehensive optimization goal in the model parameter adjustment process;

[0019] Selective Kernel Convolutional Network (SKNet), used to capture signal features at different levels;

[0020] The softmax classification output layer converts the output of the multi-task transient voltage stability assessment model into a probability value between 0 and 1, thereby representing the probability of system transient voltage stability and the stability of each busbar node voltage under system voltage instability. In step 2, the core denoising concept of the traditional denoising diffusion probability model (DDPM) is retained, while its multi-step iterative process is discarded. Instead, a single-step diffusion autoencoder (OSD-AE) and a single-step diffusion autodecoder (OSD-Dec) are constructed to achieve deep noise reduction and feature extraction of high-dimensional time-series electrical measurement data in a single forward propagation process.

[0021] In step 2, the preprocessed high-dimensional time series electrical measurement data information is recorded as x0. The process of the single-step diffusion autoencoder OSD-AE to initially extract and enhance the time series electrical measurement data information x0 is shown in formula (1):

[0022]

[0023] In formula (1), x T It is data corrupted by noise; is the diffusion coefficient, which is used to control the noise intensity; x0 is the high-dimensional time series electrical measurement data after preprocessing; is standard Gaussian noise, which simulates the random disturbance suffered by the data during the diffusion process.

[0024] Next, the single-step diffusion autoencoder OSD-AE is used to extract the noise-corrupted data x T The potential feature vector z is extracted from [1], and its mathematical expression is shown in formula (2):

[0025] z=f enc (x T ,θ enc )(2);

[0026] In formula (2): f enc is the encoder, which is embedded in the SKNet structure; θ enc is the encoder f enc A collection of parameters.

[0027] In step 2, when performing the data reconstruction task, the single-step diffusion self-decoder OSD-Dec passes through the decoder f dec Realize data reconstruction to improve the interpretability and stability of potential features; specifically, the decoder f dec Through nonlinear mapping, the potential feature vector z is restored to the original data space, as shown in formula (3):

[0028]

[0029] In formula (3): Is the data with the same distribution as the original time series electrical measurement data; fdec is the decoder, which is embedded in the SKNet structure; θ dec is the decoder f dec A collection of parameters.

[0030] In order to measure the difference between the reconstructed output and the original input and ensure that the potential features have good reconstructability and complete expression ability, the mean square error loss function (MSE) is used in the single-step diffusion self-decoder OSD-Dec. The mathematical expression of the mean square error loss function (MSE) is shown in the following formula (4):

[0031]

[0032] In formula (4): is the error loss function; f dec (z) represents the decoding function that restores the latent feature vector z to the original data space; represents the square distance between the original input and the reconstructed output in the Euclidean space for the same sample; x0 represents the time series electrical measurement data information; z represents the potential feature vector.

[0033] In step 2, the Transformer attention network, as one of the post-processing modules of the single-step diffusion autoencoder OSD-AE, is not used to model the multiple time steps of the diffusion process, but to optimize the global relationship of the latent feature vectors generated by the single-step diffusion. Specifically, the single-step diffusion autoencoder OSD-AE extracts the latent representation z through single-step diffusion. This feature not only contains the denoised information of the original input data, but also contains the deep patterns in the high-dimensional time-series electrical measurement data. The Transformer attention network uses multi-head attention (MHA) to calculate the information interaction relationship between different channels or dimensions, and further extracts discriminative features through a feedforward neural network (FFN). This process can effectively enhance the latent features extracted by the single-step diffusion autoencoder OSD-AE, making the model more discriminative in high-dimensional data space and improving the accuracy of identifying unstable nodes. In the subsequent calculation process of the Transformer attention network, the latent features are further reduced in computational complexity through global average pooling (GAP) while retaining the most discriminative information. Subsequently, multi-scale channel information is further fused through SKNet to enhance the robustness of feature expression. Finally, after being mapped to the target category through the fully connected layer (FC), the task of evaluating the transient voltage stability of the power system and dividing the voltage instability nodes or areas is completed.

[0034] In step 2, at the Softmax classification output layer, the output of the TVSDTNet feature-enhanced multi-task transient voltage stability assessment model is converted into a probability value between 0 and 1, thereby indicating the probability of whether the system transient voltage is stable or not and whether the voltage of each bus node is stable or not under system voltage instability; by setting an appropriate threshold, these probability values can be converted into clear classification decisions. In addition, the Sigmoid classifier performs stably in gradient calculation and model training, further improving the performance and reliability of the model. The mathematical expression of the Sigmoid classifier is shown in formula (5):

[0035]

[0036] In formula (5), σ1(x) represents the output of the Sigmoid function, and y is the input value, which is usually the output score of the model. Here, it refers to the input score of the high-dimensional time-series electrical measurement data after being sorted by the single-step diffusion autoencoder OSD-AE, the Transformer attention network, GPA, the selective kernel convolution network SKNet, and the fully connected layer. The output value of the Sigmoid function can be interpreted as the probability that y belongs to the transient voltage stability of the power system. For example, when σ1(y)>0.5, it usually indicates that the sample is classified as a stable sample; when σ1(y)≤0.5, it usually indicates that the sample is classified as an unstable sample.

[0037] In step 3, Task 1 is essentially a binary classification assessment, which aims to determine whether transient voltage instability occurs in the power system under specific operating conditions. Its output is a system-level stability / instability label, which is used to comprehensively grasp the transient voltage safety status of the system;

[0038] The essence of Task 2 is a multi-binary classification evaluation. It aims to further identify the nodes or areas where voltage instability occurs when the system is judged as transient voltage instability by Task 1, and to make stability / instability judgments on the busbars of each node, thereby helping operation and maintenance personnel to quickly locate key fault nodes and take corresponding control measures. The present invention selects corresponding measurement indicators for different tasks. Task 1 selects the main indicators derived from the confusion matrix: accuracy and F1 score; Task 2 uses a customized voltage instability complete coverage rate f cr , voltage stability complete coverage rate t cr And the geometric mean G of the two m , which is used to measure the performance of dividing voltage instability nodes or areas.

[0039] In step 3, the transient voltage stability of the power system under different operating conditions is judged using practical engineering criteria, as follows:

[0040] The bus node voltage must remain below 0.75 pu for no more than 1 second after a fault. Task 1 is defined as assessing the overall system transient voltage stability. If any bus node voltage under a given operating condition does not meet the transient voltage stability criteria, the system label is assigned a value of 0; otherwise, it is assigned a value of 1. Task 2 is defined as demarcating voltage instability nodes or regions under transient voltage instability conditions. In Task 2, each bus node distribution is evaluated for stability or instability. If the bus node voltage meets the practical transient voltage stability criteria, the corresponding label is assigned a value of 1; otherwise, it is assigned a value of 0. When the power system is applied to m bus nodes, a label vector of dimension m is obtained.

[0041] In step 4, in order to measure whether the constructed multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement can correctly assess the transient voltage stability of the power system, the accuracy (A) is defined according to the binary classification assessment confusion matrix shown in Table 1. c ), precision (P r ), recall rate (R e ) and F1 score (F1), as shown in Equations (9) to (12). The accuracy and F1 score are selected as the evaluation indicators of the system transient voltage stability assessment task;

[0042]

[0043] In formula (9) to formula (12), A c is the accuracy, which is used to measure the overall recognition ability of the model for all samples; R e is the recall rate, which is used to measure the proportion of correctly predicted samples in actual unstable samples; P r As the accuracy, it is used to measure how many of all samples predicted to be unstable are actually unstable; F1 is the harmonic average of precision and recall, which comprehensively evaluates the recognition effect of the model on unstable samples. In order to fully reflect the performance of the model in the transient voltage stability assessment of the power system, the present invention mainly adopts A c The two indicators of F1 are used as the core evaluation criteria.

[0044] Table 1 Confusion matrix

[0045]

[0046] In Table 1, T P and F N They represent the number of correct and incorrect judgments of transient voltage stability samples, F P and T N They represent the number of transient voltage instability samples that are incorrectly and correctly identified respectively.

[0047] In step 4, in order to measure whether the multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement can fully and correctly predict whether the voltage instability node bus in the transient voltage instability power system can be fully and correctly identified, the voltage instability node complete coverage rate f is defined. cr , as shown in the following formula (13); In order to measure whether the model can fully and correctly predict whether the voltage stability node bus in the transient voltage instability power system can be fully and correctly identified, the voltage stability node complete coverage rate t is defined cr , as shown in formula (14).

[0048]

[0049] In the above formula, f cr is the complete coverage rate of voltage instability nodes; t cr is the complete coverage rate of voltage stability nodes; N is the total number of samples identified as transient voltage instability in task 1; y ij is the true temporary stability label of the j-th bus transient voltage in the i-th sample; is an indicator function, which is 1 when all unstable nodes of sample i are correctly predicted, and 0 otherwise; is the jth predicted label of the i-th sample;

[0050] In step 5, the model is comprehensively evaluated in terms of its ability to identify both voltage instability nodes and voltage stability nodes, so as to achieve a more refined evaluation of the model's ability to distinguish the transient voltage stability of each node. cr and voltage stability node complete coverage rate t cr The geometric mean G m , providing accurate and intuitive quantitative reference for voltage instability node or area division;

[0051] Geometric mean G m The mathematical expression of is shown in equation (15).

[0052]

[0053] In formula (15), G m is the voltage instability node complete coverage rate f cr and voltage stability node complete coverage rate t cr The geometric mean of cr is the complete coverage rate of voltage instability nodes; t cr Full coverage of voltage stabilization nodes.

[0054] The present invention provides a multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement, and the technical effects are as follows: 1) The method of the present invention can effectively reduce noise and enhance data features of time series electrical measurement data based on the single-step diffusion autoencoder OSD-AE based on the traditional DDPM core denoising idea, thereby improving the transient voltage stability assessment performance of complex power systems and the accuracy of voltage instability node or area division.

[0055] 2) The present invention not only uses the confusion matrix derived index of Task 1 to intuitively evaluate the transient voltage stability of the power system, but also uses the evaluation index f cr , t cr and their geometric mean G m , objectively reflecting the characteristics of voltage instability nodes or regional divisions from the sample level. It can effectively assist operating personnel in analyzing grid stability and improve the safety and reliability of the power system. 3) The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement proposed in this invention uses the single-step diffusion autoencoder OSD-AE to perform preliminary data feature enhancement and noise reduction processing, and the Transformer attention network to further perform data noise reduction and feature enhancement processing, and combines the single-step diffusion autodecoder OSD-Dec to reconstruct data with the same distribution as the time-series electrical measurement data. It not only effectively enhances the expression of time-series electrical measurement data information, but also realizes the system transient voltage stability assessment and the precise division of voltage instability nodes or regions.

[0056] 4) Step 2 of the present invention retains the core denoising idea of the traditional denoising diffusion probability model DDPM while abandoning the multi-step iterative process and constructing a single-step diffusion autoencoder-autodecoder (OSD-AE & OSD-Dec) to achieve efficient high-dimensional time series electrical measurement data feature extraction and discrimination, and preliminarily enhance the features of high-dimensional time series electrical measurement data.

[0057] 5) Step 2 of the present invention constructs a single-step diffusion autoencoder OSD-AE based on the DDPM concept. OSD-AE can be regarded as a data feature enhancement strategy that enables network learning to adapt to noise interference and extract the noise-affected data x T The latent feature vector z is extracted from the CNN to enhance the generalization ability of the model. Secondly, only one noise perturbation is performed during the forward diffusion process, avoiding complex step-by-step iterations.

[0058] 6) Step 2 of the present invention constructs a single-step diffusion self-decoder OSD-Dec based on the DDPM concept. The single-step diffusion self-decoder OSD-Dec is constructed by the decoder f dec Achieve data reconstruction to improve the interpretability and stability of potential features.

[0059] 7) In the parameter adjustment process of TVSDTNet, the present invention adopts the weighted loss of the mean square error loss function and the focal loss function as the comprehensive optimization target of the model. By adjusting the weight coefficients of the two, the reconstructability of the potential characteristics in the reconstruction task and the accuracy and sensitivity to difficult-to-classify samples in the classification task are taken into account, thereby achieving the optimal balance of the overall performance of the model.

[0060] 8) The multi-task evaluation model based on TVSDTNet feature enhancement constructed by the present invention is used to evaluate the transient voltage stability state of the power system, and the performance of the model in identifying voltage instability nodes or areas is verified through two customized indicators: complete coverage rate of voltage instability nodes and complete coverage rate of voltage stability nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The present invention will be further described below with reference to the accompanying drawings and examples:

[0062] Figure 1 Schematic diagram of the traditional DDPM structure.

[0063] Figure 2 Schematic diagram of the selective kernel convolutional network SKNet structure.

[0064] Figure 3 Schematic diagram of the transient voltage stability assessment model based on TVSDTNet.

[0065] Figure 4 Schematic diagram of transient voltage stability assessment process based on TVSDTNet.

[0066] Figure 5 This is the modified IEEE 39-node test system topology diagram.

[0067] Figure 6 This is a schematic diagram of the sample situation of bus voltage instability at the task two node.

[0068] Figure 7 This is an analysis chart of the impact of loss function weights on performance indicators.

[0069] Figure 8 Schematic diagram of the voltage instability node or area division for voltage instability sample 1.

[0070] Figure 9 Schematic diagram of the division of voltage instability nodes or areas for voltage instability sample 2.

[0071] Figure 10 Schematic diagram of the voltage instability node or area division for voltage instability sample 3.

[0072] Figure 11 Schematic diagram of the division of voltage instability nodes or areas for voltage instability sample 4.

[0073] Figure 12 This is a visualization diagram of the timing electrical measurement data t-SEN.

[0074] Figure 13 t-SEN visualization effect diagram of the potential feature vector z data.

[0075] Figure 14 t-SEN visualization of the Transformer attention network output data.

[0076] Figure 15 t-SEN visualization of the Softmax classifier output data.

[0077] Figure 16 This is the geographical wiring diagram of the power grid in a certain area of Hebei Province. DETAILED DESCRIPTION

[0078] The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement first draws on the traditional DDPM core denoising idea to construct a single-step diffusion autoencoder and a single-step diffusion autodecoder, and embeds the selective kernel convolutional network SKNet in them to enhance the multi-scale denoising and deep potential feature extraction capabilities of the original high-dimensional time series electrical measurement data. Secondly, the Transformer attention network is combined to further capture the long-range dependencies between data features, and SKNet is again integrated into the classification branch to strengthen multi-scale convolution and channel attention. At the same time, during the training process, the mean square error loss of the diffusion model and the focal loss of the classification task are combined according to the weights, taking into account the classification performance and the robustness of feature expression. The model's accurate identification and classification performance of voltage instability nodes or areas is evaluated by customizing the complete coverage rate of voltage instability nodes and the complete coverage rate of voltage stable nodes. Finally, through the IEEE 39-node test system with high penetration of wind and solar turbines and the verification analysis of a regional power grid in Hebei Province, the effectiveness and practicality of the proposed method in high-dimensional data denoising, feature enhancement and power system transient voltage stability assessment are verified.

[0079] (1) Construction of a multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement:

[0080] This paper proposes a multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement. The model consists of six parts: input electrical measurement data, a single-step diffusion autoencoder (OSD-AE) based on the DDPM concept, a single-step diffusion autodecoder (OSD-Dec) based on the DDPM concept, a Transformer attention network, a selective kernel convolutional network (SKNet), and a softmax classification output layer.

[0081] 1) Input electrical measurement data: This provides basic training data for the model, ensuring that the model can learn and identify the transient stability characteristics of the power system under different operating conditions.

[0082] The selected raw high-dimensional time-series electrical measurement data should cover data from steady-state, fault periods, and post-fault removal to reflect the dynamic characteristics of the system at different stages. Specifically, the electrical quantities of the raw high-dimensional time-series electrical measurement data include, but are not limited to: bus-level data, such as bus voltage amplitude and phase angle; branch-level data, such as active power and reactive power of AC branches; and generator-unit-level data, such as generator power angle, generator active power and reactive power, and generator angular velocity, among other key electrical quantities.

[0083] The transient voltage stability assessment problem is divided into two interrelated subtasks. Task 1 is essentially a binary classification assessment, aiming to determine whether the power system has experienced transient voltage instability under specific operating conditions. Its output is a system-level stability / instability label, which is used to comprehensively understand the transient voltage safety status of the system. Task 2 is essentially a multi-binary classification assessment. When the system is judged to have transient voltage instability by Task 1, it aims to further identify the node or area where voltage instability has occurred and perform stability / instability judgments on the busbars of each node, thereby helping operation and maintenance personnel quickly locate critical fault nodes and implement appropriate control measures.

[0084] The transient voltage stability of power systems operating under different operating conditions is assessed using practical engineering criteria: the busbar node voltage must remain below 0.75 pu for no more than 1 second after a fault. Task 1 is defined as assessing the overall transient voltage stability of the system. In Task 1, if any busbar node voltage under a given operating condition does not meet the transient voltage stability criteria, the system label is assigned a value of 0; otherwise, it is assigned a value of 1. Task 2 is defined as delineating unstable voltage nodes or regions under transient voltage instability conditions. In Task 2, each busbar node distribution is assessed for stability or instability. If the busbar node voltage meets the practical transient voltage stability criteria, the corresponding label is assigned a value of 1; otherwise, it is assigned a value of 0. When the power system is applied to m busbar nodes, a label vector of dimension m is obtained.

[0085] 2) Single-step diffusion autoencoder OSD-AE based on DDPM idea: The schematic diagram of the traditional DDPM structure is as follows Figure 1 As shown in Figure 3, the step-by-step diffusion-denoising mechanism requires multiple rounds of calculation to recover the data, while OSD-AE only performs noise perturbation once during the forward diffusion process, avoiding complex step-by-step iterations.

[0086] Specifically, given high-dimensional time-series electrical measurement data x0, its single-step diffusion process is shown in formula (1).

[0087]

[0088] In the above formula, x T It is data corrupted by noise; is the diffusion coefficient, which is used to control the noise intensity; x0 is the original high-dimensional time series electrical measurement data; is standard Gaussian noise, which simulates the random disturbance suffered by the data during the diffusion process.

[0089] This process can be regarded as a data feature enhancement strategy, which enables the network to adapt to noise interference during learning, thereby enhancing the generalization ability of the model.

[0090] Then, OSD-AE passes through the encoder f enc From the noise-affected data x T Extract the potential feature vector z, and its mapping relationship is shown in formula (2).

[0091] z=f enc (x T ,θ enc ) (2);

[0092] In the above formula, z is the potential feature vector representation; f enc It is an encoder, using SKNet structure; x T is the data interfered by noise; θ enc is the encoder f enc A collection of parameters.

[0093] 3) Single-step diffusion self-decoder OSD-Dec based on DDPM idea: OSD-Dec passes through the decoder f dec Achieve data reconstruction to improve the interpretability and stability of potential features.

[0094] Specifically, the decoder f dec Through nonlinear mapping, the latent feature vector z is restored to the original data space, as shown in formula (3).

[0095]

[0096] In the above formula, Is the data with the same distribution as the original time series electrical measurement data; f dec is the decoder, which is embedded in the SKNet structure; z is the potential feature vector representation; θ dec is the decoder f dec A collection of parameters.

[0097] In order to ensure that the potential feature vector z can be used for classification while retaining the original information of the data features, the present invention uses the mean square error (MSE) loss to constrain the reconstructability of the potential features, and its mathematical expression is shown in formula (4).

[0098]

[0099] In the above formula, is the error loss function; x0 is the original time series electrical measurement data; f dec It is a decoder, which is embedded with SKNet structure; is standard Gaussian noise, simulating the random perturbations suffered by data during the diffusion process; z is the potential eigenvector.

[0100] 4) Transformer Attention Network: As a post-processing module of the single-step diffusion autoencoder (OSD-AE), it optimizes the latent feature vector z generated by the single-step diffusion autoencoder (OSD-AE). OSD-AE extracts the latent representation z through single-step diffusion. This feature not only contains the denoised information of the original input data, but also contains the deep patterns in the high-dimensional time-series electrical measurement data. The Transformer Attention Network calculates the information interaction between different channels or dimensions of the latent feature vector z through multi-head attention (MHA) and further extracts discriminative features through a feedforward neural network (FNN).

[0101] 5) Selective Kernel Convolutional Network (SKNet): This network can extract both global and local information, enabling collaborative modeling of multi-scale features. Its core concept is to set up multiple convolutional branches in parallel within the same network layer, each branch using a different convolution kernel size or receptive field, thereby capturing signal features at different levels. Subsequently, an attention mechanism is used to perform a weighted fusion of these branch outputs, dynamically selecting the most effective branch output.

[0102] Specifically, the structure diagram of the selective kernel convolutional network SKNet is as follows Figure 2 The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement applies three selective kernel convolutional networks: first, it is embedded in the single-step diffusion autoencoder OSD-AE to enhance the encoder f enc First, it has the potential to represent the original time-series electrical measurement data information; second, it is embedded in the single-step diffusion self-decoder OSD-Dec to capture multi-scale information in the reconstruction stage, improving the recovery effect of key transient features; third, after the Transformer attention network, the deep features extracted by attention are further fused at multiple scales, significantly improving the model's adaptability and expression ability for complex time-series electrical measurement data information.

[0103] 6) Softmax Classification Output Layer: This classification output layer converts the output of the TVSDTNet feature-enhanced multi-task transient voltage stability assessment model into probability values between 0 and 1, representing the probability of system transient voltage stability and the stability of each busbar node voltage under system voltage instability. By setting appropriate thresholds, these probability values can be converted into clear classification decisions. In addition, the Sigmoid classifier performs stably in gradient calculation and model training, further improving the performance and reliability of the model. The mathematical expression of the Sigmoid classifier is shown in Equation (5).

[0104]

[0105] In the above formula, σ1(x) represents the output of the Sigmoid function, and y is the input value, which is usually the output score of the model. Here, it refers to the input score of the high-dimensional time-series electrical measurement data information after being sorted by the single-step diffusion autoencoder OSD-AE, the Transformer attention network, GPA, the selective kernel convolution network SKNet, and the fully connected layer. The output value of the Sigmoid function can be interpreted as the probability that y belongs to the transient voltage stability of the power system. For example: when σ1(y)>0.5, it usually means that the sample is classified as a stable sample; when σ1(y)≤0.5, it usually means that the sample is classified as an unstable sample;

[0106] The classifier ultimately outputs a total of m+1 transient voltage stability labels, with the last m labels corresponding to the transient voltage stability status of each busbar in the system, and the first label reflecting the overall transient voltage stability of the power system. This output format allows the model to quickly determine the overall transient voltage stability of the system under specific operating conditions during online evaluation, while also accurately identifying nodes or areas experiencing voltage instability.

[0107] In summary, the structural diagram of the multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement proposed in this paper can be expressed as follows: Figure 3 shown.

[0108] (2) Multi-task transient voltage stability assessment process based on TVSDTNet feature enhancement:

[0109] 1) Time domain transient stability simulation stage: The present invention performs time domain transient stability simulation on the power system under different specific operating conditions to obtain historical steady-state data information and fault information of the power system, and divides the simulation results into transient stability label information according to the practical transient voltage stability criterion.

[0110] 2) Sample set construction: In order to ensure the completeness of sample data in the time domain, the present invention selects the steady-state time t0 before the fault and the fault occurrence time t fand the current time t after the fault is cleared c+n Specifically, t0 corresponds to the steady-state data obtained in the cycle before the fault occurs, thus fully covering the initial states of different operating conditions.

[0111] 3) Data preprocessing: To facilitate model training and evaluation, the present invention uses Z-score normalization to normalize the data input into the model so that all data are in the same dimension and the negative impact of dimensional differences is eliminated. The mathematical expression of Z-score normalization is shown in Formula (6).

[0112]

[0113] In the above formula, x0 represents the normalized high-dimensional time-series electrical measurement data information; x represents the original high-dimensional time-series electrical measurement data information; μ represents the mean of the data in the sample; σ is the standard deviation of the sample data; ζ represents a minimum value.

[0114] 4) Model Optimization Phase: During model training, this paper uses the Adam optimizer and introduces ReduceLROnPlateau as a learning rate scheduling strategy. If the test set loss does not decrease significantly after 10 consecutive iterations, the learning rate is automatically reduced to 1 / 10 of its original value to ensure better convergence stability in later training stages. Furthermore, to prevent parameter update anomalies caused by gradient explosion, this paper uses gradient clipping to limit the gradient norm to less than 1.0, thereby reducing instability during training and effectively mitigating the risk of overfitting.

[0115] 5) Hyperparameter Optimization: During model training, the present invention simultaneously performs multi-objective optimization on both the "reconstruction subtask" and the "classification subtask." The reconstruction subtask corresponds to the OSD-Dec module, using the MSE loss to measure the accuracy of noise reduction reconstruction. The classification subtask focuses on transient voltage stability assessment and unstable node classification, using the focal loss (FL) function to improve the discrimination performance of unbalanced electrical measurement data. The mathematical expression of the overall loss function is shown in Equation (7).

[0116]

[0117] In the above formula, is the weighted loss function; λ is The weight coefficient of is the mean square error loss function; is the focal loss function.

[0118] The mathematical expression of the FL loss function is shown in the following formula (8).

[0119]

[0120] In the above formula, is the focal loss function; α is the balancing factor; p is the sample prediction probability; γ is the focusing factor.

[0121] 6) Model Comprehensive Evaluation Indicators: After training, the test data set is passed to the model, and the model performance is comprehensively evaluated based on the comprehensive evaluation indicators output by the model. The present invention selects corresponding measurement indicators for different tasks.

[0122] Task 1 defines the accuracy (A) based on the binary classification evaluation confusion matrix shown in Table 1. c ), precision (P r ), recall rate (R e ) and F1 score (F1), as shown in Equations (9) to (12). Accuracy and F1 score are selected as evaluation indicators for the power system transient voltage stability assessment task.

[0123]

[0124] In formula (9) to formula (12), A c is the accuracy, which is used to measure the overall recognition ability of the model for all samples; R e is the recall rate, which is used to measure the proportion of correctly predicted samples in actual unstable samples; P r As the accuracy, it is used to measure how many of all samples predicted to be unstable are actually unstable; F1 is the harmonic average of precision and recall, which comprehensively evaluates the recognition effect of the model on unstable samples. In order to fully reflect the performance of the model in the transient voltage stability assessment of the power system, the present invention mainly adopts A c The two indicators of F1 are used as the core evaluation criteria.

[0125] Task 2: Select the customized voltage instability node full coverage rate f cr , voltage stability node full coverage rate t cr , and their geometric mean G m As shown in formula (13)-formula (14).

[0126]

[0127] In equations (13) and (14), N is the total number of samples identified as transient voltage instability in task 1; y ij is the true temporary stability label of the j-th bus transient voltage in the i-th sample; is the jth predicted label of the i-th sample, is an indicator function, which is 1 when all unstable nodes of sample i are correctly predicted, and 0 otherwise.

[0128] By customizing the voltage instability node full coverage fcr and voltage stability node complete coverage rate t cr The geometric mean G m Provides accurate and intuitive quantitative reference for voltage instability node or area division. Geometric mean G m The mathematical expression of is shown in equation (15).

[0129]

[0130] In the above formula, G m is the voltage instability node complete coverage rate f cr and voltage stability node complete coverage rate t cr The geometric mean of cr is the complete coverage rate of voltage instability nodes; t cr Full coverage of voltage stabilization nodes.

[0131] In summary, the overall flow chart of the multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement proposed in this paper can be expressed as follows: Figure 4 shown.

[0132] The technical effects of the present invention will be further illustrated by examples below:

[0133] like Figure 5 As shown. The modified IEEE 39-node test system includes 8 traditional generator sets, 2 photovoltaic power stations and 3 wind farms. In addition, a two-terminal DC model is added at the position of node 16. The DC model structure is a bipolar LCC-HVDC DC model. The single-pole DC power is 500MW, totaling 1000MW. In order to enhance the adaptability of the TVSDTNet feature enhancement model to transient voltage stability assessment and voltage instability node or area division under various system conditions, the present invention sets three wind and solar penetration rate scenarios of 0%, 25% and 50% under the premise of a system frequency of 50Hz, and changes the load level step by step between 90% and 110%. Subsequently, a three-phase short-circuit grounding fault is applied to each AC line in turn. The fault duration is 3 to 11 cycles, and the fault clearing time is the end of 1 second. Time domain transient simulation is performed to collect key electrical quantities of the power system. The specific sampling volume includes bus voltage amplitude and phase angle, AC branch active and reactive power, generator power angle and angular velocity, as well as generator active and reactive power, generating a total of 13,365 sample data under different working conditions. Among them, in Task 1, there are 2,163 samples of power system voltage instability and 11,202 samples of voltage stability; in Task 2, the number of bus voltage instability samples is as follows: Figure 6 shown.

[0134] To facilitate the adjustment of the model structure, the changes in the feature dimensions of each main hidden measurement in the TVSDTNet feature enhancement model of high-dimensional time-series electrical measurement data are described in detail, as shown in Table 2. The batch size of the data before being passed into the model is set to B = 32.

[0135] Table 2 Changes in feature dimensions of each main hidden layer

[0136]

[0137] In order to verify the influence of the weighted combination coefficient λ of the MSE loss function of the reconstruction branch and the FL loss function of the classification task on the model performance, training and testing were performed under different λ. The results are shown in the figure. Figure 7 shown.

[0138] Depend on Figure 7 It can be seen that when the MSE reconstruction loss and FL classification loss are weighted and fused under the same training framework, the model shows an obvious synergistic effect in taking into account both the reconstruction quality and the improvement of the voltage instability identification ability. c and F1 and voltage instability node or area division performance evaluation index f cr Experimental results show that when the MSE weight parameter is 0.6, the performance of both overall system transient stability assessment and voltage instability node or area classification reaches optimal levels. This fully demonstrates that the strategy of combining reconstruction and classification has good synergy and robustness, providing effective support for both transient voltage stability assessment and voltage instability node or area classification.

[0139] Subsequently, in order to verify the comprehensive performance advantages of the multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement proposed in this invention over the existing transient voltage stability assessment methods, the present invention selected four existing models, namely SKNet, CNN-BiLSTM, SE-CGRU and Transformer, as benchmarks, and conducted a comprehensive performance comparison under the same data set and experimental environment. The results of the horizontal comparison of the comprehensive performance are shown in Table 3.

[0140] Table 3 Horizontal comparison of comprehensive performance

[0141]

[0142] As shown in Table 3, the TVSDTNet model performs best in all indicators, especially in f cr and t cr These two key indicators reached 95.66% and 95.36% respectively, and the geometric mean G mCompared with other benchmark models, TVSDTNet can more accurately identify transient voltage instability nodes or areas while maintaining the overall system discrimination performance. This intuitively demonstrates that the feature enhancement model proposed in this paper not only has excellent system transient voltage stability assessment capabilities, but also shows significant advantages in delineating voltage instability nodes or areas.

[0143] Next, in order to further verify the accuracy of the TVSDTNet model in identifying voltage instability nodes or areas, 4 samples were randomly selected from the test set sample data, and transient voltage stability evaluation was performed on them. The voltage instability nodes or areas were drawn according to the model output results, as shown in the figure. Figures 8 to 11 shown. Figure 8 、 Figure 9 、 Figure 10 and Figure 11 The following diagrams show the voltage instability node or area division for Sample 1, Sample 2, Sample 3, and Sample 4, which are judged as voltage instability in Task 1. The red busbars in the figure represent the busbars judged as voltage instability in the actual measurement simulation, and the black busbars represent the busbars with stable voltage in the actual measurement simulation.

[0144] Depend on Figure 8 、 Figure 9 、 Figure 10 and Figure 11 It can be seen that the voltage instability nodes or areas mapped based on the identification results provided by the TVSDTNet feature-enhanced multi-task transient voltage stability assessment model are generally consistent with the time-domain simulation results, and can accurately determine the majority of voltage-instability buses while maintaining the correct identification of voltage-stable buses. This further demonstrates that the multi-task assessment model does help the model extract more discriminative features from raw high-dimensional electrical measurement data, achieving precise location of voltage instability nodes.

[0145] It should be pointed out that the multi-task framework designed in the present invention faces two sample imbalance problems at the same time: first, in the system transient voltage stability assessment, the power grid is in a "transient voltage stable state" for a long time and the "transient voltage instability state" is relatively rare, resulting in an insufficient proportion of unstable samples in the overall sample data; second, in the voltage instability node or area division task, different buses present their own fault distribution and timing characteristics, further exacerbating the imbalance between samples. In addition, due to the influence of uncertainties such as differences in operating conditions, the model may still make misjudgments or miss judgments on a few nodes, resulting in individual unstable buses not being fully identified or being misjudged as stable buses. However, the proportion of these error cases in the huge sample space is extremely low, and the TVSDTNet feature enhancement model can still maintain a high consistency in the positioning of unstable nodes, which fully demonstrates its good robustness and practical value in transient voltage stability assessment and unstable node or area division.

[0146] In the transient stability assessment of power systems, three-phase short-circuit faults are one of the typical disturbances that cause drastic changes in the system state. Their occurrence often leads to drastic fluctuations in the bus voltage amplitude and phase angle, and further affects the dynamic response of the AC branch power, generator power angle, etc. However, in actual operation, problems such as limited accuracy of measurement equipment, communication delays or noise interference may also be encountered, causing the original key fault feature quantities to be missing, distorted, or even seriously deviate from the ideal distribution. If the model is trained only based on ideal data, the model performance is very likely to decline significantly when encountering such abnormal features. In order to simulate the above-mentioned real-life scenarios and verify the robustness of the proposed model to "abnormal features", the present invention simulates the data on t c+n The characteristic portion of each moment is obscured by 10%, 20%, and 30% and replaced with random values. The newly added eigenvectors follow a Gaussian distribution with a mean of 0 and a variance of 1. This is used to assess the transient voltage stability of the power system. The results of Task 1 are shown in Table 4, and the results of Task 2 are shown in Table 5.

[0147] Table 4 Performance comparison of models with abnormal features (Task 1)

[0148]

[0149] Table 4 shows the A of each model under different abnormal feature proportions c The results show that as the proportion of abnormal features increases from 10% to 30%, the evaluation performance of all models decreases. However, the proposed TVSDTNet feature enhancement model maintains the best performance throughout the entire range and still achieves an accuracy of 98% when the proportion of abnormal features is 30%, significantly outperforming baseline models such as Transformer and CNN-BiLSTM.

[0150] Table 5 Performance comparison of models with abnormal features (Task 2)

[0151]

[0152]

[0153] Table 5 shows the performance comparison of each model in voltage instability node or area division. From the results, it can be seen that as the proportion of abnormal features increases, the f of all methods increases. cr and t cr The TVSDTNet model not only maintains the highest coverage under 10% and 20% abnormal feature conditions, but also achieves more than 95% f under the most extreme conditions. cr and 95% of t cr .

[0154] Combining Tables 4 and 5 demonstrates that the TVSDTNet model can effectively filter or compensate for the interference of abnormal features on system temporary stability assessment, demonstrating strong noise resistance and robustness. It also further demonstrates that the model is also highly robust in multi-binary classification scenarios, and can accurately locate voltage instability nodes or areas even in the presence of defective features or high levels of noise.

[0155] Finally, in order to more intuitively demonstrate the effectiveness of the proposed TVSDTNet feature enhancement model for system transient voltage stability assessment, the present invention uses the T-SEN algorithm to visualize the feature outputs of each stage of the model. T-SEN can map high-dimensional structures to low-dimensional space while better retaining the similarity between data, making it easier to observe the distribution of features in low-dimensional space. The present invention sequentially displays high-dimensional time series electrical measurement data, the potential feature vector z output by OSD-AE, the Transformer output data features, and the final classifier output, and performs t-SNE visualization. The experimental results are shown in Figure 2. Figures 12 to 15 As shown. Figure 12 、 Figure 13 、 Figure 14 and Figure 15 They represent the feature distribution of high-dimensional time-series electrical measurement data, the feature distribution of the latent feature vector z obtained by OSD-AE processing, the feature distribution of the Transformer attention network output data, and the feature distribution of the final Softmax classifier output data.

[0156] Depend on Figure 12 It can be seen that the two types of samples, “voltage stability” and “voltage instability” in the original data features, are highly mixed in the low-dimensional space, and the boundary between the classes is not obvious; Figure 13 It can be seen that after OSD-AE processing, although the inter-class clustering effect is enhanced to a certain extent, there is still local overlap. Figure 14 It can be seen that under the long-range dependency of Transformer, a clearer inter-class separation trend is shown. Figure 15 It can be seen that the two major clusters are almost completely separated in the classifier output, and the two types of samples almost no longer overlap.

[0157] This visualization directly confirms that OSD-AE is effective in denoising and extracting deep features from raw high-dimensional time series data. Furthermore, when combined with the Transformer's global sequence information modeling capabilities, the discriminability of features is further amplified, providing a clearer discriminant boundary for the final classification. This "step-by-step enhancement" process is precisely why the proposed TVSDTNet model achieves excellent performance in multi-task transient voltage stability assessment and the classification of unstable nodes or regions.

[0158] also, Figure 16A geographical connection diagram of a regional power grid in Hebei Province, my country, is presented. This grid is a receiving-end grid with a high penetration of wind and solar power. The regional grid includes six thermal power plants, three photovoltaic power plants, two wind farms, four 500kV substations and their equivalent circuits, 49 220kV substations and their equivalent circuits, and 17 110kV substations and their equivalent circuits. During sample collection, based on the specific configuration of the Hebei regional power grid, three-phase short-circuit faults occurring on one line of a single-circuit and double-circuit line were simulated. Specifically, the fault locations were set at 2%, 50%, and 98% of the line length to simulate the impact of different locations on grid stability. Three load levels of 80%, 100%, and 120% were considered, along with six fault durations ranging from 5 to 11 cycles. 6,000 samples were randomly generated, of which 5,057 showed voltage stability and 943 showed voltage instability. The performance of the proposed model was again compared with the baseline model, as shown in Table 6.

[0159] Table 6 Horizontal comparison of comprehensive performance of power grid in a certain region

[0160]

[0161] As shown in Table 6, the TVSDTNet feature enhancement model outperforms the baseline model in all indicators. c , also excelling in delineating voltage instability nodes or regions. While other models also achieved high values on some metrics, they were inferior to the TVSDTNet model. This demonstrates the model's excellent generalization capabilities, greater accuracy and stability in complex power system fault identification, and its high application value.

Claims

1. A multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement, characterized by The following steps are involved: Step 1: Construct time-series electrical measurement data information; Step 2: Construct a multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement; Step 3: Split the transient voltage stability assessment problem into two interrelated subtasks: Task 1: Determine whether the power system has transient voltage instability under specific operating conditions; Task 2: If the system is determined to have transient voltage instability by Task 1, further identify the node or area where voltage instability has occurred. Step 4: Define the voltage instability node complete coverage rate f cr , voltage stability node full coverage rate t cr and their geometric mean G m , in order to evaluate the ability of the multi-task transient voltage stability assessment model to accurately identify and divide transient voltage instability nodes or areas; Step 5: Based on the multi-task transient voltage stability assessment model constructed in step 2, the transient voltage stability state is assessed, and the voltage instability nodes or areas are accurately identified and divided.

2. The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement according to claim 1 is characterized by: In step 2, the multi-task transient voltage stability assessment model includes: Single-step diffusion autoencoder OSD-AE is used to extract and enhance the time series electrical measurement data information and generate the latent feature vector z; The single-step diffusion self-decoder OSD-Dec is used for the reconstruction of time-series electrical measurement data information, using a mean square error loss function to constrain the reconstructability of latent features; Transformer attention network, used for classification tasks, uses a focal loss function to constrain the model's attention to difficult-to-classify samples; Selective Kernel Convolutional Network (SKNet), used to capture signal features at different levels; Softmax classification output layer: converts the output of the multi-task transient voltage stability assessment model into a probability value between 0 and 1, thereby indicating the probability of system transient voltage stability and whether the voltage of each bus node is stable under system voltage instability.

3. The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement according to claim 2 is characterized by: In step 2, a single-step diffusion autoencoder OSD-AE and a single-step diffusion autodecoder OSD-Dec are constructed, and the pre-processed high-dimensional time series electrical measurement data information is recorded as x0; The process of the single-step diffusion autoencoder OSD-AE to initially extract and enhance the time series electrical measurement data information x0 is shown in formula (1): In formula (1), x T It is data corrupted by noise; is the diffusion coefficient, which is used to control the noise intensity; x0 is the high-dimensional time series electrical measurement data after preprocessing; is standard Gaussian noise, which simulates the random disturbance suffered by the data during the diffusion process; Next, the single-step diffusion autoencoder OSD-AE is used to extract the noise-corrupted data x T The potential feature vector z is extracted from [1], and its mathematical expression is shown in formula (2): z=f enc (x T ,θ enc )(2); In formula (2): f enc is the encoder, which is embedded in the SKNet structure; θ enc is the encoder f enc A collection of parameters.

4. The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement according to claim 3 is characterized by: When performing data reconstruction tasks, the single-step diffusion self-decoder OSD-Dec passes through the decoder f dec Realize data reconstruction to improve the interpretability and stability of potential features; specifically, the decoder f dec Through nonlinear mapping, the potential feature vector z is restored to the original data space, as shown in formula (3): In formula (3): Is the data with the same distribution as the original time series electrical measurement data; f dec is the decoder, which is embedded in the SKNet structure; θ dec is the decoder f dec A collection of parameters.

5. The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement according to claim 4 is characterized in that: The mean square error loss function is used in the single-step diffusion self-decoder OSD-Dec. The mathematical expression of the mean square error loss function is shown in the following formula (4): In formula (4): is the error loss function; f dec (z) represents the decoding function that restores the latent feature vector z to the original data space; represents the square distance between the original input and the reconstructed output in the Euclidean space for the same sample; x0 represents the time series electrical measurement data information; z represents the potential feature vector.

6. The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement according to claim 5 is characterized by: At the Softmax classification output layer, the output of the multi-task transient voltage stability assessment model enhanced with TVSDTNet features is converted into a probability value between 0 and 1, thereby indicating the probability of whether the system transient voltage is stable or not and whether the voltage of each bus node is stable or not under system voltage instability. By setting appropriate thresholds, these probability values can be converted into clear classification decisions. The mathematical expression of the Sigmoid classifier is shown in Equation (5): In formula (5), σ1(x) represents the output of the Sigmoid function, and y is the input value, which refers to the input score of the high-dimensional time series electrical measurement data information after being sorted by the single-step diffusion autoencoder OSD-AE, Transformer attention network, GPA, selective kernel convolutional network SKNet, and fully connected layer; the output value of the Sigmoid function is the probability that y belongs to the transient voltage stability of the power system.

7. The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement according to claim 1 is characterized by: In step 3, corresponding measurement indicators are selected for different tasks; Task 1 selects the main indicators derived from the confusion matrix: accuracy and F1 score; Task 2 uses the customized voltage instability complete coverage rate f cr , voltage stability complete coverage rate t cr And the geometric mean G of the two m , used to measure the performance of dividing voltage instability nodes or areas; to judge the transient voltage stability of power systems operating under different working conditions, as follows: The bus node voltage shall not exceed 0.75pu for more than 1s after a fault; and the overall transient voltage stability assessment of the system is defined as Task 1. In Task 1, if any bus node voltage does not meet the transient voltage stability criterion under a certain working condition, the system label is recorded as 0, otherwise it is recorded as 1; the voltage instability nodes or areas under transient voltage instability working conditions are divided into Task 2. In Task 2, the stability and instability of each bus node distribution are judged. If the bus node voltage meets the practical criterion for transient voltage stability, the corresponding label is recorded as 1, otherwise it is recorded as 0; when the power system is used for m bus nodes, a label vector with a dimension of m can be obtained.

8. The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement according to claim 1 is characterized by: In step 4, in order to measure whether the constructed multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement can correctly assess the transient voltage stability of the power system, the accuracy (A) is defined according to the binary classification assessment confusion matrix. c ), precision (P r ), recall rate (R e ) and F1 score (F1), as shown in Equations (9) to (12); and the accuracy and F1 score are selected as the evaluation indicators of the system transient voltage stability assessment task; In formula (9) to formula (12), A c is the accuracy, which is used to measure the overall recognition ability of the model for all samples; R e is the recall rate, which is used to measure the proportion of correctly predicted samples in actual unstable samples; P r is the accuracy, which is used to measure how many of all samples predicted to be unstable are actually unstable; F1 is the harmonic average of precision and recall, which comprehensively evaluates the recognition effect of the model on unstable samples; A is used to fully reflect the performance of the model in the transient voltage stability assessment of the power system. c and F1 as the core evaluation criteria; T P and F N They represent the number of correct and incorrect judgments of transient voltage stability samples, F P and T N They represent the number of transient voltage instability samples that are incorrectly and correctly identified respectively.

9. The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement according to claim 8 is characterized by: In step 4, in order to measure whether the multi-task transient voltage stability assessment model based on TVSDTNet feature enhancement can fully and correctly predict whether the voltage instability node bus in the transient voltage instability power system can be fully and correctly identified, the voltage instability node complete coverage rate f is defined. cr , as shown in the following formula (13); In order to measure whether the model can fully and correctly predict whether the voltage stability node bus in the transient voltage instability power system can be fully and correctly identified, the voltage stability node complete coverage rate t is defined cr , as shown in the following formula (14); In the above formula, f cr is the complete coverage rate of voltage instability nodes; t cr is the complete coverage rate of voltage stability nodes; N is the total number of samples identified as transient voltage instability in task 1; y ij is the true temporary stability label of the j-th bus transient voltage in the i-th sample; is an indicator function, which is 1 when all unstable nodes of sample i are correctly predicted, and 0 otherwise; is the jth predicted label of the i-th sample.

10. The multi-task transient voltage stability assessment method based on TVSDTNet feature enhancement according to claim 1 is characterized in that: In step 5, the voltage instability node complete coverage rate f is customized cr and voltage stability node complete coverage rate t cr The geometric mean G m , providing accurate and intuitive quantitative reference for voltage instability node or area division; Geometric mean G m The mathematical expression of is shown in formula (15): In formula (15), G m is the voltage instability node complete coverage rate f cr and voltage stability node complete coverage rate t cr The geometric mean of cr is the complete coverage rate of voltage instability nodes; t cr Full coverage of voltage stabilization nodes.

Images