A grid-connected low-voltage ride-through capability evaluation method, system, device and medium based on deep learning

Abstract

The present application relates to a kind of based on deep learning's grid-connected low-voltage ride-through capability evaluation method, system, equipment and medium, method includes: obtaining real off-grid sample set and low-voltage ride-through physical constraint boundary, under the guidance of physical constraint, generate synthetic off-grid sample and with normal sample merge and construct equalization training data set;Based on graph convolution-time convolution double-flow fusion network, simultaneously extract power grid topology spatial feature and fault transient time sequence feature, complete model training under physical constraint and generate the low-voltage ride-through risk score and cognitive uncertainty variance of each grid-connected point;The risk score is carried out physical consistency check and topological consistency evaluation, combined with distribution matching degree index and cognitive uncertainty variance generates multidimensional reliability index;Comprehensive reliability score is obtained by fusion, according to the comparison result of comprehensive reliability score and preset threshold value, output low-voltage ride-through capability evaluation result, and when low reliability, feedback signal is generated and returned to sample generation step, form closed loop optimization.

Description

Technical Field

[0001] This invention relates to the field of power system security assessment technology, specifically to a method, system, equipment, and medium for assessing grid-connected low-voltage ride-through capability based on deep learning. Background Technology

[0002] With the continuous increase in the penetration rate of new energy sources, the power system exhibits significant characteristics of low inertia and weak support, posing increasingly severe challenges to the safe and stable operation of the power grid. Among various power grid faults, voltage dips are one of the most common forms of disturbance. When a voltage dip occurs at the grid connection point, if the new energy power generation equipment does not have the ability to withstand the voltage dip and disconnects from the grid prematurely, it may trigger cascading faults or even large-scale power outages. Therefore, grid low-voltage ride-through capability—that is, the ability of new energy power plants and inverter equipment to maintain continuous operation without disconnecting from the grid during grid voltage dips—has become a key indicator for measuring the safety of new energy grid connection. In recent years, academia and industry have conducted extensive research on the assessment of low-voltage ride-through capability. Among them, deep learning-based assessment methods, due to their ability to automatically learn fault transient characteristics from massive operational data, have shown significant advantages in assessment efficiency and are gradually becoming the mainstream technology in this field.

[0003] However, existing deep learning-based low-voltage ride-through capability assessment methods face a fundamental dilemma in practical engineering applications. On the one hand, actual grid disconnection incidents involving new energy power generation equipment are extremely rare, while samples of normal operation and successful ride-through constitute the vast majority. This extreme class imbalance causes deep neural networks to be severely biased towards the majority class during training, resulting in extremely limited ability to identify the minority class—the grid disconnection scenarios that truly require attention. On the other hand, existing methods generally treat the assessment model as a "black box," only outputting a grid disconnection risk score or a binary judgment result, without providing a quantitative description of the reliability of the assessment result itself. The model may still give a high-confidence risk score when there are insufficient training samples to support the input, or it may show extremely high confidence when the prediction is completely wrong. Due to the lack of an effective mechanism for evaluating the credibility of model output, dispatchers find it difficult to determine in which scenarios the model's assessment results are trustworthy and in which scenarios the model is on the failure boundary, making it difficult for such methods to gain genuine engineering trust in actual power dispatching decisions. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide a deep learning-based method, system, device, and medium for evaluating the reliability of grid-connected low-voltage ride-through capability under extremely small sample unbalanced conditions.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a deep learning-based method for evaluating grid-connected low-voltage ride-through capability, comprising the following steps:

[0007] S1: Obtain the real offline sample set and the low-voltage crossing physical constraint boundary. Based on the real offline sample set, generate the synthetic offline sample set under the guidance of the low-voltage crossing physical constraint boundary. Merge the synthetic offline sample set with the normal sample set to generate a balanced training dataset.

[0008] S2: Input the equalized training dataset into the preset graph convolution-temporal convolution dual-stream fusion network, extract the grid topology spatial features and fault transient temporal features, and complete the model training under the constraint of the low-voltage crossing physical constraint boundary to generate the low-voltage crossing risk score and cognitive uncertainty variance for each grid connection point;

[0009] S3: Based on the low-voltage crossing risk score and the low-voltage crossing physical constraint boundary, the degree of violation is verified, and a physical consistency index is generated. The spatial smoothness of the low-voltage crossing risk score and the obtained power grid topology is evaluated to generate a topology consistency index. At the same time, the distribution matching index between synthetic samples and real samples is calculated based on the equalized training dataset, and the cognitive uncertainty index is generated by combining the cognitive uncertainty variance.

[0010] S4: Based on preset weights, the physical consistency index, distribution matching index, topological consistency index and cognitive uncertainty index are weighted and summed to generate a comprehensive credibility score. The comprehensive credibility score is then mapped to a preset first threshold and second threshold to generate a low-pressure crossing capability assessment result, where the first threshold is less than the second threshold.

[0011] S5: When the overall credibility score is lower than the first threshold, mark the current evaluation result as a high credibility evaluation result and output the low-pressure crossing capability evaluation result; when the overall credibility score is higher than the second threshold, generate a low credibility sample feedback signal and return to step S1 to update the synthesized offline sample set.

[0012] Secondly, the present invention provides a deep learning-based system for evaluating grid-connected low-voltage ride-through capability, which is configured with the following modules:

[0013] The balanced sample construction module is used to obtain the real off-grid sample set and the low-voltage crossing physical constraint boundary. Based on the real off-grid sample set, a synthetic off-grid sample set is generated under the guidance of the low-voltage crossing physical constraint boundary. The synthetic off-grid sample set is then merged with the normal sample set to generate a balanced training dataset.

[0014] The dual-stream feature modeling module is used to input the equalized training dataset into the preset graph convolution-temporal convolution dual-stream fusion network, extract the topological spatial features of the power grid and the transient temporal features of the fault, and complete the model training under the constraint of the physical constraint boundary of low-voltage crossing, generating the low-voltage crossing risk score and cognitive uncertainty variance for each grid connection point;

[0015] The multi-dimensional consistency verification module is used to verify the degree of violation based on the low-voltage crossing risk score and the physical constraint boundary of low-voltage crossing, generate a physical consistency index, and evaluate the spatial smoothness of the low-voltage crossing risk score and the obtained power grid topology to generate a topology consistency index. At the same time, it calculates the distribution matching index between synthetic samples and real samples based on the equalized training dataset, and generates a cognitive uncertainty index by combining the cognitive uncertainty variance.

[0016] The comprehensive credibility assessment module is used to generate a comprehensive credibility score by weighted summation of the physical consistency index, distribution matching index, topological consistency index and cognitive uncertainty index based on preset weights. The comprehensive credibility score is then mapped to a preset first threshold and second threshold to generate a low-pressure crossing capability assessment result, wherein the first threshold is less than the second threshold.

[0017] The decision iteration feedback module is used to mark the current evaluation result as a high-confidence evaluation result and output the low-pressure crossing capability evaluation result when the comprehensive confidence score is lower than the first threshold; when the comprehensive confidence score is higher than the second threshold, it generates a low-confidence sample feedback signal and returns to step S1 to update the synthesized offline sample set.

[0018] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned deep learning-based methods for evaluating grid-connected low-voltage ride-through capability.

[0019] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the aforementioned deep learning-based methods for evaluating grid-connected low-voltage ride-through capability.

[0020] In summary, the deep learning-based method for evaluating grid-connected low-voltage cross-traffic capability provided in this application generates synthetic off-grid samples under the guidance of physical constraints on low-voltage cross-traffic and merges them with normal samples to construct a balanced training dataset. This effectively alleviates the class imbalance problem caused by the extreme scarcity of real off-grid samples, thereby improving the model's ability to identify minority off-grid scenarios. Secondly, by constructing a multi-dimensional credibility evaluation index system through physical consistency verification, topological space smoothness evaluation, distribution matching degree analysis, and cognitive uncertainty quantification, a systematic quantification of the reliability of the model's output results can be achieved, thus addressing the fundamental deficiency of existing methods in being unable to evaluate their own prediction confidence. Thirdly, by mapping the comprehensive credibility score to a preset threshold and dynamically adjusting the output strategy accordingly, an "undecidable" state can be proactively triggered in low-credibility scenarios instead of outputting potentially erroneous predictions, ensuring the evaluation security in high-risk decision-making scenarios. Finally, by feeding back low-credibility samples to the sample generation step to form a closed-loop optimization link, the evaluation system can continuously improve itself during long-term operation, thereby achieving a highly reliable evaluation of grid-connected low-voltage cross-traffic capability for actual engineering deployment.

[0021] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a deep learning-based method for evaluating grid-connected low-voltage ride-through capability, provided in an embodiment of this application;

[0023] Figure 2 This is a schematic diagram of a deep learning-based grid-connected low-voltage ride-through capability assessment system, provided as another embodiment of this application. Detailed Implementation

[0024] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] In one embodiment, such as Figure 1As shown, a deep learning-based method for assessing grid-connected low-voltage ride-through capability is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0027] S1: Obtain the real off-grid sample set and the low-voltage crossing physical constraint boundary. Based on the real off-grid sample set, generate the synthetic off-grid sample set under the guidance of the low-voltage crossing physical constraint boundary. Merge the synthetic off-grid sample set with the normal sample set to generate a balanced training dataset.

[0028] Specifically, the system acquires a real off-grid sample set and the physical constraint boundary for low-voltage crossing. The system extracts complete operational data on off-grid disconnections caused by grid voltage dips from the power grid dispatch center database, the new energy power plant monitoring system, and inverter operation and maintenance records, forming a real off-grid sample set. The system preprocesses the extracted real off-grid samples, including outlier removal, missing value completion, and normalization, to ensure the validity and consistency of the sample data. Based on power system operation patterns, new energy equipment technical standards, and actual engineering needs, the system determines the physical constraint boundary for low-voltage crossing. The constraints cover voltage dips, equipment operation, grid topology, and off-grid triggering constraints, and are stored in the form of a parameter threshold matrix. The system uses a constraint-based generative adversarial network as the sample generation model, training on the real off-grid sample set, and generates a synthetic off-grid sample set guided by the physical constraint boundary for low-voltage crossing.

[0029] The system incorporates a physical constraint penalty term into the model loss function to ensure that the synthesized samples meet all physical constraints. After generation, the synthesized samples are validated to ensure that their feature distribution is similar to that of real off-grid samples. The system acquires actual monitoring data of the normal operation and successful voltage dip crossing of new energy power generation equipment as a normal sample set. The normal sample set undergoes the same preprocessing as the real off-grid samples. Then, the validated synthetic off-grid sample set is merged with the preprocessed normal sample set, shuffled, and divided into training and validation sets to generate a balanced training dataset, thus solving the class bias problem in deep neural network training.

[0030] S2: Input the equalized training dataset into the preset graph convolution-temporal convolution dual-stream fusion network, extract the grid topology spatial features and fault transient temporal features, and complete the model training under the constraints of the low-voltage crossing physical constraint boundary to generate the low-voltage crossing risk score and cognitive uncertainty variance for each grid connection point.

[0031] Specifically, the system inputs a balanced training dataset into a pre-defined graph convolutional-temporal convolutional dual-stream fusion network. This network consists of a graph convolutional stream, a temporal convolutional stream, a feature fusion layer, and an output layer. The system extracts grid topology features through the graph convolutional stream, transforming the grid topology into an adjacency matrix. After inputting node feature parameters, the graph convolutional layer captures the topological relationships between nodes. The system also extracts fault transient temporal features through the temporal convolutional stream. After inputting the temporal feature sequences of each grid-connected point, the temporal convolutional layer captures fault transient features at different time scales. Finally, the feature fusion layer employs an attention mechanism to fuse the two types of features, emphasizing features that significantly impact low-voltage ride-through capability assessment.

[0032] Furthermore, the system outputs a risk score and cognitive uncertainty variance through two branches in the output layer. The risk score branch outputs the low-voltage crossing risk score for each grid connection point through a fully connected layer, while the uncertainty branch, based on the Monte Carlo dropout method, calculates the variance of the risk score as the cognitive uncertainty variance through multiple random inferences. The system employs a composite loss function for model training, which includes classification loss, physical constraint loss, and uncertainty loss. During training, the system calls the low-voltage crossing physical constraint boundary parameters in real time to constrain and verify the model output, adjusts the model parameters through the loss function, and completes model training using an optimizer and related strategies. After training, the system outputs the low-voltage crossing risk score and cognitive uncertainty variance for each grid connection point.

[0033] S3: Based on the low-voltage crossing risk score and the low-voltage crossing physical constraint boundary, the degree of violation is verified, and a physical consistency index is generated. The spatial smoothness of the low-voltage crossing risk score and the obtained power grid topology is evaluated to generate a topology consistency index. At the same time, the distribution matching index between synthetic samples and real samples is calculated based on the equalized training dataset, and the cognitive uncertainty index is generated by combining the cognitive uncertainty variance.

[0034] Specifically, the system verifies the degree of violation based on low-voltage ride-through risk scores and low-voltage ride-through physical constraint boundaries, generating a physical consistency index. The system first determines the expected range of risk scores under different operating conditions based on the low-voltage ride-through physical constraint boundaries, then calculates the degree of violation of the risk score for each grid connection point. Through normalization, the degree of violation is converted into a physical consistency index, which measures the degree of conformity between the risk score and the physical constraints. The system also performs a spatial smoothness assessment on the low-voltage ride-through risk scores and the acquired grid topology, generating a topology consistency index.

[0035] The system determines the neighboring nodes of each grid-connected point based on the grid topology adjacency matrix, calculates the spatial smoothness index of the risk score of each grid-connected point and the average risk score of its neighboring nodes, and averages all indices to obtain the topology consistency index, which is used to evaluate the spatial matching between the risk score and the grid topology. The system calculates the distribution matching index between synthetic and real samples based on a balanced training dataset, selecting core feature parameters as distribution comparison indicators. By calculating the distribution distance in the feature space and normalizing it, the distribution matching index is obtained, which measures the similarity of feature distributions between synthetic and real samples. The system generates a cognitive uncertainty index based on cognitive uncertainty variance, normalizes the cognitive uncertainty variance, and calculates the cognitive uncertainty index, which quantifies the model's cognitive uncertainty about the evaluation results.

[0036] S4: Based on preset weights, the physical consistency index, distribution matching index, topological consistency index and cognitive uncertainty index are weighted and summed to generate a comprehensive credibility score. The comprehensive credibility score is then mapped to a preset first threshold and second threshold to generate a low-pressure crossing capability assessment result, where the first threshold is less than the second threshold.

[0037] Specifically, the system generates a comprehensive credibility score by weighted summation of the physical consistency index, distribution matching index, topological consistency index, and cognitive uncertainty index based on preset weights. The weights are determined using the analytic hierarchy process (AHP) combined with engineering expert experience and can be adaptively adjusted according to the actual operating conditions of different renewable energy power plants to ensure the sum of the weights is 1. The system presets a first threshold and a second threshold, with the first threshold being less than the second threshold. These thresholds are optimized based on the false positive and false negative rates of the validation set data and are used to distinguish the credibility levels of the assessment results. The system performs interval mapping between the comprehensive credibility score and the preset first and second thresholds to generate low-voltage ride-through capability assessment results. During the interval mapping process, the system generates assessment results with different credibility levels according to the different intervals in which the comprehensive credibility score falls, and simultaneously outputs the corresponding low-voltage ride-through risk score and risk level information, thus binding the assessment results with credibility and providing a basis for power grid dispatching decisions.

[0038] S5: When the overall credibility score is lower than the first threshold, mark the current evaluation result as a high credibility evaluation result and output the low-pressure crossing capability evaluation result; when the overall credibility score is higher than the second threshold, generate a low credibility sample feedback signal and return to step S1 to update the synthesized offline sample set.

[0039] Specifically, the system categorizes the evaluation results based on different intervals of the comprehensive credibility score. When the comprehensive credibility score is below the first threshold, the system marks the current evaluation result as a high-credibility evaluation result and outputs grid connection point related information, low-voltage crossing risk score, risk level, comprehensive credibility score, and details of the four major credibility indices. The output format adopts a standardized format for easy reading and display by the power grid dispatching system. When the comprehensive credibility score is above the second threshold, the system generates a low-credibility sample feedback signal. The feedback signal includes the operating condition parameters of the current sample and the risk score and cognitive uncertainty variance output by the model. The system then returns this feedback signal to step S1 to update the synthetic off-grid sample set. The system extracts the operating condition parameters of the low-credibility samples from the feedback signal, adds them to the input constraints of the sample generation model, adjusts the model generation parameters, regenerates synthetic off-grid samples that conform to the operating condition and satisfy the physical constraints, adds the newly generated samples to the original synthetic off-grid sample set and removes invalid samples, merges them again to generate a new balanced training dataset, and then re-executes the subsequent steps to achieve iterative optimization of the sample set and model. When the overall credibility score is between the first and second thresholds, the system outputs the evaluation result and the overall credibility score, and generates auxiliary prompts to provide reference for dispatchers.

[0040] In summary, the deep learning-based method for evaluating grid-connected low-voltage cross-traffic capability provided in this application generates synthetic off-grid samples under the guidance of physical constraints on low-voltage cross-traffic and merges them with normal samples to construct a balanced training dataset. This effectively alleviates the class imbalance problem caused by the extreme scarcity of real off-grid samples, thereby improving the model's ability to identify minority off-grid scenarios. Secondly, by constructing a multi-dimensional credibility evaluation index system through physical consistency verification, topological space smoothness evaluation, distribution matching degree analysis, and cognitive uncertainty quantification, a systematic quantification of the reliability of the model's output results can be achieved, thus addressing the fundamental deficiency of existing methods in being unable to evaluate their own prediction confidence. Thirdly, by mapping the comprehensive credibility score to a preset threshold and dynamically adjusting the output strategy accordingly, an "undecidable" state can be proactively triggered in low-credibility scenarios instead of outputting potentially erroneous predictions, ensuring the evaluation security in high-risk decision-making scenarios. Finally, by feeding back low-credibility samples to the sample generation step to form a closed-loop optimization link, the evaluation system can continuously improve itself during long-term operation, thereby achieving a highly reliable evaluation of grid-connected low-voltage cross-traffic capability for actual engineering deployment.

[0041] In one embodiment, S1 of the deep learning-based method for evaluating grid-connected low-voltage ride-through capability provided by the present invention specifically includes the following steps:

[0042] S11: Obtain the real off-grid sample set and the reactive current response time constraint, voltage recovery time constraint and reactive current injection depth constraint in the physical constraint boundary of low voltage crossing. Decompose each sample in the real off-grid sample set into voltage drop waveform segment, current response waveform segment and control parameter segment, and perform sliding window segmentation and normalized scale compression processing on each waveform segment to generate a standardized off-grid sample feature sequence.

[0043] Specifically, the system acquires a real grid disconnection sample set and the reactive current response time constraint, voltage recovery time constraint, and reactive current injection depth constraint within the physical constraint boundary of low-voltage crossing. The real grid disconnection sample set originates from the historical fault database of the power grid dispatch center, the monitoring system of new energy power plants, and fault records provided by equipment manufacturers. The sample data covers grid disconnection events under different types of new energy power generation equipment, different power grid fault scenarios, and different voltage drop conditions. The system decomposes each sample in the real grid disconnection sample set into a voltage drop waveform segment, a current response waveform segment, and a control parameter segment. The voltage drop waveform segment contains time-series variation data of the grid connection point voltage, the current response waveform segment contains time-series variation data of the inverter output current, and the control parameter segment contains the control strategy parameters and operating status parameters of the new energy power generation equipment.

[0044] The system performs sliding window segmentation on each waveform segment, traversing the entire segment with a fixed step size to divide each segment into multiple continuous sub-waveform segments. The system then performs normalized scaling on each sub-waveform segment, mapping its parameter values ​​to a uniform range through a linear transformation to eliminate dimensional differences between parameters. Finally, the system integrates all processed sub-waveform segments, arranging them according to the original sample's temporal order to generate a standardized offline sample feature sequence. This standardized offline sample feature sequence preserves the fault transient characteristics and control parameter characteristics of the original samples.

[0045] S12: Construct a conditional variational autoencoder that uses the low-voltage crossing physical constraint boundary as the feasible domain constraint of the decoding output. The conditional variational autoencoder encodes the latent space of the standardized off-grid sample feature sequence. The latent space encoding result is then subjected to joint loss minimization iterative training with the constraint violation penalty term in the low-voltage crossing physical constraint boundary to generate a latent space distribution modulated by physical constraints.

[0046] Specifically, the system constructs a conditional variational autoencoder (CVA), which uses low-voltage crossing of physical constraint boundaries as the feasible region constraint for the decoded output. The CVA consists of an encoder and a decoder. The encoder maps the input standardized off-grid sample feature sequence to the latent space, generating a latent space encoded result; the decoder maps the latent space encoded result back to the original data space, outputting a reconstructed sample feature sequence. The system inputs the standardized off-grid sample feature sequence into the encoder of the CVA. The encoder, through multi-layer neural network computation, transforms the sample feature sequence into a probability distribution in the latent space, obtaining the latent space encoded result.

[0047] The system introduces a constraint violation penalty term from the low-voltage crossing of physical constraints, fusing this penalty term with the original loss function of the conditional variational autoencoder to construct a joint loss function. This joint loss function measures the reasonableness of the latent space encoding result, the reconstruction error between the decoded output and the original sample, and the degree to which the decoded output violates the physical constraints. The system iteratively trains by minimizing the joint loss, continuously adjusting the network parameters of the conditional variational autoencoder to ensure that the latent space encoding result accurately reflects the characteristics of the off-grid samples while ensuring that the decoded output meets the requirements of low-voltage crossing of physical constraints. After multiple rounds of iterative training, the system generates a latent space distribution modulated by physical constraints, which can simultaneously consider the characteristic patterns of off-grid samples and the physical constraints.

[0048] S13: Perform multiple random samplings from the latent space distribution modulated by physical constraints. After each sampling, use low-voltage crossing of the physical constraint boundary to filter the decoding output for constraint violation thresholds. Retain the decoding outputs that do not trigger the constraint violation thresholds as candidate synthetic samples. After waveform segment splicing and scale restoration processing of all candidate synthetic samples, merge them with the normal sample set to generate a balanced training dataset.

[0049] Specifically, the system performs multiple random samplings from the physically constrained latent space distribution, obtaining a latent space encoded vector with each sampling. The system inputs the latent space encoded vector obtained from each sampling into the decoder of a conditional variational autoencoder. The decoder decodes the latent space encoded vector and outputs a reconstructed sample feature sequence. The system uses a low-voltage crossing of the physical constraint boundary to filter the sample feature sequence from the decoded output by applying constraint violation thresholds, determining whether the operating parameters corresponding to the decoded output sample feature sequence meet the reactive current response time constraint, voltage recovery time constraint, and reactive current injection depth constraint. The system retains decoded outputs that do not trigger constraint violation thresholds as candidate synthetic samples and discards decoded outputs that do. The system performs waveform segment splicing on all candidate synthetic samples, splicing the voltage drop waveform segment, current response waveform segment, and control parameter segment of each candidate synthetic sample in their original time sequence to restore a complete sample structure. The system performs scale restoration processing on the spliced ​​candidate synthetic samples, restoring the normalized and compressed parameter values ​​to the original parameter range, resulting in a synthetic off-grid sample that conforms to the actual operating scenario. The system merges all synthetic off-grid samples with the normal sample set, which contains historical data on the normal operation of new energy equipment and successful low-voltage ride-through. The preprocessing method for the normal sample set is consistent with that of the real off-grid samples. The system then integrates and merges the sample sets to generate a balanced training dataset.

[0050] In one embodiment, S2 of the deep learning-based method for evaluating grid-connected low-voltage ride-through capability provided by the present invention specifically includes the following steps:

[0051] S21: Obtain the set of bus nodes and transmission line edges in the power grid topology. Construct a weighted graph adjacency matrix with each grid connection point bus as a graph node and the reciprocal of the transmission line impedance parameter as the edge weight. Extract the three-phase voltage amplitude, three-phase voltage phase, active power, reactive power and inverter controller parameters of each node as the initial feature matrix of the node, and generate the node feature representation of the weighted graph structure.

[0052] Specifically, the system acquires the set of bus nodes and the set of transmission line edges in the power grid topology. The power grid topology is derived from the power grid model database of the power grid dispatch center, containing the identification information of all bus nodes and the connection relationship information of transmission lines. The system uses each grid-connected bus as a graph node and the reciprocal of the transmission line impedance parameter as the edge weight between corresponding graph nodes to construct a weighted graph adjacency matrix. The weighted graph adjacency matrix is ​​used to characterize the connection strength between each grid-connected bus in the power grid topology; the edge weight is inversely proportional to the transmission line impedance parameter. The system extracts the three-phase voltage amplitude, three-phase voltage phase, active power, reactive power, and inverter controller parameters corresponding to each graph node. These parameters are all derived from the operation records of the power grid real-time monitoring system and the new energy equipment controller. The system integrates all extracted node parameters according to a unified dimension to construct an initial node feature matrix. The rows of the initial node feature matrix correspond to each graph node, and the columns correspond to various types of node parameters, which can comprehensively characterize the electrical operating status and equipment control status of each grid-connected bus. The system combines the weighted graph adjacency matrix and the initial node feature matrix to generate a node feature representation of a weighted graph structure. This representation can simultaneously reflect the connection relationship of the power grid topology and the operating characteristics of each node.

[0053] S22: Based on the weighted graph structure, the node feature representation adopts the spectral domain graph convolutional layer to perform layer-by-layer feature aggregation along the propagation direction of the graph adjacency matrix. The aggregated node features are compared with the reactive current injection constraints in the low-voltage crossing physical constraint boundary using regularization constraints. After fusing the multi-layer aggregation results through residual connection, the power grid topology space feature matrix is ​​generated.

[0054] Specifically, the system uses spectral domain graph convolutional layers for feature aggregation based on the node feature representation of a weighted graph structure. The spectral domain graph convolutional layers aggregate node features layer by layer along the node propagation direction represented by the adjacency matrix of the weighted graph. Each layer of the spectral domain graph convolutional layer achieves spatial propagation and fusion of node features by weighted summation of the features of adjacent nodes, gradually uncovering the correlation features between nodes in the power grid topology. The system performs a regularized constraint comparison between the aggregated node features of each layer and the reactive current injection constraints in the low-voltage crossing physical constraint boundary to determine whether the reactive current injection parameters corresponding to the aggregated node features meet the constraint requirements. The regularized constraint comparison process calculates the deviation between the node features and the constraint standard, adjusting features that do not meet the constraint requirements to ensure that the aggregated node features conform to the physical operating laws. The system fuses the aggregation results of multiple spectral domain graph convolutional layers through residual connections. Residual connections can alleviate the gradient vanishing problem in deep networks and retain the effective features obtained from each aggregation layer. The integrated and fused feature results of the system generate a power grid topology spatial feature matrix, which can comprehensively characterize the spatial correlation features of the power grid topology and the operating status features of each node.

[0055] S23: Obtain the voltage drop waveform sequence and current response waveform sequence collected by the original fault recording equipment. Perform stacked convolution operation on the voltage drop waveform sequence and current response waveform sequence according to the low voltage ride-through assessment time window. Introduce gating units between each expanded convolution layer to control the retention and forgetting of timing information. Use causal dilated convolution kernels to progressively expand the receptive field of the timing dimension to generate multi-scale fault transient timing features.

[0056] Specifically, the system acquires voltage dip waveform sequences and current response waveform sequences from the original fault recording equipment. The original fault recording equipment is deployed at each grid connection point to collect voltage and current time-series data in real time during grid fault occurrences. The collected data comprehensively reflects the voltage dip process and equipment current response process. The system divides the voltage dip waveform sequences and current response waveform sequences according to the low-voltage ride-through assessment time window, ensuring that the divided waveform sequences fully cover the low-voltage ride-through assessment period. The system uses causal-dilated convolution kernels to perform stacked convolution operations on the divided waveform sequences. Causal-dilated convolution kernels can expand the receptive field of the convolutional layer without increasing computational load. The system introduces gating units between each expanded convolutional layer. These gating units filter the time-series information output from the previous convolutional layer, controlling the retention and forgetting of time-series information, selecting time-series features that play a crucial role in assessing grid disconnection risk, and suppressing interference from invalid time-series information. The system progressively expands the receptive field along the temporal dimension using causal dilated convolution kernels. Each convolutional operation expands the receptive field, gradually capturing transient fault features at different time scales. After multiple rounds of stacked convolutions and receptive field expansion, the system generates multi-scale transient fault temporal features. These features can cover the transient change patterns of faults across different time dimensions, providing temporal support for subsequent spatiotemporal feature fusion.

[0057] S24: Based on the cross-modal attention mechanism, the power grid topology spatial feature matrix and the multi-scale fault transient temporal features are mutually attentionally weighted. The response time constraint in the low-voltage crossing physical constraint boundary is used as the auxiliary loss term for weight update. The weighted spatiotemporal features are mapped to the degrid risk probability interval through the fully connected layer to generate the low-voltage crossing risk score for each grid connection point.

[0058] Specifically, the system uses a cross-modal attention mechanism to assign mutual attention weights to the power grid topology spatial feature matrix and multi-scale fault transient temporal features. The cross-modal attention mechanism calculates the correlation between the two types of features to determine the contribution of each feature to the grid disconnection risk assessment, and then assigns corresponding attention weights. The correlation calculation is based on the similarity of the feature vectors of the two features; the higher the similarity, the greater the attention weight of the corresponding feature, ensuring that the fused features highlight key influencing factors. The system uses the response time constraint in the low-voltage crossing physical constraint boundary as an auxiliary loss term for weight updates, integrating the response time constraint into the attention weight update process. The weight allocation results are adjusted through the auxiliary loss term to ensure that the allocation of attention weights conforms to physical operating constraints. The system inputs the weighted fused spatiotemporal features into a fully connected layer. The fully connected layer performs dimensionality transformation and feature mapping on the fused features, mapping them to a grid disconnection risk probability interval. The grid disconnection risk probability interval characterizes the probability range of a grid disconnection event occurring at the grid connection point. Based on the mapping results, the system generates a low-voltage crossing risk score for each grid connection point, which quantifies the degree of grid disconnection risk for each grid connection point under the current fault scenario.

[0059] S25: During the model inference phase, the forward propagation operation of steps S21 to S24 is repeated on the same input sample at a preset random discard ratio. The low-pressure crossing risk score obtained in each forward propagation is recorded, and the cumulative amount of the squared deviation between each risk score and the mean risk score is calculated to generate the cognitive uncertainty variance.

[0060] Specifically, during the model inference phase, the system performs repeated forward propagation operations on the same input sample. These forward propagation operations correspond to steps S21 to S24. During repeated forward propagation, the system randomly discards some nodes in the network according to a preset random dropout ratio. This random dropout operation simulates the uncertainty of model parameters, thereby capturing the model's cognitive uncertainty regarding the input sample's evaluation results. The system records the low-pressure crossing risk score obtained in each forward propagation operation, and all recorded scores constitute the risk score set for the input sample. The system calculates the mean of the risk score set, which serves as the baseline risk score for the input sample. The system calculates the deviation between each risk score and the baseline risk score, squaring each deviation to obtain the squared deviation value. The system accumulates all squared deviation values ​​to obtain the cumulative squared deviation value. The system uses the cumulative squared deviation value as the cognitive uncertainty variance. This variance quantifies the model's cognitive uncertainty regarding the current input sample's evaluation results. The magnitude of the variance is inversely proportional to the reliability of the model's evaluation results, providing a quantitative basis for subsequent comprehensive reliability score calculations.

[0061] In one embodiment, S3 of the deep learning-based method for evaluating grid-connected low-voltage ride-through capability provided by the present invention specifically includes the following steps:

[0062] S31: Obtain the grid connection point voltage sequence, reactive current sequence, and response delay sequence from the low-voltage ride-through risk score and the original fault recording data. Calculate the negative correlation deviation between the low-voltage ride-through risk score and the grid connection point voltage sequence, the negative correlation deviation between the low-voltage ride-through risk score and the reactive current sequence, and the positive correlation deviation between the low-voltage ride-through risk score and the response delay sequence. Weight the three deviations according to the strictness of each constraint in the low-voltage ride-through physical constraint boundary to generate a physical consistency index.

[0063] Specifically, the system acquires a low-voltage ride-through risk score and extracts the grid connection point voltage sequence, reactive current sequence, and response delay sequence from the original fault recording data. The system calculates the correlation deviation between the low-voltage ride-through risk score and the three sequences. The low-voltage ride-through risk score shows a negative correlation with both the grid connection point voltage sequence and the reactive current sequence, and a positive correlation with the response delay sequence. The system calculates the negative correlation deviation using the following formula:

[0064]

[0065] in, The negative correlation deviation between the low-voltage ride-through risk score and the grid connection point voltage series is represented by S, where S is the low-voltage ride-through risk score series and U is the grid connection point voltage series. The covariance calculation function, This is the variance calculation function. The formula for calculating the negative correlation deviation between the low-voltage ride-through risk score and the reactive current sequence is the same as the formula above, except that U is replaced by the reactive current sequence I. The system calculates the positive correlation deviation using the following formula:

[0066]

[0067] in, The deviation between the low-voltage crossing risk score and the response delay sequence is represented by T, where T is the response delay sequence. The system extracts the stringency of each constraint in the physical constraint boundary of the low-voltage crossing and converts it into a weighted weight of the corresponding deviation, using the formula... The three deviations are weighted and fused to generate a physical consistency index. This is a physical consistency index. The weights are the values ​​corresponding to each degree of deviation. The weighted average of the corresponding deviations is 1.

[0068] S32: Obtain the synthetic off-net sample set and the real off-net sample set from the equalized training dataset. Map the synthetic off-net sample and the real off-net sample to the regenerable kernel Hilbert space through kernel functions. Calculate the norm square of the kernel mean embedding difference between the two sample sets in the regenerable kernel Hilbert space to generate the distribution matching index.

[0069] Specifically, the system acquires a synthetic off-grid sample set and a real off-grid sample set from the balanced training dataset. Both sample sets contain standardized off-grid-related feature data. The system selects a preset kernel function and maps the synthetic and real off-grid samples to a regenerable kernel Hilbert space through this kernel function, achieving high-dimensional spatial mapping of sample features. The kernel function is used to characterize the similarity between sample features. The system calculates the kernel mean embedding of the two sample sets in the regenerable kernel Hilbert space. The kernel mean embedding is used to describe the distribution characteristics of the sample sets in the high-dimensional space. The calculation formula is as follows:

[0070]

[0071] in, Let n be the kernel mean embedding of the sample set X, k be the preset kernel function, and x be a single sample in the sample set. The system calculates the norm square of the difference between the kernel mean embeddings of the two sample sets and generates the distribution matching index using the formula:

[0072]

[0073] in, The distribution matching index, To synthesize the kernel mean embedding of the off-net sample set, For the kernel mean embedding of the real off-net sample set, ||·|| is the norm calculation symbol, which can characterize the distribution similarity between two sample sets.

[0074] S33: Obtain the node adjacency relationship and electrical distance between nodes in the power grid topology. Using the reciprocal of the electrical distance as the spatial smoothness weight, calculate the mean of the weighted squared difference of the low-voltage crossing risk score between each node and its adjacent nodes. Sum the mean values ​​of all nodes and divide by the total number of nodes to generate the topology consistency index.

[0075] Specifically, the system acquires the node adjacency relationships and electrical distances between nodes in the power grid topology. Node adjacency relationships are used to determine the neighboring nodes of each node, and electrical distances characterize the degree of electrical connection between nodes. The system performs a reciprocal calculation on the electrical distances between nodes, using the result as a spatial smoothness weight; a larger weight value indicates a stronger electrical connection between the corresponding two nodes. The system calculates the weighted squared difference in low-voltage ride-through risk scores between each node and its neighboring nodes, using the following formula:

[0076]

[0077] in, Let be the weighted squared mean of the differences at the i-th node. Let be the number of neighboring nodes of the i-th node. Let be the set of neighboring nodes of the i-th node. The spatial smoothness weights between the i-th node and the j-th node are... Score the low-pressure crossing risk for the i-th node. The low-pressure crossing risk score is assigned to the j-th node. The system sums the weighted squared differences of all nodes, then divides the sum by the total number of nodes using the formula... Generate a topology consistency index. V is the topology consistency index, where V is the total number of nodes in the power grid topology. This index can characterize the spatial matching between the low-voltage ride-through risk score and the power grid topology.

[0078] S34: Obtain the cognitive uncertainty variance and the preset maximum prediction variance of the validation set. Then, sum the ratio of the cognitive uncertainty variance of each node to the maximum prediction variance of the validation set at the node level and divide it by the total number of nodes to generate the cognitive uncertainty index.

[0079] Specifically, the system acquires the cognitive uncertainty variance and a preset maximum prediction variance for the validation set, where the maximum predicted variance is the highest observed cognitive uncertainty variance in the preset validation set data. The system calculates the ratio of the cognitive uncertainty variance for each node to the maximum predicted variance for the validation set; this ratio is used to normalize the cognitive uncertainty variance and eliminate the influence of dimensions. The system introduces an exponential decay factor, which adjusts the influence of the cognitive uncertainty variance on the cognitive uncertainty index. The exponential decay factor is determined by the confidence index and time decay coefficient output by the model for the current evaluation result. Preferably, the system generates the uncertainty index using a preset formula, the formula for calculating the cognitive uncertainty index being:

[0080]

[0081] in, To understand the uncertainty index, Let be the total number of nodes in the node set of the power grid topology, and i be the node index. Let be the variance of cognitive uncertainty at node i. The maximum prediction variance observed on the predefined validation set. It is an exponential decay factor. This is the confidence index output by the model for the current evaluation results. This is the time decay coefficient. The system accumulates the relevant calculation results of each node at the node level, and then divides the accumulated result by the total number of nodes to obtain the final cognitive uncertainty index. This index is used to quantify the degree of cognitive uncertainty of the model regarding the evaluation results.

[0082] In one embodiment, S4 of the deep learning-based method for evaluating grid-connected low-voltage ride-through capability provided by the present invention specifically includes the following steps:

[0083] S41: Obtain the physical consistency index, distribution matching index, topological consistency index, and cognitive uncertainty index. Assign fusion weights to each index according to their importance under extreme small sample imbalance conditions. Sum the products of each index and its corresponding weights to generate a comprehensive credibility score.

[0084] Specifically, the system acquires four separately generated indices: physical consistency index, distribution matching index, topological consistency index, and cognitive uncertainty index. These four indices quantify the reliability of the model evaluation results from different dimensions, collectively supporting the calculation of the comprehensive credibility score. For evaluation needs under extreme small-sample, disequilibrium conditions, the system analyzes the impact of the four indices on the reliability of the evaluation results and assigns corresponding fusion weights according to the importance of each index under these conditions. The allocation of fusion weights must adhere to the principle that the sum of the weights is 1. The weight allocation result depends on the supporting role of each index in the accuracy of the disconnection risk assessment under extreme small-sample scenarios. The system generates a comprehensive credibility score by weighted summation of the indices using a preset formula. The calculation formula is as follows:

[0085]

[0086] in, To assess overall credibility, The fusion weights are the four credibility indices. The credibility indices for the four dimensions are, in order: physical consistency index, distribution matching index, topological consistency index, and cognitive uncertainty index. The system first multiplies each credibility index by its corresponding fusion weight to obtain a weighted contribution value for each index. Then, all weighted contribution values ​​are summed to obtain a comprehensive credibility score. This comprehensive credibility score integrates reliability information from all four dimensions, comprehensively reflecting the overall reliability of the model evaluation results and providing a core basis for subsequent confidence level classification. The system fully adapts to the characteristics of extremely small, imbalanced datasets during weight allocation, ensuring that the weighted fusion results accurately reflect the differences in importance among the indices.

[0087] S42: Obtain the comprehensive confidence score and the preset first threshold and the preset second threshold. Compare the comprehensive confidence score with the first threshold and the second threshold in turn. Based on the judgment result, divide the confidence level of the current evaluation result into three level intervals: high confidence interval, medium confidence interval and low confidence interval, and generate a confidence level mapping table.

[0088] Specifically, the system obtains a comprehensive confidence score, along with a preset first threshold and a preset second threshold. These two thresholds are used to classify the confidence level of the assessment result, with the first threshold being lower than the second threshold. The system compares the comprehensive confidence score sequentially with the first and second thresholds, determining the range within which the comprehensive confidence score falls through this continuous comparison process, and then classifying the confidence level of the current assessment result based on this range. The system uses a preset formula to classify the confidence level:

[0089]

[0090] Where Level represents the confidence level. To assess overall credibility, The first threshold is preset. The system uses a preset second threshold. Based on the division rules of the above formula, the system clearly divides the confidence level of the current evaluation result into three levels: high confidence, medium confidence, and low confidence. Each level corresponds to a different level of reliability for the evaluation result. The system associates and organizes the range of the comprehensive confidence score with the corresponding confidence level, generating a confidence level mapping table. This table clarifies the confidence level standards corresponding to different comprehensive confidence score ranges, providing a unified basis for the subsequent classification and output of evaluation results. This ensures that the confidence level division logic of the evaluation results is consistent and traceable, meeting the standardization requirements in engineering applications.

[0091] S43: When the overall credibility score is in the high credibility range, the low-voltage crossing risk score is directly output as the evaluation result; when the overall credibility score is in the medium credibility range, the low-voltage crossing risk score, the overall credibility score, and the credibility indicators of each dimension are encapsulated together and output, and an active review signal is triggered; when the overall credibility score is in the low credibility range, the feature vector of the current input sample and the placeholder of the label to be labeled are encapsulated together to generate a low credibility sample feedback signal.

[0092] Specifically, the system calls the confidence level mapping table, combines the specific value of the comprehensive confidence score, determines its confidence level interval, and then performs corresponding evaluation result processing and output operations according to different confidence level intervals. When the comprehensive confidence score is in the high confidence interval, it indicates that the reliability of the model evaluation result meets the preset standard and can directly support power grid dispatching decisions. The system directly outputs the low-voltage ride-through risk score as the evaluation result, and the output content only includes the low-voltage ride-through risk score, meeting the need for rapid acquisition of core evaluation information in dispatching decisions. When the comprehensive confidence score is in the medium confidence interval, it indicates that the reliability of the model evaluation result is at a medium level, and more comprehensive information is needed to assist in the judgment. The system integrates and encapsulates the low-voltage ride-through risk score with the comprehensive confidence score and confidence indicators of various dimensions. The encapsulated output content contains complete evaluation data and reliability information. At the same time, the system triggers an active review signal, which is used to prompt relevant personnel to conduct further manual review of the evaluation result to ensure the accuracy of the evaluation result.

[0093] When the overall credibility score is in the low confidence range, it indicates that the reliability of the model evaluation results has not met the preset standard, and the model needs to be optimized by supplementing samples. The system extracts the feature vector of the current input sample, which contains all the running parameters and feature information of the current sample. At the same time, it generates placeholders for labels to be labeled, which are used for subsequent manual labeling of the sample. The system generates a low confidence sample feedback signal using a formula:

[0094]

[0095] in, This is a feedback signal from a low-confidence sample. The feature vector of the current input sample. This serves as a placeholder for the labels to be labeled. The system encapsulates the feature vector of the current input sample with the placeholder labels to be labeled, generating a low-confidence sample feedback signal. This feedback signal is used for subsequent sample set updates and optimizations, improving the evaluation reliability of the model under extremely small sample imbalance conditions.

[0096] In one embodiment, S5 of the deep learning-based method for evaluating grid-connected low-voltage ride-through capability provided by the present invention specifically includes the following steps:

[0097] S51: When the overall credibility score is lower than the first threshold, the low-pressure crossing risk score of the current sample is comprehensively packaged with the overall credibility score, physical consistency index, distribution matching index, topological consistency index and cognitive uncertainty index, and the packaged data is used as the high credibility assessment result. The low-pressure crossing capability assessment result is output as a high credibility assessment report containing node identifiers, risk scores and detailed values ​​of credibility indicators for each dimension.

[0098] Specifically, the system acquires a comprehensive credibility score and a first threshold. It compares the comprehensive credibility score with the first threshold. When the comprehensive credibility score is lower than the first threshold, the system performs the encapsulation and output of the high-credibility assessment result. The system acquires the low-pressure crossing risk score of the current sample, as well as the physical consistency index, distribution matching index, topological consistency index, and cognitive uncertainty index. These parameters together constitute the complete data system of the assessment result. The system performs comprehensive encapsulation processing on all the above parameters. The encapsulation process follows a preset data format specification to ensure that the encapsulated data structure is unified and parsable. The expression for the encapsulated data is:

[0099]

[0100] in, This is the encapsulated, high-reliability evaluation data. This is the node identifier corresponding to the current sample. Assess the low-pressure crossing risk score for the current sample. This is a physical consistency index. The distribution matching index, This is the topological consistency index. To address the uncertainty index, the system uses the encapsulated data as a high-reliability assessment result, further generating a high-reliability assessment report. This report includes node identifiers, low-voltage ride-through risk scores, and detailed values ​​for reliability indicators across various dimensions, clearly presenting the assessment results and the basis for reliability in each dimension. The system organizes the assessment report according to a preset output format, ensuring completeness and logical clarity, enabling direct provision to the power grid dispatching system and relevant maintenance personnel. This provides reliable data support for power grid safety dispatching decisions and achieves standardized output of high-reliability assessment results.

[0101] S52: When the overall credibility score is higher than the second threshold, the feature vector of the current input sample and the placeholder of the label to be labeled are encapsulated to generate a low credibility sample feedback signal.

[0102] Specifically, the system acquires a comprehensive credibility score and a second threshold. It compares the comprehensive credibility score with the second threshold; when the comprehensive credibility score is higher than the second threshold, the system generates a low-credibility sample feedback signal. The system extracts the feature vector of the current input sample. This feature vector contains all the operating parameters and feature information of the current sample, covering the grid connection point voltage sequence, current response sequence, control parameters, and grid topology-related features, comprehensively reflecting the current sample's operating condition. The system generates placeholder labels to be labeled, reserving space for subsequent manual labeling to facilitate the addition of labels to the low-credibility sample and improve its information.

[0103] Furthermore, the system encapsulates the feature vector and placeholder labels of the current input sample. The encapsulation process follows a preset signal format specification to ensure that the encapsulated signal can be recognized and processed by subsequent steps. The expression for generating the low-confidence sample feedback signal is:

[0104]

[0105] in, This is a feedback signal from a low-confidence sample. The feature vector of the current input sample. This is a placeholder for the labels to be labeled. After the system is encapsulated, a low-confidence sample feedback signal is generated. This signal is used to trigger subsequent sample set updates and model parameter optimization operations, solving the problem of low evaluation reliability caused by insufficient samples under extremely small and imbalanced conditions.

[0106] S53: Return the low-confidence sample feedback signal to step S1, triggering the conditional variational autoencoder to use the feature vector in the low-confidence sample feedback signal as the new real off-grid sample for iterative update of generator parameters. That is, based on the joint distribution change of the new real off-grid sample and the original real off-grid sample set, the encoder and decoder parameters of the generator are alternately optimized by maximizing the variational lower bound term in the joint loss function. The projection distribution of the new real off-grid sample in the latent space is used as the density guiding signal of the sampling prior. By adjusting the latent space sampling weight and the strength coefficient of the constraint violation penalty term of the conditional variational autoencoder, the generator is made to preferentially sample the latent space region where the low-confidence sample is located to generate an updated synthetic off-grid sample set.

[0107] Specifically, the system returns the low-confidence sample feedback signal generated in step S52 to step S1, triggering the conditional variational autoencoder to perform parameter iterative update processing. The system extracts the feature vector from the low-confidence sample feedback signal and adds this feature vector as a new real off-grid sample to the original real off-grid sample set, forming an updated real off-grid sample set. The system calculates the joint distribution change between the new real off-grid sample and the original real off-grid sample set, and adjusts the training strategy of the conditional variational autoencoder based on the joint distribution change. By maximizing the variational lower bound term in the joint loss function, the encoder and decoder parameters of the generator are alternately optimized. The expression for the variational lower bound term of the joint loss function is:

[0108]

[0109] in, For the variational lower bound term, Let be the posterior distribution of the encoder. For the generation distribution of the decoder, Let x be the prior distribution, KL be the KL divergence calculation function, x be the sample features, z be the latent space vector, and y be the physical constraints.

[0110] Furthermore, the system adds real off-grid samples as input to the encoder of the conditional variational autoencoder to obtain the projected distribution of the sample in the latent space. This projected distribution is used as the density guiding signal for sampling prior. The system adjusts the latent space sampling weights and the strength coefficients of the constraint violation penalty terms of the conditional variational autoencoder, so that the generator prioritizes sampling in the latent space region where low-confidence samples are located during the sampling process, ensuring that the generated synthetic off-grid samples can cover the working scenarios corresponding to the low-confidence samples. Through the above parameter optimization and sampling adjustment, the system generates an updated synthetic off-grid sample set. The updated synthetic off-grid sample set can supplement the deficiencies of the original sample set, improve the completeness and representativeness of the sample set, provide more reliable data support for subsequent model training, and further improve the reliability of the evaluation method.

[0111] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0112] Based on the same inventive concept, this application also provides a deep learning-based grid-connected low-voltage ride-through capability assessment system for implementing the aforementioned deep learning-based grid-connected low-voltage ride-through capability assessment method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more deep learning-based grid-connected low-voltage ride-through capability assessment system embodiments provided below can be found in the limitations of the deep learning-based grid-connected low-voltage ride-through capability assessment method described above, and will not be repeated here.

[0113] Preferably, such as Figure 2 As shown, this invention provides a deep learning-based grid-connected low-voltage ride-through capability assessment system 600, which is configured with the following modules:

[0114] The balanced sample construction module 610 is used to obtain the real off-grid sample set and the low-voltage crossing physical constraint boundary. Based on the real off-grid sample set, a synthetic off-grid sample set is generated under the guidance of the low-voltage crossing physical constraint boundary. The synthetic off-grid sample set is then merged with the normal sample set to generate a balanced training dataset.

[0115] The dual-stream feature modeling module 620 is used to input the equalized training dataset into the preset graph convolution-temporal convolution dual-stream fusion network, extract the topological spatial features of the power grid and the transient temporal features of the fault, and complete the model training under the constraint of the physical constraint boundary of low-voltage crossing, generating the low-voltage crossing risk score and cognitive uncertainty variance for each grid connection point.

[0116] The multi-dimensional consistency verification module 630 is used to verify the degree of violation based on the low-voltage crossing risk score and the low-voltage crossing physical constraint boundary, generate a physical consistency index, evaluate the spatial smoothness of the low-voltage crossing risk score and the obtained power grid topology, generate a topology consistency index, and calculate the distribution matching index between synthetic samples and real samples based on the equalized training dataset, and generate a cognitive uncertainty index by combining the cognitive uncertainty variance.

[0117] The comprehensive credibility assessment module 640 is used to generate a comprehensive credibility score by weighted summation of the physical consistency index, distribution matching index, topological consistency index and cognitive uncertainty index based on preset weights, and to map the comprehensive credibility score to a preset first threshold and second threshold to generate a low-pressure crossing capability assessment result, wherein the first threshold is less than the second threshold.

[0118] The decision iteration feedback module 650 is used to mark the current evaluation result as a high-confidence evaluation result and output the low-pressure crossing capability evaluation result when the comprehensive confidence score is lower than the first threshold; when the comprehensive confidence score is higher than the second threshold, it generates a low-confidence sample feedback signal and returns to step S1 to update the synthesized offline sample set.

[0119] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described deep learning-based method for evaluating grid-connected low-voltage ride-through capability.

[0120] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described deep learning-based method for evaluating grid-connected low-voltage ride-through capability.

[0121] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0122] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0123] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based method for evaluating grid-connected low-voltage ride-through capability, characterized in that, Includes the following steps: S1: Obtain the real offline sample set and the low-voltage crossing physical constraint boundary. Based on the real offline sample set, generate a synthetic offline sample set under the guidance of the low-voltage crossing physical constraint boundary. Merge the synthetic offline sample set with the normal sample set to generate a balanced training dataset. S2: Input the equalized training dataset into a preset graph convolution-temporal convolution dual-stream fusion network, extract the grid topology spatial features and fault transient temporal features, and complete the model training under the constraints of the low-voltage crossing physical constraint boundary to generate the low-voltage crossing risk score and cognitive uncertainty variance for each grid connection point; S3: Based on the low-voltage crossing risk score and the low-voltage crossing physical constraint boundary, the degree of violation is verified, a physical consistency index is generated, and the spatial smoothness of the low-voltage crossing risk score and the obtained power grid topology is evaluated to generate a topology consistency index. At the same time, the distribution matching degree index between the synthetic sample and the real sample is calculated based on the equalized training dataset, and the cognitive uncertainty index is generated by combining the cognitive uncertainty variance. S4: Based on preset weights, the physical consistency index, the distribution matching degree index, the topological consistency index, and the cognitive uncertainty index are weighted and summed to generate a comprehensive credibility score. The comprehensive credibility score is then mapped to a preset first threshold and a second threshold to generate a low-voltage crossing capability assessment result, wherein the first threshold is less than the second threshold. S5: When the overall credibility score is lower than the first threshold, mark the current evaluation result as a high credibility evaluation result and output the low-pressure ride-through capability evaluation result; When the overall credibility score is higher than the second threshold, a low credibility sample feedback signal is generated and the process returns to step S1 to update the synthetic off-net sample set.

2. The method according to claim 1, characterized in that, S1 includes: S11: Obtain the real off-grid sample set and the reactive current response time constraint, voltage recovery time constraint and reactive current injection depth constraint in the physical constraint boundary of low voltage crossing. Decompose each sample in the real off-grid sample set into voltage drop waveform segment, current response waveform segment and control parameter segment, and perform sliding window segmentation and normalized scale compression processing on each waveform segment to generate a standardized off-grid sample feature sequence. S12: Construct a conditional variational autoencoder that uses the low-voltage crossing physical constraint boundary as the feasible domain constraint of the decoding output. Encode the standardized off-grid sample feature sequence using the conditional variational autoencoder. Perform joint loss minimization iterative training with the latent space encoding result and the constraint violation penalty term in the low-voltage crossing physical constraint boundary to generate a latent space distribution modulated by physical constraints. S13: Perform multiple random samplings from the latent space distribution modulated by physical constraints. After each sampling, use the low-voltage crossing of the physical constraint boundary to filter the decoding output for constraint violation thresholds. Retain the decoding outputs that do not trigger the constraint violation thresholds as candidate synthetic samples. After performing waveform segment splicing and scale restoration processing on all candidate synthetic samples, merge them with the normal sample set to generate an equalized training dataset.

3. The method according to claim 1, characterized in that, S2 includes: S21: Obtain the set of bus nodes and the set of transmission line edges in the power grid topology. Construct a weighted graph adjacency matrix with each grid connection point bus as a graph node and the reciprocal of the transmission line impedance parameter as the edge weight. Extract the three-phase voltage amplitude, three-phase voltage phase, active power, reactive power and inverter controller parameters of each node as the initial feature matrix of the node, and generate the node feature representation of the weighted graph structure. S22: Based on the node feature representation of the weighted graph structure, the spectral domain graph convolutional layer is used to perform layer-by-layer feature aggregation along the propagation direction of the graph adjacency matrix. The aggregated node features are compared with the reactive current injection constraints in the low-voltage crossing physical constraint boundary using regularization constraints. After fusing the multi-layer aggregation results through residual connections, a power grid topology space feature matrix is ​​generated. S23: Obtain the voltage drop waveform sequence and current response waveform sequence collected by the original fault recording equipment. Perform stacked convolution operation on the voltage drop waveform sequence and the current response waveform sequence according to the low voltage ride-through assessment time window. Introduce a gating unit between each expanded convolution layer to control the retention and forgetting of timing information. Use the causal dilated convolution kernel to progressively expand the receptive field of the timing dimension to generate multi-scale fault transient timing features. S24: Based on the cross-modal attention mechanism, the power grid topology spatial feature matrix and the multi-scale fault transient temporal features are mutually attentionally weighted. The response time constraint in the low-voltage crossing physical constraint boundary is used as the auxiliary loss term for weight update. The weighted and fused spatiotemporal features are mapped to the degrid risk probability interval through the fully connected layer to generate a low-voltage crossing risk score for each grid connection point. S25: During the model inference phase, the forward propagation operation of steps S21 to S24 is repeated on the same input sample at a preset random discard ratio. The low-pressure crossing risk score obtained in each forward propagation is recorded, and the cumulative amount of the squared deviation between each risk score and the mean risk score is calculated to generate the cognitive uncertainty variance.

4. The method according to claim 1, characterized in that, S3 includes: S31: Obtain the grid connection point voltage sequence, reactive current sequence, and response delay sequence from the low-voltage ride-through risk score and the original fault recording data. Calculate the negative correlation deviation between the low-voltage ride-through risk score and the grid connection point voltage sequence, the negative correlation deviation between the low-voltage ride-through risk score and the reactive current sequence, and the positive correlation deviation between the low-voltage ride-through risk score and the response delay sequence, respectively. Weight and fuse the three deviations according to the strictness of each constraint in the low-voltage ride-through physical constraint boundary to generate a physical consistency index. S32: Obtain the synthetic off-net sample set and the real off-net sample set in the equalized training dataset, map the synthetic off-net sample and the real off-net sample to the regenerable kernel Hilbert space through the kernel function respectively, calculate the norm square of the kernel mean embedding difference of the two sample sets in the regenerable kernel Hilbert space, and generate the distribution matching index. S33: Obtain the node adjacency relationship and electrical distance between nodes in the power grid topology. Using the reciprocal of the electrical distance as the spatial smoothness weight, calculate the mean of the weighted square difference of the low-voltage crossing risk score between each node and its adjacent nodes. Sum the mean values ​​of all nodes and divide by the total number of nodes to generate the topology consistency index. S34: Obtain the cognitive uncertainty variance and the preset maximum prediction variance of the validation set. Then, sum the ratio of the cognitive uncertainty variance of each node to the maximum prediction variance of the validation set at the node level and divide it by the total number of nodes to generate the cognitive uncertainty index.

5. The method according to claim 4, characterized in that, The formula for calculating the cognitive uncertainty index is as follows: in, To understand the uncertainty index, Let be the total number of nodes in the node set of the power grid topology, and i be the node index. Let be the variance of cognitive uncertainty at node i. The maximum prediction variance observed on the predefined validation set. It is an exponential decay factor. This is the confidence index output by the model for the current evaluation results. This is the time decay coefficient.

6. The method according to claim 1, characterized in that, S4 includes: S41: Obtain the physical consistency index, the distribution matching index, the topological consistency index, and the cognitive uncertainty index; assign fusion weights to each index according to their importance under extreme small sample imbalance conditions; sum the products of each index and its corresponding weights to generate a comprehensive credibility score. S42: Obtain the comprehensive confidence score and the preset first threshold and the preset second threshold, compare the comprehensive confidence score with the first threshold and the second threshold in turn, and divide the confidence level of the current evaluation result into three level intervals: high confidence interval, medium confidence interval and low confidence interval according to the judgment result, and generate a confidence level mapping table. S43: When the overall credibility score is in the high credibility range, the low-pressure crossing risk score is directly output as the evaluation result; when the overall credibility score is in the medium credibility range, the low-pressure crossing risk score, the overall credibility score, and the credibility indicators of each dimension are encapsulated together and output, and an active verification signal is triggered; when the overall credibility score is in the low credibility range, the feature vector of the current input sample and the placeholder of the label to be labeled are encapsulated together to generate a low credibility sample feedback signal.

7. The method according to any one of claims 1-6, characterized in that, S5 includes: S51: When the comprehensive credibility score is lower than the first threshold, the low-pressure crossing risk score of the current sample is comprehensively packaged with the comprehensive credibility score, physical consistency index, distribution matching index, topological consistency index and cognitive uncertainty index, and the packaged data is used as the high credibility assessment result. The low-pressure crossing capability assessment result is output as a high credibility assessment report containing node identifier, risk score and subdivided values ​​of credibility indicators of each dimension. S52: When the overall credibility score is higher than the second threshold, the feature vector of the current input sample and the placeholder of the label to be labeled are encapsulated to generate a low credibility sample feedback signal. S53: Return the low-confidence sample feedback signal to step S1, triggering the conditional variational autoencoder to use the feature vector in the low-confidence sample feedback signal as the new real off-grid sample for generator parameter iterative update processing. That is, based on the joint distribution change of the new real off-grid sample and the original real off-grid sample set, the encoder and decoder parameters of the generator are alternately optimized by maximizing the variational lower bound term in the joint loss function. The projection distribution of the new real off-grid sample in the latent space is used as the density guiding signal of the sampling prior. By adjusting the latent space sampling weight and the strength coefficient of the constraint violation penalty term of the conditional variational autoencoder, the generator is made to preferentially sample the latent space region where the low-confidence sample is located to generate an updated synthetic off-grid sample set.

8. A deep learning-based system for evaluating grid-connected low-voltage ride-through capability, characterized in that, The system includes: The balanced sample construction module is used to obtain the real off-grid sample set and the low-voltage crossing physical constraint boundary. Based on the real off-grid sample set, a synthetic off-grid sample set is generated under the guidance of the low-voltage crossing physical constraint boundary. The synthetic off-grid sample set is then merged with the normal sample set to generate a balanced training dataset. The dual-stream feature modeling module is used to input the equalized training dataset into a preset graph convolutional-temporal convolutional dual-stream fusion network, extract the grid topology spatial features and fault transient temporal features, and complete model training under the constraints of the low-voltage crossing physical constraint boundary to generate a low-voltage crossing risk score and cognitive uncertainty variance for each grid connection point. The multi-dimensional consistency verification module is used to verify the degree of violation based on the low-voltage crossing risk score and the low-voltage crossing physical constraint boundary, generate a physical consistency index, and evaluate the spatial smoothness of the low-voltage crossing risk score and the obtained power grid topology to generate a topology consistency index. At the same time, it calculates the distribution matching degree index between the synthetic sample and the real sample based on the equalized training dataset, and generates a cognitive uncertainty index by combining the cognitive uncertainty variance. The comprehensive credibility assessment module is used to generate a comprehensive credibility score by weighted summation of the physical consistency index, the distribution matching index, the topological consistency index and the cognitive uncertainty index based on preset weights, and to perform interval mapping between the comprehensive credibility score and preset first threshold and second threshold to generate a low-voltage crossing capability assessment result, wherein the first threshold is less than the second threshold; The decision iteration feedback module is used to mark the current evaluation result as a high-confidence evaluation result and output the low-pressure crossing capability evaluation result when the comprehensive confidence score is lower than the first threshold; when the comprehensive confidence score is higher than the second threshold, it generates a low-confidence sample feedback signal and returns to step S1 to update the synthetic off-grid sample set.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images