Stability evaluation method and device of power system, computer equipment, readable storage medium and program product

Abstract

The invention relates to a stability evaluation method and device of a power system, computer equipment, a computer readable storage medium and a computer program product. Comprising the following steps: for any iteration training, acquiring data from a stability data set of a power system, and performing feature extraction on steady-state operation data, fault process data and power grid topology data to obtain fusion latent variable features; inputting the fused latent variable features into a stability evaluation network to obtain a stability result and a margin result; determining stability loss and margin loss based on the stability result, the stability true value, the margin result and the margin true value; determining a stability weight and a margin weight; based on the stability loss, the margin loss, the KL divergence regular term, the stability weight and the margin weight, determining the total loss of the model, further judging whether a preset condition is met or not, and if not, continuing iterative training; if yes, the training is completed. According to the method, stability prediction and margin prediction can be collaboratively optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a stability evaluation method and device of a power system, a computer device, a computer readable storage medium and a computer program product. BACKGROUND

[0002] With the continuous expansion of the scale of the power system and the deep access of new energy and power electronic devices, the operation state of the power system presents high dynamicity, nonlinearity and multi-dimensional complexity. As an important indicator for measuring whether the power system can maintain synchronous operation after suffering from serious disturbance, transient stability is one of the core problems of power dispatching and safety analysis.

[0003] In recent years, with the development of information technology and artificial intelligence, data-driven TSA (Transient Stability Assessment) methods based on deep learning have gradually become a research hotspot in this field. The existing methods usually model transient stability discrimination and margin prediction as independent tasks, which fails to fully exploit the potential correlation features between the two, making it difficult to ensure the reliability and stability of the evaluation results. SUMMARY

[0004] Therefore, it is necessary to provide a stability evaluation method, device, computer device, computer readable storage medium and computer program product of a power system, which can cooperatively optimize transient stability prediction and margin prediction, in view of the above technical problems.

[0005] In a first aspect, the present application provides a stability evaluation method of a power system, comprising:

[0006] For each iteration training, steady-state operation data, fault process data and power grid topology data are obtained from a pre-established stability data set of the power system, and features are extracted from the steady-state operation data, the fault process data and the power grid topology data to obtain steady-state features, fault features and topology features; the steady-state features, the fault features and the topology features are respectively subjected to linear transformation, latent variable sampling, low-rank interaction and fusion operation to obtain fused latent variable features;

[0007] The fused latent variable features are input into a feature extraction layer in a stability evaluation network of the previous iteration process to obtain power system features; the power system features are respectively input into a stability discrimination layer and a margin prediction layer in the stability evaluation network of the previous iteration process to obtain stability results and margin results;

[0008] The model obtains the true stability and margin values ​​from a pre-established power system stability dataset. Based on the stability results, the true stability values, the margin results, and the true margin values, it determines the stability loss and margin loss for the current iteration. Based on the Gaussian distribution of stability weights and margin weights from the previous iteration, it determines the stability weights and margin weights for the current iteration through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.

[0009] Based on the total loss of the model, it is determined whether the preset conditions have been met. If not, the stability evaluation network of the previous iteration process, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and the next iteration training is performed. If the preset conditions are met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are obtained. The obtained steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system.

[0010] In one embodiment, the process of performing linear transformation, latent variable sampling, low-rank interaction, and fusion operations on the steady-state features, the fault features, and the topological features to obtain fused latent variable features includes:

[0011] Linear transformations and latent variable sampling are performed on the steady-state features, fault features, and topological features respectively to obtain steady-state latent variables, fault latent variables, and topological latent variables; two-modal low-rank interactions and three-modal low-rank interactions are performed on the steady-state latent variables, fault latent variables, and topological latent variables to obtain multiple interaction results; the steady-state features, fault features, topological features, and multiple interaction results are input into a fusion neural network to obtain fused latent variable features.

[0012] In one embodiment, the process of performing linear transformations and latent variable sampling on the steady-state features, the fault features, and the topological features to obtain steady-state latent variables, fault latent variables, and topological latent variables includes:

[0013] The steady-state features, fault features, and topological features are respectively input into a preset neural network to obtain steady-state mean, steady-state variance, fault mean, fault variance, topological mean, and topological variance. Latent variables are sampled from the steady-state mean and steady-state variance to obtain steady-state latent variables. Latent variables are sampled from the fault mean and fault variance to obtain fault latent variables. Latent variables are sampled from the topological mean and topological variance to obtain topological latent variables.

[0014] In one embodiment, the two-modal low-rank interaction and three-modal low-rank interaction are performed on the steady-state latent variable, the fault latent variable, and the topological latent variable to obtain multiple interaction results, including:

[0015] A first interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the fault latent variable; a second interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the topological latent variable; a third interaction result is obtained by performing a two-modal low-rank interaction between the fault latent variable and the topological latent variable; and a fourth interaction result is obtained by performing a three-modal low-rank interaction between the steady-state latent variable, the fault latent variable, and the topological latent variable. The first, second, third, and fourth interaction results together constitute multiple interaction results.

[0016] In one embodiment, steady-state operation data, fault process data, and grid topology data of the power system are acquired. These acquired data are then input into a trained stability assessment network to obtain the stability results of the power system, including:

[0017] Steady-state test data, fault test data, and topology test data of the power system are acquired. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state test features, fault test features, and topology test features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are performed on the steady-state features, fault test features, and topology test features respectively to obtain fused test features. The fused test features are input into the trained stability assessment network to obtain stability prediction results and margin prediction results. Based on the stability prediction results and margin prediction results, it is determined whether the power system is stable.

[0018] In one embodiment, the method further includes:

[0019] If the power system is in a stable state, monitoring continues; if the power system is in an unstable state, an alarm is issued.

[0020] Secondly, this application also provides a power system stability assessment device, comprising:

[0021] The acquisition module is used to acquire steady-state operation data, fault process data, and grid topology data from a pre-established power system stability dataset for any given training iteration; extract features from the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features; and perform linear transformation, latent variable sampling, low-rank interaction, and fusion operations on the steady-state features, the fault features, and the topology features to obtain fused latent variable features.

[0022] The first extraction module is used to input the fused latent variable features into the feature extraction layer of the stability evaluation network of the previous iteration process to obtain power system features; and to input the power system features into the stability discrimination layer and margin prediction layer of the stability evaluation network of the previous iteration process to obtain stability results and margin results.

[0023] The second extraction module is used to obtain the true stability value and true margin value from a pre-established power system stability dataset; based on the stability result, the true stability value, the margin result, and the true margin value, it determines the stability loss and margin loss of the current iteration process; based on the Gaussian distribution of stability weights and margin weights of the previous iteration process, it determines the stability weights and margin weights of the current iteration process through reparameterized sampling; based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.

[0024] The judgment module is used to determine whether a preset condition has been met based on the total loss of the model. If not, it optimizes the stability evaluation network of the previous iteration process and the Gaussian distribution parameters of the stability weights and margin weights based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and performs the next iteration training. If the preset condition has been met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are acquired. The acquired steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system.

[0025] Thirdly, 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 perform the following steps:

[0026] For any given training iteration, steady-state operation data, fault process data, and grid topology data are obtained from a pre-established power system stability dataset. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are then performed on the steady-state features, the fault features, and the topology features to obtain fused latent variable features.

[0027] The fused latent variable features are input into the feature extraction layer of the stability assessment network of the previous iteration process to obtain power system features; the power system features are then input into the stability discrimination layer and margin prediction layer of the stability assessment network of the previous iteration process to obtain stability results and margin results.

[0028] The model obtains the true stability and margin values ​​from a pre-established power system stability dataset. Based on the stability results, the true stability values, the margin results, and the true margin values, it determines the stability loss and margin loss for the current iteration. Based on the Gaussian distribution of stability weights and margin weights from the previous iteration, it determines the stability weights and margin weights for the current iteration through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.

[0029] Based on the total loss of the model, it is determined whether the preset conditions have been met. If not, the stability evaluation network of the previous iteration process, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and the next iteration training is performed. If the preset conditions are met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are obtained. The obtained steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system.

[0030] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0031] For any given training iteration, steady-state operation data, fault process data, and grid topology data are obtained from a pre-established power system stability dataset. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are then performed on the steady-state features, the fault features, and the topology features to obtain fused latent variable features.

[0032] The fused latent variable features are input into the feature extraction layer of the stability assessment network of the previous iteration process to obtain power system features; the power system features are then input into the stability discrimination layer and margin prediction layer of the stability assessment network of the previous iteration process to obtain stability results and margin results.

[0033] The model obtains the true stability and margin values ​​from a pre-established power system stability dataset. Based on the stability results, the true stability values, the margin results, and the true margin values, it determines the stability loss and margin loss for the current iteration. Based on the Gaussian distribution of stability weights and margin weights from the previous iteration, it determines the stability weights and margin weights for the current iteration through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.

[0034] Based on the total loss of the model, it is determined whether the preset conditions have been met. If not, the stability evaluation network of the previous iteration process, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and the next iteration training is performed. If the preset conditions are met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are obtained. The obtained steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system.

[0035] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0036] For any given training iteration, steady-state operation data, fault process data, and grid topology data are obtained from a pre-established power system stability dataset. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are then performed on the steady-state features, the fault features, and the topology features to obtain fused latent variable features.

[0037] The fused latent variable features are input into the feature extraction layer of the stability assessment network of the previous iteration process to obtain power system features; the power system features are then input into the stability discrimination layer and margin prediction layer of the stability assessment network of the previous iteration process to obtain stability results and margin results.

[0038] The model obtains the true stability and margin values ​​from a pre-established power system stability dataset. Based on the stability results, the true stability values, the margin results, and the true margin values, it determines the stability loss and margin loss for the current iteration. Based on the Gaussian distribution of stability weights and margin weights from the previous iteration, it determines the stability weights and margin weights for the current iteration through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.

[0039] Based on the total loss of the model, it is determined whether the preset conditions have been met. If not, the stability evaluation network of the previous iteration process, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and the next iteration training is performed. If the preset conditions are met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are obtained. The obtained steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system.

[0040] The aforementioned power system stability assessment method, apparatus, computer equipment, computer-readable storage medium, and computer program product, for any given iteration of training, acquire steady-state operating data, fault process data, and grid topology data from a pre-established power system stability dataset. Feature extraction is performed on the steady-state operating data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are then performed on the steady-state features, fault features, and topology features respectively to obtain fused latent variable features. These fused latent variable features are input into the feature extraction layer of the stability assessment network from the previous iteration to obtain power system features. Finally, these power system features are input into the stability discrimination layer and margin prediction layer of the stability assessment network from the previous iteration to obtain... The model obtains stability and margin results; acquires true stability and margin values ​​from a pre-established power system stability dataset; determines the stability loss and margin loss for the current iteration based on the stability results, true stability values, margin results, and true margin values; determines the stability weights and margin weights for the current iteration through reparameterized sampling based on the Gaussian distribution of stability weights and margin weights from the previous iteration; determines the total model loss based on the stability loss, margin loss, KL divergence regularization term, stability weights, and margin weights; by determining the stability weights and margin weights, the model can dynamically balance the training priorities of the two tasks while co-optimizing the stability prediction task and the margin prediction task, enabling the model to simultaneously possess accurate classification and continuous value prediction capabilities. Based on the total loss of the model, it is determined whether a preset condition has been met. If not, the stability evaluation network of the previous iteration, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration, as well as the Gaussian distributions of the stability weights and margin weights of the current iteration, and training is performed for the next iteration. If the preset condition has been met, the stability evaluation network of the current iteration is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, steady-state operation data, fault process data, and grid topology data of the power system are acquired. The acquired steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system. Through the collaborative optimization dual-task modeling and dynamic weight mechanism of this application, the evaluation accuracy and robustness of the model in heterogeneous data scenarios can be improved while mining the correlation features between stability and margin, achieving efficient and reliable online judgment of power system stability. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating a power system stability assessment method in one embodiment;

[0043] Figure 2 A detailed flowchart of a power system stability assessment method in one embodiment;

[0044] Figure 3 This is a structural block diagram of a power system stability assessment device in one embodiment;

[0045] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] In one embodiment, such as Figure 1 As shown, a method for evaluating the stability of a power system 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:

[0048] Step 102: For any iteration of training, obtain steady-state operation data, fault process data, and grid topology data from the pre-established power system stability dataset. Extract features from the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features. Perform linear transformation, latent variable sampling, low-rank interaction, and fusion operations on the steady-state features, the fault features, and the topology features respectively to obtain fused latent variable features.

[0049] Optionally, for the steady-state operation data of the power system, a multi-layer perceptron (MLP) is used to achieve nonlinear mapping and feature compression, mapping the high-dimensional static input into a low-dimensional feature representation, thereby effectively extracting the global features (steady-state features) of the system's steady-state operation.

[0050]

[0051] In the formula, The static feature input vector includes steady-state operating condition data such as generator power and bus voltage. This is a high-dimensional representation vector of the static features after nonlinear mapping by an MLP, with dimensions of... For time-series data reflecting the transient response of power system faults (fault process data), a Transformer-based encoder structure is used for deep abstract feature extraction (only the encoder part is used, without a decoder structure, because this patent focuses on extracting representative time-series features of steady states from the input sequence, not on sequence-to-sequence generation). A multi-head self-attention mechanism is used to model the dependencies between time steps, capturing the dynamic evolution of state variables before and after the fault, thus achieving efficient time-series feature representation (fault features).

[0052]

[0053] In the formula, The multivariate time-series data input matrix includes time-series data such as generator power angle, bus voltage amplitude, and phase angle before and after the fault; This is a high-dimensional representation vector of the multivariate temporal features after learning and encoding using the Transformer encoder structure, with dimensions of... For power grid topology data (including node attributes and adjacency matrices) describing the structural relationships of a power system, a graph attention network (GAT) is used for feature learning. By introducing an attention mechanism, the importance of each neighboring node to the central node is calculated, the electrical correlation between nodes is modeled, the influence of topology on transient stability (topology features) is extracted, and global topology information and local node states are fused.

[0054]

[0055] In the formula, This is the adjacency matrix of the power grid topology, representing the connection relationships between nodes (buses); This is a node feature matrix, containing information such as voltage and phase angle for each node; The topological graph domain feature encoding representation has a dimension of . .

[0056] Step 104: Input the fused latent variable features into the feature extraction layer of the stability assessment network of the previous iteration process to obtain power system features; input the power system features into the stability discrimination layer and margin prediction layer of the stability assessment network of the previous iteration process respectively to obtain stability results and margin results.

[0057] The stability discrimination layer outputs a judgment result on whether the system is stable, which is a classification task; the margin prediction layer outputs a margin prediction, which is a regression task.

[0058] Step 106: Obtain the true stability value and true margin value from the pre-established power system stability dataset. Based on the stability result, the true stability value, the margin result, and the true margin value, determine the stability loss and margin loss of the current iteration process. Based on the Gaussian distribution of stability weights and Gaussian distribution of margin weights of the previous iteration process, determine the stability weight and margin weight of the current iteration process through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weight, and the margin weight, determine the total model loss.

[0059] Optionally, assign weights to each task. The task weights are treated as a learnable latent variable, and their distribution is modeled using a variational inference framework. During training, the task weights are dynamically adjusted based on the difficulty of each task and the uncertainty of the data. Specifically, the weights of each task k follow a Gaussian distribution. Its posterior distribution is obtained by variational inference, and the formula is as follows:

[0060]

[0061] In the formula, The average weight of task k. Let V be the variance of the weights of task k, both of which are obtained through linear transformation during the network forward propagation process; It uses a Gaussian distribution of task weights to capture the uncertainty differences between tasks. Through reparameterization techniques, the weights of classification tasks (stable discriminative tasks) can be... Weights of regression task (margin prediction task) They are sampled from their respective variational distributions.

[0062] The loss function design includes three terms: classification loss (stability loss) for stable discrimination tasks, regression loss (margin loss) for stability margin prediction, and KL divergence regularization term for shared latent variables. Cross-entropy loss is used to measure the stability loss, and the function is:

[0063]

[0064] In the formula, Let be the stable label of the i-th sample; This represents the stability prediction probability of the model output; N is the sample size. The margin loss is calculated using the mean squared error loss function, and its formula is:

[0065]

[0066] In the formula, , These are the model's predicted stability margin and the actual stability margin for the i-th sample, respectively.

[0067] To constrain the posterior distribution of the fused latent variables Without deviating from the prior distribution (Usually assumed to be a standard Gaussian distribution), this invention introduces a KL divergence regularization term to constrain the distribution of latent variables. This regularization term can suppress model overfitting and ensure that the feature fusion results have better generalization and stability under complex power conditions. Its mathematical expression is derived from the analytical expression of the KL divergence of two Gaussian distributions. The simplified result after derivation is given below:

[0068]

[0069] In the formula, and All of these are distribution parameters derived from the fusion of latent variables. and For prior distribution parameters (usually) =0, =1). Combining the stability discrimination task, margin prediction task, and KL divergence regularization term, the multi-task loss function of this invention is defined as:

[0070]

[0071] In the formula, This is the regularization coefficient, used to balance the task loss and the KL divergence regularization term. During model training, the task weights... and It will be achieved by minimizing the loss function We will optimize this process by updating it through backpropagation.

[0072] Step 108: Based on the total loss of the model, determine whether the preset conditions are met. If not, optimize the stability evaluation network of the previous iteration process and the Gaussian distribution parameters of the stability weights and margin weights based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and perform the next iteration training. If the preset conditions are met, use the stability evaluation network of the current iteration process as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, obtain the steady-state operation data, fault process data, and grid topology data of the power system, and input the obtained steady-state operation data, fault process data, and grid topology data into the trained stability evaluation network to obtain the stability result of the power system.

[0073] Optionally, the preset condition can be to determine whether the total loss of the current iteration is less than the total loss of the previous iteration. If it is not less, the preset condition is determined to be met; otherwise, the preset condition is not met.

[0074] The aforementioned power system stability assessment method, for any given iteration of training, acquires steady-state operation data, fault process data, and grid topology data from a pre-established power system stability dataset. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are then performed on the steady-state features, fault features, and topology features respectively to obtain fused latent variable features. These fused latent variable features are input into the feature extraction layer of the stability assessment network from the previous iteration to obtain power system features. These power system features are then input into the stability discrimination layer and margin prediction layer of the stability assessment network from the previous iteration to obtain stability results and margin results. From the pre-established power system stability dataset... First, the true stability and margin values ​​are centrally obtained from the established power system stability dataset. Based on the stability results, the true stability values, the margin results, and the true margin values, the stability loss and margin loss for the current iteration are determined. Based on the Gaussian distribution of stability weights and margin weights from the previous iteration, the stability weights and margin weights for the current iteration are determined through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, the total model loss is determined. By determining the stability weights and margin weights, the training priorities of the two tasks can be dynamically balanced while co-optimizing the stability prediction task and the margin prediction task, allowing the model to simultaneously possess accurate classification and continuous value prediction capabilities. Based on the total loss of the model, it is determined whether a preset condition has been met. If not, the stability evaluation network of the previous iteration, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration, as well as the Gaussian distributions of the stability weights and margin weights of the current iteration, and training is performed for the next iteration. If the preset condition has been met, the stability evaluation network of the current iteration is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, steady-state operation data, fault process data, and grid topology data of the power system are acquired. The acquired steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system. Through the collaborative optimization dual-task modeling and dynamic weight mechanism of this application, the evaluation accuracy and robustness of the model in heterogeneous data scenarios can be improved while mining the correlation features between stability and margin, achieving efficient and reliable online judgment of power system stability.

[0075] In an exemplary embodiment, the step of performing linear transformation, latent variable sampling, low-rank interaction, and fusion operations on the steady-state features, the fault features, and the topological features to obtain fused latent variable features includes:

[0076] Linear transformations and latent variable sampling are performed on the steady-state features, fault features, and topological features respectively to obtain steady-state latent variables, fault latent variables, and topological latent variables; two-modal low-rank interactions and three-modal low-rank interactions are performed on the steady-state latent variables, fault latent variables, and topological latent variables to obtain multiple interaction results; the steady-state features, fault features, topological features, and multiple interaction results are input into a fusion neural network to obtain fused latent variable features.

[0077] Optionally, a fusion neural network is used to fuse multiple features and can be any neural network.

[0078] For example, linear transformations and latent variable sampling are performed on the steady-state features, fault features, and topological features respectively to obtain steady-state latent variables, fault latent variables, and topological latent variables. After obtaining the latent variables of each mode, a low-rank interaction method is used to model the high-order feature interactions to obtain multiple interaction results. Then, a fusion neural network is used to assign attention weights to the steady-state features, fault features, topological features, and multiple interaction results, and outputs a fusion latent variable feature representation. .

[0079] In this embodiment, a variational inference framework is used to perform probabilistic modeling and latent variable sampling of heterogeneous features. By combining low-rank interactions to capture multimodal feature associations and dynamically weighting them through a fusion network, steady-state, fault, and topology information can be efficiently fused while quantifying feature uncertainty. This improves the model's learning ability and generalization performance for key features under complex power conditions.

[0080] In an exemplary embodiment, the step of performing linear transformations and latent variable sampling on the steady-state features, the fault features, and the topological features to obtain steady-state latent variables, fault latent variables, and topological latent variables includes:

[0081] The steady-state features, fault features, and topological features are respectively input into a preset neural network to obtain steady-state mean, steady-state variance, fault mean, fault variance, topological mean, and topological variance. Latent variables are sampled from the steady-state mean and steady-state variance to obtain steady-state latent variables. Latent variables are sampled from the fault mean and fault variance to obtain fault latent variables. Latent variables are sampled from the topological mean and topological variance to obtain topological latent variables.

[0082] Optionally, the default neural network can be any neural network.

[0083] For example, the encoding representation of each modality (in These represent the probability distribution spaces of latent variables mapped to static, topological, and temporal modes, respectively. To address this, a variational inference framework is introduced, employing learnable parameterized linear mappings to generate the mean and variance of the latent variable distributions. This enables probabilistic modeling of feature uncertainties, thus adapting to data uncertainties in power systems caused by factors such as renewable energy fluctuations and fault disturbances. First, the encoding representation of heterogeneous data is addressed. A linear transformation is performed during the forward propagation of the network to obtain the mean vectors of the latent variable distributions. Sum of logarithmic variance vector Its mathematical representation is as follows:

[0084]

[0085] In the formula, , These are the weight matrices that map the encoded representation to the latent variable distribution parameters. , This represents the corresponding bias term. Based on the above distribution parameters, the posterior distribution of the latent variables of heterogeneous data is defined as a multidimensional Gaussian distribution:

[0086]

[0087] Where I is the identity matrix, used to assume that the dimensions of the latent variables are independent of each other. Probabilistic modeling can reveal the uncertainty of quantified features, thus maintaining the model's stability and generalization ability under complex power operating conditions. To ensure the differentiability of the model, this invention employs a reparameterization technique to sample the posterior distribution of the latent variables in the equation, thereby obtaining the latent variables. :

[0088]

[0089] In the formula, Let be a noise vector sampled from a standard normal distribution, and ⊙ denote an element-wise multiplication operation. This reparameterization method transfers randomness to noise. In, thus making latent variables Relative to model parameters and Maintaining differentiability facilitates end-to-end training using gradient descent.

[0090] In this embodiment, a preset neural network is used to perform linear transformation on each modal feature to generate the mean and variance of the latent variable distribution. Then, variational inference and reparameterized sampling are combined to obtain the latent variables of each modality. This not only quantifies the uncertainty of the heterogeneous features (steady state, fault, topology) of the power system to adapt to complex operating conditions, but also ensures the differentiability of the latent variables with respect to the model parameters, laying a reliable foundation for subsequent feature interaction fusion and end-to-end model training.

[0091] In an exemplary embodiment, the two-modal low-rank interaction and three-modal low-rank interaction are performed on the steady-state latent variable, the fault latent variable, and the topological latent variable to obtain multiple interaction results, including:

[0092] A first interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the fault latent variable; a second interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the topological latent variable; a third interaction result is obtained by performing a two-modal low-rank interaction between the fault latent variable and the topological latent variable; and a fourth interaction result is obtained by performing a three-modal low-rank interaction between the steady-state latent variable, the fault latent variable, and the topological latent variable. The first, second, third, and fourth interaction results together constitute multiple interaction results.

[0093] After obtaining the latent variables of each mode Subsequently, this invention employs a low-rank interaction method to model high-order feature interactions. Low-rank decomposition effectively reduces the dimensionality of interaction representations while preserving key interaction information. Specifically, the low-rank interactions of two modalities in heterogeneous data are calculated using the following formulas:

[0094]

[0095]

[0096]

[0097] For the interaction characteristics of the three modes:

[0098]

[0099] In the formula, R is the rank of the low-rank decomposition, which controls the interaction complexity; , , Let be the learnable projection vector corresponding to the r-th rank, which acts on the latent variables of static, temporal, and topological modes, respectively. These low-rank interactions effectively capture the second- and third-order relationships between different modes, while greatly reducing the dimensionality of feature interactions and avoiding the dimensionality explosion problem that may occur in tensor fusion.

[0100] In this embodiment, by performing two-modal and three-modal low-rank interactions on steady-state, fault, and topological latent variables, the low-rank decomposition captures the second-order and third-order correlations between multimodal features while controlling dimensionality complexity. This avoids the dimensionality explosion problem of traditional tensor fusion and effectively mines deep interaction information between different features, providing support for the accuracy of subsequent feature fusion.

[0101] In an exemplary embodiment, steady-state operation data, fault process data, and grid topology data of the power system are acquired. These acquired data are then input into a trained stability assessment network to obtain the stability results of the power system, including:

[0102] Steady-state test data, fault test data, and topology test data of the power system are acquired. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state test features, fault test features, and topology test features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are performed on the steady-state features, fault test features, and topology test features respectively to obtain fused test features. The fused test features are input into the trained stability assessment network to obtain stability prediction results and margin prediction results. Based on the stability prediction results and margin prediction results, it is determined whether the power system is stable.

[0103] For example, steady-state test data, fault test data, and topology test data of a power system are acquired. Feature extraction is performed on the steady-state operation data, fault process data, and power grid topology data to obtain steady-state test features, fault test features, and topology test features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are performed on the steady-state features, the fault test features, and the topology test features, respectively, to obtain fused test features. The fused test features are input into the trained stability assessment network to obtain stability prediction results and margin prediction results. Based on the stability prediction results, it is determined whether the power system is stable, and based on the margin prediction results, the redundancy time for the power system to recover from the current fault state to stable operation is determined.

[0104] In this embodiment, the test data is standardized by reusing the feature processing logic (linear transformation, latent variable sampling, low-rank interaction) of the training phase. Then, the stability and margin prediction results of the trained network are combined to ensure the consistency between the testing process and the training process to reduce errors. At the same time, the dual output of "classification (whether it is stable) + numerical prediction (margin)" provides a more comprehensive and reliable basis for judging the stability of the power system.

[0105] In one exemplary embodiment, the method further includes:

[0106] If the power system is in a stable state, monitoring continues; if the power system is in an unstable state, an alarm is issued.

[0107] For example, when the stability assessment network outputs a stability prediction result of "stable" and a margin prediction result > 0, it is determined that the power system is currently in a stable state. At this time, the system does not need to intervene and continues to collect subsequent steady-state operation data and fault monitoring data in real time to maintain continuous monitoring to track changes in operating conditions. If the network outputs a stability prediction result of "unstable" or a margin prediction result ≤ 0, it is determined that the system has a risk of instability or is already in an unstable state. At this time, the alarm mechanism is immediately triggered, and an audible and visual alarm is issued through the alarm platform of the power dispatch center. At the same time, specific instability risk information is pushed to the dispatcher to assist the dispatcher in taking emergency control measures such as load shedding and adjusting generator output in a timely manner.

[0108] In this embodiment, a dual criterion is constructed by combining stability prediction results and margin prediction results (a stable state requires both predicted stability and margin > 0, while an unstable state corresponds to "predicted instability" or "margin ≤ 0"), thereby achieving accurate determination of the power system's operating status. At the same time, a response mechanism of "continuous monitoring when stable and immediate alarm when unstable" is established, which not only ensures real-time tracking during normal system operation but also quickly triggers intervention procedures when there is a risk of instability, providing timely and reliable decision support for power dispatch and enhancing the power system's security and defense capabilities.

[0109] In one embodiment, such as Figure 2As shown, a method for evaluating the stability of a power system is provided, comprising: for any given training iteration, acquiring steady-state operating data, fault process data, and grid topology data from a pre-established power system stability dataset; extracting features from the steady-state operating data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features; inputting the steady-state features, fault features, and topology features into a preset neural network to obtain steady-state mean, steady-state variance, fault mean, fault variance, topology mean, and topology variance; sampling latent variables from the steady-state mean and steady-state variance to obtain steady-state latent variables; sampling latent variables from the fault mean and fault variance to obtain fault latent variables; and sampling latent variables from the topology mean and topology variance to obtain topology latent variables. A first interaction result is obtained by performing a two-modal low-rank interaction on the steady-state latent variable and the fault latent variable; a second interaction result is obtained by performing a two-modal low-rank interaction on the steady-state latent variable and the topological latent variable; a third interaction result is obtained by performing a two-modal low-rank interaction on the fault latent variable and the topological latent variable; and a fourth interaction result is obtained by performing a three-modal low-rank interaction on the steady-state latent variable, the fault latent variable, and the topological latent variable. The first, second, third, and fourth interaction results together constitute multiple interaction results. The steady-state features, the fault features, the topological features, and the multiple interaction results are input into a fusion neural network to obtain fused latent variable features. The fused latent variable features are input into the feature extraction layer of the stability assessment network of the previous iteration process to obtain power system features; the power system features are then input into the stability discrimination layer and the margin prediction layer of the stability assessment network of the previous iteration process to obtain stability results and margin results. The model obtains the true stability value and true margin value from a pre-established power system stability dataset. Based on the stability result, the true stability value, the margin result, and the true margin value, it determines the stability loss and margin loss of the current iteration process. Based on the Gaussian distribution of stability weights and margin weights of the previous iteration process, it determines the stability weights and margin weights of the current iteration process through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.Based on the total loss of the model, it is determined whether the preset conditions have been met. If not, the stability evaluation network of the previous iteration process, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and the next iteration training is performed. If the preset conditions are met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are obtained. The obtained steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system. Acquiring steady-state operation data, fault process data, and grid topology data of a power system, and inputting these data into a trained stability assessment network to obtain the stability results of the power system, includes: acquiring steady-state test data, fault test data, and topology test data of the power system; extracting features from the steady-state operation data, fault process data, and grid topology data to obtain steady-state test features, fault test features, and topology test features; performing linear transformation, latent variable sampling, low-rank interaction, and fusion operations on the steady-state features, fault test features, and topology test features respectively to obtain fused test features; inputting the fused test features into the trained stability assessment network to obtain stability prediction results and margin prediction results; and determining whether the power system is stable based on the stability prediction results and margin prediction results. When the stability assessment network outputs a stability prediction result of "stable" and a margin prediction result > 0 after training, the power system is determined to be in a stable state. At this time, the system does not need to be intervened and continues to collect subsequent steady-state operation data and fault monitoring data in real time to keep track of changes in operating conditions. If the network outputs a stability prediction result of "unstable" or a margin prediction result ≤ 0, the system is determined to have a risk of instability or is already in an unstable state. At this time, the alarm mechanism is immediately triggered, and an audible and visual alarm is issued through the alarm platform of the power dispatch center. At the same time, specific instability risk information is pushed to the dispatcher to assist the dispatcher in taking emergency control measures such as load shedding and adjusting generator output in a timely manner.

[0110] 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 in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0111] In one exemplary embodiment, such as Figure 3 As shown, a power system stability assessment device is provided, comprising: an acquisition module 301, a first extraction module 302, a second extraction module 303, and a judgment module 304, wherein:

[0112] The acquisition module is used to acquire steady-state operation data, fault process data, and grid topology data from a pre-established power system stability dataset for any given training iteration; extract features from the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features; and perform linear transformation, latent variable sampling, low-rank interaction, and fusion operations on the steady-state features, fault features, and topology features respectively to obtain fused latent variable features.

[0113] The first extraction module is used to input the fused latent variable features into the feature extraction layer of the stability evaluation network of the previous iteration process to obtain power system features; and to input the power system features into the stability discrimination layer and margin prediction layer of the stability evaluation network of the previous iteration process to obtain stability results and margin results.

[0114] The second extraction module is used to obtain the true stability value and true margin value from a pre-established power system stability dataset; based on the stability result, the true stability value, the margin result, and the true margin value, it determines the stability loss and margin loss of the current iteration process; based on the Gaussian distribution of stability weights and margin weights of the previous iteration process, it determines the stability weights and margin weights of the current iteration process through reparameterized sampling; based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.

[0115] The judgment module is used to determine whether a preset condition has been met based on the total loss of the model. If not, it optimizes the stability evaluation network of the previous iteration process and the Gaussian distribution parameters of the stability weights and margin weights based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and performs the next iteration training. If the preset condition has been met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are acquired. The acquired steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system.

[0116] In one embodiment, the first extraction module is further configured to:

[0117] Linear transformations and latent variable sampling are performed on the steady-state features, fault features, and topological features respectively to obtain steady-state latent variables, fault latent variables, and topological latent variables; two-modal low-rank interactions and three-modal low-rank interactions are performed on the steady-state latent variables, fault latent variables, and topological latent variables to obtain multiple interaction results; the steady-state features, fault features, topological features, and multiple interaction results are input into a fusion neural network to obtain fused latent variable features.

[0118] In one embodiment, the first extraction module is further configured to:

[0119] The steady-state features, fault features, and topological features are respectively input into a preset neural network to obtain steady-state mean, steady-state variance, fault mean, fault variance, topological mean, and topological variance. Latent variables are sampled from the steady-state mean and steady-state variance to obtain steady-state latent variables. Latent variables are sampled from the fault mean and fault variance to obtain fault latent variables. Latent variables are sampled from the topological mean and topological variance to obtain topological latent variables.

[0120] In one embodiment, the first extraction module is further configured to:

[0121] A first interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the fault latent variable; a second interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the topological latent variable; a third interaction result is obtained by performing a two-modal low-rank interaction between the fault latent variable and the topological latent variable; and a fourth interaction result is obtained by performing a three-modal low-rank interaction between the steady-state latent variable, the fault latent variable, and the topological latent variable. The first, second, third, and fourth interaction results together constitute multiple interaction results.

[0122] In one embodiment, the determining module is further configured to:

[0123] Steady-state test data, fault test data, and topology test data of the power system are acquired. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state test features, fault test features, and topology test features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are performed on the steady-state features, fault test features, and topology test features respectively to obtain fused test features. The fused test features are input into the trained stability assessment network to obtain stability prediction results and margin prediction results. Based on the stability prediction results and margin prediction results, it is determined whether the power system is stable.

[0124] In one embodiment, the determining module is further configured to:

[0125] If the power system is in a stable state, monitoring continues; if the power system is in an unstable state, an alarm is issued.

[0126] Each module in the aforementioned power system stability assessment device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0127] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores a power system stability dataset. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a power system stability assessment method.

[0128] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0129] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0130] For any given training iteration, steady-state operation data, fault process data, and grid topology data are obtained from a pre-established power system stability dataset. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are then performed on the steady-state features, the fault features, and the topology features to obtain fused latent variable features.

[0131] The fused latent variable features are input into the feature extraction layer of the stability assessment network of the previous iteration process to obtain power system features; the power system features are then input into the stability discrimination layer and margin prediction layer of the stability assessment network of the previous iteration process to obtain stability results and margin results.

[0132] The model obtains the true stability and margin values ​​from a pre-established power system stability dataset. Based on the stability results, the true stability values, the margin results, and the true margin values, it determines the stability loss and margin loss for the current iteration. Based on the Gaussian distribution of stability weights and margin weights from the previous iteration, it determines the stability weights and margin weights for the current iteration through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.

[0133] Based on the total loss of the model, it is determined whether the preset conditions have been met. If not, the stability evaluation network of the previous iteration process, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and the next iteration training is performed. If the preset conditions are met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are obtained. The obtained steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system.

[0134] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0135] Linear transformations and latent variable sampling are performed on the steady-state features, fault features, and topological features respectively to obtain steady-state latent variables, fault latent variables, and topological latent variables; two-modal low-rank interactions and three-modal low-rank interactions are performed on the steady-state latent variables, fault latent variables, and topological latent variables to obtain multiple interaction results; the steady-state features, fault features, topological features, and multiple interaction results are input into a fusion neural network to obtain fused latent variable features.

[0136] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0137] The steady-state features, fault features, and topological features are respectively input into a preset neural network to obtain steady-state mean, steady-state variance, fault mean, fault variance, topological mean, and topological variance. Latent variables are sampled from the steady-state mean and steady-state variance to obtain steady-state latent variables. Latent variables are sampled from the fault mean and fault variance to obtain fault latent variables. Latent variables are sampled from the topological mean and topological variance to obtain topological latent variables.

[0138] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0139] A first interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the fault latent variable; a second interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the topological latent variable; a third interaction result is obtained by performing a two-modal low-rank interaction between the fault latent variable and the topological latent variable; and a fourth interaction result is obtained by performing a three-modal low-rank interaction between the steady-state latent variable, the fault latent variable, and the topological latent variable. The first, second, third, and fourth interaction results together constitute multiple interaction results.

[0140] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0141] Steady-state test data, fault test data, and topology test data of the power system are acquired. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state test features, fault test features, and topology test features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are performed on the steady-state features, fault test features, and topology test features respectively to obtain fused test features. The fused test features are input into the trained stability assessment network to obtain stability prediction results and margin prediction results. Based on the stability prediction results and margin prediction results, it is determined whether the power system is stable.

[0142] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0143] If the power system is in a stable state, monitoring continues; if the power system is in an unstable state, an alarm is issued.

[0144] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0145] For any given training iteration, steady-state operation data, fault process data, and grid topology data are obtained from a pre-established power system stability dataset. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are then performed on the steady-state features, the fault features, and the topology features to obtain fused latent variable features.

[0146] The fused latent variable features are input into the feature extraction layer of the stability assessment network of the previous iteration process to obtain power system features; the power system features are then input into the stability discrimination layer and margin prediction layer of the stability assessment network of the previous iteration process to obtain stability results and margin results.

[0147] The model obtains the true stability and margin values ​​from a pre-established power system stability dataset. Based on the stability results, the true stability values, the margin results, and the true margin values, it determines the stability loss and margin loss for the current iteration. Based on the Gaussian distribution of stability weights and margin weights from the previous iteration, it determines the stability weights and margin weights for the current iteration through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.

[0148] Based on the total loss of the model, it is determined whether the preset conditions have been met. If not, the stability evaluation network of the previous iteration process, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and the next iteration training is performed. If the preset conditions are met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are obtained. The obtained steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system.

[0149] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0150] Linear transformations and latent variable sampling are performed on the steady-state features, fault features, and topological features respectively to obtain steady-state latent variables, fault latent variables, and topological latent variables; two-modal low-rank interactions and three-modal low-rank interactions are performed on the steady-state latent variables, fault latent variables, and topological latent variables to obtain multiple interaction results; the steady-state features, fault features, topological features, and multiple interaction results are input into a fusion neural network to obtain fused latent variable features.

[0151] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0152] The steady-state features, fault features, and topological features are respectively input into a preset neural network to obtain steady-state mean, steady-state variance, fault mean, fault variance, topological mean, and topological variance. Latent variables are sampled from the steady-state mean and steady-state variance to obtain steady-state latent variables. Latent variables are sampled from the fault mean and fault variance to obtain fault latent variables. Latent variables are sampled from the topological mean and topological variance to obtain topological latent variables.

[0153] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0154] A first interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the fault latent variable; a second interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the topological latent variable; a third interaction result is obtained by performing a two-modal low-rank interaction between the fault latent variable and the topological latent variable; and a fourth interaction result is obtained by performing a three-modal low-rank interaction between the steady-state latent variable, the fault latent variable, and the topological latent variable. The first, second, third, and fourth interaction results together constitute multiple interaction results.

[0155] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0156] Steady-state test data, fault test data, and topology test data of the power system are acquired. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state test features, fault test features, and topology test features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are performed on the steady-state features, fault test features, and topology test features respectively to obtain fused test features. The fused test features are input into the trained stability assessment network to obtain stability prediction results and margin prediction results. Based on the stability prediction results and margin prediction results, it is determined whether the power system is stable.

[0157] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0158] If the power system is in a stable state, monitoring continues; if the power system is in an unstable state, an alarm is issued.

[0159] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0160] For any given training iteration, steady-state operation data, fault process data, and grid topology data are obtained from a pre-established power system stability dataset. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state features, fault features, and topology features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are then performed on the steady-state features, the fault features, and the topology features to obtain fused latent variable features.

[0161] The fused latent variable features are input into the feature extraction layer of the stability assessment network of the previous iteration process to obtain power system features; the power system features are then input into the stability discrimination layer and margin prediction layer of the stability assessment network of the previous iteration process to obtain stability results and margin results.

[0162] The model obtains the true stability and margin values ​​from a pre-established power system stability dataset. Based on the stability results, the true stability values, the margin results, and the true margin values, it determines the stability loss and margin loss for the current iteration. Based on the Gaussian distribution of stability weights and margin weights from the previous iteration, it determines the stability weights and margin weights for the current iteration through reparameterized sampling. Based on the stability loss, the margin loss, the KL divergence regularization term, the stability weights, and the margin weights, it determines the total model loss.

[0163] Based on the total loss of the model, it is determined whether the preset conditions have been met. If not, the stability evaluation network of the previous iteration process, as well as the Gaussian distribution parameters of the stability weights and margin weights, are optimized based on the total loss to obtain the stability evaluation network of the current iteration process, as well as the Gaussian distribution of the stability weights and the Gaussian distribution of the margin weights of the current iteration process, and the next iteration training is performed. If the preset conditions are met, the stability evaluation network of the current iteration process is used as the trained stability evaluation network. When it is necessary to evaluate the stability of the power system, the steady-state operation data, fault process data, and grid topology data of the power system are obtained. The obtained steady-state operation data, fault process data, and grid topology data are input into the trained stability evaluation network to obtain the stability result of the power system.

[0164] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0165] Linear transformations and latent variable sampling are performed on the steady-state features, fault features, and topological features respectively to obtain steady-state latent variables, fault latent variables, and topological latent variables; two-modal low-rank interactions and three-modal low-rank interactions are performed on the steady-state latent variables, fault latent variables, and topological latent variables to obtain multiple interaction results; the steady-state features, fault features, topological features, and multiple interaction results are input into a fusion neural network to obtain fused latent variable features.

[0166] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0167] The steady-state features, fault features, and topological features are respectively input into a preset neural network to obtain steady-state mean, steady-state variance, fault mean, fault variance, topological mean, and topological variance. Latent variables are sampled from the steady-state mean and steady-state variance to obtain steady-state latent variables. Latent variables are sampled from the fault mean and fault variance to obtain fault latent variables. Latent variables are sampled from the topological mean and topological variance to obtain topological latent variables.

[0168] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0169] A first interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the fault latent variable; a second interaction result is obtained by performing a two-modal low-rank interaction between the steady-state latent variable and the topological latent variable; a third interaction result is obtained by performing a two-modal low-rank interaction between the fault latent variable and the topological latent variable; and a fourth interaction result is obtained by performing a three-modal low-rank interaction between the steady-state latent variable, the fault latent variable, and the topological latent variable. The first, second, third, and fourth interaction results together constitute multiple interaction results.

[0170] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0171] Steady-state test data, fault test data, and topology test data of the power system are acquired. Feature extraction is performed on the steady-state operation data, fault process data, and grid topology data to obtain steady-state test features, fault test features, and topology test features. Linear transformation, latent variable sampling, low-rank interaction, and fusion operations are performed on the steady-state features, fault test features, and topology test features respectively to obtain fused test features. The fused test features are input into the trained stability assessment network to obtain stability prediction results and margin prediction results. Based on the stability prediction results and margin prediction results, it is determined whether the power system is stable.

[0172] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0173] If the power system is in a stable state, monitoring continues; if the power system is in an unstable state, an alarm is issued.

[0174] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0175] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0176] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0177] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method of stability assessment of an electric power system, characterized by, The method comprises: For each iteration training, steady-state operation data, fault process data and power grid topology data are obtained from a pre-established power system stability dataset, feature extraction is performed on the steady-state operation data, the fault process data and the power grid topology data to obtain steady-state features, fault features and topology features, linear transformation, latent variable sampling, low-rank interaction and fusion operations are performed on the steady-state features, the fault features and the topology features to obtain fused latent variable features; The fused latent variable features are input into a feature extraction layer in a stability evaluation network of the previous iteration process to obtain power system features, and the power system features are input into a stability discrimination layer and a margin prediction layer in the stability evaluation network of the previous iteration process to obtain stability results and margin results; Stability true values and margin true values are obtained from the pre-established power system stability dataset, and based on the stability results, the stability true values, the margin results and the margin true values, stability loss and margin loss of the current iteration process are determined, and based on stability weight Gaussian distribution and margin weight Gaussian distribution of the previous iteration process, stability weight and margin weight of the current iteration process are determined by reparameterization sampling, and based on the stability loss, the margin loss, a KL divergence regularization term, the stability weight and the margin weight, model total loss is determined; Based on the model total loss, it is determined whether a preset condition is reached, if not, the stability evaluation network of the previous iteration process and Gaussian distribution parameters of the stability weight and the margin weight are optimized based on the total loss to obtain the stability evaluation network of the current iteration process, stability weight Gaussian distribution and margin weight Gaussian distribution of the current iteration process, and the next iteration training is performed, if the preset condition is reached, the stability evaluation network of the current iteration process is taken as the trained stability evaluation network, when it is necessary to evaluate the stability of the power system, steady-state operation data, fault process data and power grid topology data of the power system are obtained, and the obtained steady-state operation data, fault process data and power grid topology data are input into the trained stability evaluation network to obtain the stability of the power system.

2. The method of claim 1, wherein, The linear transformation, latent variable sampling, low-rank interaction and fusion operations are performed on the steady-state features, the fault features and the topology features to obtain the fused latent variable features, comprising: The linear transformation and the latent variable sampling are performed on the steady-state features, the fault features and the topology features to obtain steady-state latent variables, fault latent variables and topology latent variables; Two-modal low-rank interaction and three-modal low-rank interaction are performed on the steady-state latent variables, the fault latent variables and the topology latent variables to obtain a plurality of interaction results; The steady-state features, the fault features, the topology features and the plurality of interaction results are input into a fusion neural network to obtain the fused latent variable features.

3. The method of claim 2, wherein, The linear transformation and the latent variable sampling are performed on the steady-state features, the fault features and the topology features to obtain steady-state latent variables, fault latent variables and topology latent variables, comprising: inputting the steady-state feature, the fault feature and the topology feature into a preset neural network respectively to obtain a steady-state mean value, a steady-state variance, a fault mean value, a fault variance, a topology mean value and a topology variance; sampling latent variables of the steady-state mean value and the steady-state variance to obtain a steady-state latent variable; sampling latent variables of the fault mean value and the fault variance to obtain a fault latent variable; and sampling latent variables of the topology mean value and the topology variance to obtain a topology latent variable.

4. The method of claim 2, wherein, the steady-state latent variable, the fault latent variable and the topology latent variable to obtain a plurality of interaction results, including: two-mode low-rank interaction of the steady-state latent variable and the fault latent variable to obtain a first interaction result; two-mode low-rank interaction of the steady-state latent variable and the topology latent variable to obtain a second interaction result; two-mode low-rank interaction of the fault latent variable and the topology latent variable to obtain a third interaction result; three-mode low-rank interaction of the steady-state latent variable, the fault latent variable and the topology latent variable to obtain a fourth interaction result; and the first interaction result, the second interaction result, the third interaction result and the fourth interaction result together constitute the plurality of interaction results.

5. The method of claim 1, wherein, obtaining steady-state operation data, fault process data and power grid topology data of a power system, inputting the obtained steady-state operation data, fault process data and power grid topology data into a trained stability evaluation network to obtain a stability result of the power system, including: obtaining steady-state test data, fault test data and topology test data of the power system, extracting features from the steady-state operation data, the fault process data and the power grid topology data to obtain steady-state test features, fault test features and topology test features; performing linear transformation, latent variable sampling, low-rank interaction and fusion operation on the steady-state features, the fault test features and the topology test features respectively to obtain fused test features; inputting the fused test features into the trained stability evaluation network to obtain a stability prediction result and a margin prediction result; and determining whether the power system is stable based on the stability prediction result and the margin prediction result.

6. The method of claim 5, wherein, The method further includes: if the power system is in a stable state, continuing to detect; if the power system is in an unstable state, issuing an alarm.

7. A stability assessment device of a power system characterized by comprising: The device includes: an obtaining module configured to, for each iteration training, obtain steady-state operation data, fault process data and power grid topology data from a pre-established stability data set of the power system, extract features from the steady-state operation data, the fault process data and the power grid topology data to obtain steady-state features, fault features and topology features, and perform linear transformation, latent variable sampling, low-rank interaction and fusion operation on the steady-state features, the fault features and the topology features respectively to obtain fused latent variable features; and The first extraction module is configured to input the fusion latent variable feature into a feature extraction layer in a stability evaluation network of a previous iteration process to obtain power system features; and input the power system features into a stability discrimination layer and a margin prediction layer in the stability evaluation network of the previous iteration process to obtain a stability result and a margin result. The second extraction module is configured to obtain stability true values and margin true values from a pre-established stability data set of the power system; determine stability loss and margin loss of a current iteration process based on the stability result, the stability true values, the margin result and the margin true values; determine stability weight and margin weight of the current iteration process by reparameterization sampling based on stability weight Gaussian distribution and margin weight Gaussian distribution of the previous iteration process; and determine a model total loss based on the stability loss, the margin loss, a KL divergence regularization term, the stability weight and the margin weight. The determination module is configured to determine whether a preset condition is met based on the model total loss; if not, optimize the stability evaluation network of the previous iteration process and Gaussian distribution parameters of the stability weight and the margin weight based on the total loss to obtain the stability evaluation network of the current iteration process and stability weight Gaussian distribution and margin weight Gaussian distribution of the current iteration process, and perform next iteration training; if the preset condition is met, take the stability evaluation network of the current iteration process as a trained stability evaluation network, and when it is necessary to evaluate the stability of the power system, input steady-state operation data, fault process data and power grid topology data of the power system into the trained stability evaluation network to obtain a stability result of the power system.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method in any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 6. The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 6.

Images