CN-122021057-A - Method, system, equipment and medium for reconstructing transient simulation track of power system

CN122021057ACN 122021057 ACN122021057 ACN 122021057ACN-122021057-A

Abstract

The invention relates to the technical field of transient simulation, and discloses a method, a system, equipment and a medium for reconstructing transient simulation track of an electric power system, wherein the method inputs current noise data containing current node fault characteristics into a pre-trained DiT network, outputs predictive noise contained in the current noise data through the pre-trained DiT network, carries out inverse denoising treatment on the current noise data according to the predictive noise, the method comprises the steps of obtaining denoised node failure high-dimensional characteristics, performing dimension reduction on the denoised node failure high-dimensional characteristics to obtain reconstructed node failure characteristics, and forming a transient simulation track according to the reconstructed node failure characteristics, so that efficient modeling and rapid simulation track generation of a transient process of a power system are realized by introducing a generation mechanism of a DiT network diffusion model, and the generation efficiency and track precision are improved.

Inventors

Liang Zhuohang
Yin Lulan
MAO ZHIYU
LIAO MENGJUN
LI CHENGXIANG
ZHU YIHUA
CHANG DONGXU
ZHU YUKUN
WU MINGKANG
WANG YULIN
CHEN LEI

Assignees

南方电网科学研究院有限责任公司

Dates

Publication Date: 20260512
Application Date: 20260320

Claims (10)

1. The method for reconstructing the transient simulation track of the power system is characterized by comprising the following steps of: Acquiring current noise data containing current node fault characteristics in an electric power system, inputting the current noise data into a pre-trained DiT network, and outputting predicted noise contained in the current noise data through the pre-trained DiT network, wherein the pre-trained DiT network is obtained by training an initial DiT network by combining a condition characterization matrix output by a condition characterization network based on the noisy node fault high-dimensional characteristics and the predicted noise corresponding to the noisy node fault high-dimensional characteristics as mapping samples; Performing reverse denoising processing on the current noise data according to the predicted noise to obtain denoised node failure high-dimensional characteristics; And decoding and dimension-reducing the denoised node failure high-dimension characteristic to obtain a reconstructed node failure characteristic, and forming a transient simulation track according to the reconstructed node failure characteristic.
2. The method for reconstructing a transient simulation trajectory of a power system according to claim 1, wherein the training process of the pre-trained DiT network comprises: acquiring a denoised node failure high-dimensional characteristic sample and a prediction noise label corresponding to the denoised node failure high-dimensional characteristic sample, and forming a training sample set; Inputting the training sample set into an initial DiT network, carrying out feature extraction on the denoised node fault high-dimensional feature samples according to a condition characterization matrix and a forward prediction time step output by a condition characterization network, optimizing network parameters in the initial DiT network by minimizing a mean square error between prediction noise and real noise until the network parameters converge or the iteration number reaches a preset maximum iteration number, and obtaining a DiT network after training as the pretrained DiT network.
3. The method for reconstructing a transient simulation trajectory of a power system of claim 2, further comprising: Acquiring an original node fault feature matrix of the power system after faults, masking the original node fault feature matrix, and performing high-dimensional representation coding on the original node fault feature matrix after masking to obtain a high-dimensional feature matrix; and carrying out forward diffusion noise adding processing on the high-dimensional feature matrix, combining preset controllable noise to obtain a noise-added node failure high-dimensional feature sample, and marking the noise-added node failure high-dimensional feature sample based on the preset controllable noise to obtain a prediction noise label corresponding to the noise-added node failure high-dimensional feature sample.
4. The method for reconstructing a transient simulation track of a power system according to claim 3, wherein the obtaining the original node fault feature matrix of the power system after the fault, performing mask processing on the original node fault feature matrix, and performing high-dimensional representation encoding on the mask processed original node fault feature matrix to obtain a high-dimensional feature matrix, includes: Masking the original node fault feature matrix based on a preset transparent mask matrix, and respectively carrying out up-dimension mapping on the original node fault feature matrix after masking and the transparent mask matrix to a high-dimension space to obtain a node fault feature matrix after up-dimension and a transparent mask matrix after up-dimension; The node fault feature matrix after dimension increase and the transmission mask matrix after dimension increase are fused and overlapped to obtain an initial high-dimension feature matrix after superposition; And carrying out neighborhood attention calculation and residual error connection treatment on the initial high-dimensional feature matrix after superposition by adopting a stacked graph characterization unit, fusing a neighborhood attention calculation result and a residual error connection result, carrying out layer normalization on the fused result, and carrying out nonlinear transformation on the feature matrix after layer normalization through a feedforward network and an activation function to obtain the high-dimensional feature matrix.
5. The method for reconstructing a transient simulation track of a power system according to claim 2, wherein the initial DiT network comprises a plurality of cascaded DiT modules, a layer normalization module, a linear layer and a dimension reconstruction layer; Inputting the training sample set into an initial DiT network, extracting features of the denoised node fault high-dimensional feature samples according to a condition characterization matrix and a forward prediction time step output by a condition characterization network, optimizing network parameters in the initial DiT network by minimizing a mean square error between prediction noise and real noise until the network parameters converge or the iteration number reaches a preset maximum iteration number, and obtaining a DiT network after training, wherein the 5225 network is used as the pretrained DiT network and comprises the following steps: The noisy node fault high-dimensional feature sample, the condition characterization matrix output by the condition characterization network and the forward prediction time step are input to an initial DiT network in a combined mode, feature extraction is carried out on the input features through each DiT module based on a multi-head self-attention mechanism, and cross attention calculation is carried out on the extracted features by combining the condition characterization matrix output by the condition characterization network and the forward prediction time step, so that fused features are obtained; Inputting the fused features to the layer normalization module for normalization processing, linearly mapping the normalized interaction features through the linear layer, and performing dimension reconstruction on the linearly mapped features through the dimension reconstruction layer to output prediction noise; and performing back propagation optimization on parameters of the initial DiT network by minimizing the mean square error between the predicted noise and the real noise, and completing network training when the loss function converges or the iteration number reaches the preset maximum iteration number, so as to obtain the pre-trained DiT network.
6. The method for reconstructing a transient simulation trajectory of a power system of claim 5, further comprising: Acquiring a plurality of node constraint condition characteristics of the power system after faults, and inputting each node constraint condition characteristic into the condition characterization network to perform lifting and coding to obtain a high-dimensional characterization corresponding to each node constraint condition characteristic; carrying out fusion superposition on each high-dimensional representation to obtain a fused high-dimensional representation; Obtaining the simulation step number in the forward prediction time step, and carrying out ascending and position coding on the simulation step number to obtain time step vectors corresponding to a plurality of time points; Selecting a corresponding time step vector from time step vectors corresponding to a plurality of time points according to the time stamp to be predicted, and taking the time step vector as the time stamp vector to be predicted after dimension rising; and splicing the fused high-dimensional representation and the timestamp vector to be predicted after the dimension rise to obtain a condition representation matrix of the condition representation network output.
7. The method for reconstructing a transient simulation track of a power system according to claim 1, wherein the performing inverse denoising processing on the current noise data according to the predicted noise to obtain denoised node fault high-dimensional features comprises: based on a DDIM denoising diffusion implicit model, gradually denoising the current noise data in combination with the predicted noise, and updating the current noise data in each iteration until all denoising steps are completed, so as to obtain the denoised node fault high-dimensional characteristics.
8. A power system transient simulation trajectory reconstruction system, comprising: The noise prediction module is used for acquiring current noise data containing current node fault characteristics in the power system, inputting the current noise data into a pre-trained DiT network, and outputting predicted noise contained in the current noise data through the pre-trained DiT network, wherein the pre-trained DiT network is obtained by training an initial DiT network by combining a condition characterization matrix output by a condition characterization network based on the noisy node fault high-dimensional characteristics and the predicted noise corresponding to the noisy node fault high-dimensional characteristics as mapping samples; The reverse denoising module is used for carrying out reverse denoising processing on the current noise data according to the predicted noise to obtain denoised node failure high-dimensional characteristics; And the reconstruction module is used for decoding and dimension-reducing the denoised node failure high-dimensional characteristics to obtain reconstructed node failure characteristics, and forming a transient simulation track according to the reconstructed node failure characteristics.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the power system transient simulation trajectory reconstruction method of any one of claims 1-7.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed, implements the steps of the power system transient simulation trajectory reconstruction method according to any one of claims 1-7.

Description

Method, system, equipment and medium for reconstructing transient simulation track of power system Technical Field The invention relates to the technical field of transient simulation, in particular to a method, a system, equipment and a medium for reconstructing a transient simulation track of a power system. Background With the wide access of high-proportion new energy grid connection and power electronic equipment, the dynamic behavior of a power system is more complex, and the rapid evaluation of transient stability becomes a key for guaranteeing the safe operation of a power grid. The traditional transient simulation method for solving the differential-algebraic equation based on numerical integration has high precision, but long calculation time is consumed, and the requirements of on-line analysis and real-time decision making are difficult to meet. At present, research attempts are made to learn system dynamic characteristics from historical simulation data by using a generation type model such as a generation countermeasure network and a variation self-encoder, generate alternative simulation tracks, and realize quick deduction of a system transient stability result based on the generated tracks. For example, using GAN (GENERATIVE ADVERSARIAL Network, generating timing data against the Network) to generate critical electrical quantities, or using VAEs (Variational Autoencoder, variational self-encoders) to model stable/unstable trajectories to augment training samples or build a fast evaluation proxy model. However, these generative models still have problems such as pattern collapse, unstable training, and unsatisfactory generation quality when generating long-time-series and multivariable-coupled power system transient trajectories. In particular, in the case of simulating extreme scenes such as instability, the generated data often lacks physical consistency, and the requirements of high-reliability power system analysis are difficult to meet. Disclosure of Invention In view of the above, the present invention provides a method, a system, a device and a medium for reconstructing transient simulation trajectories of an electric power system in order to solve the above technical problems. The first aspect of the invention provides a method for reconstructing a transient simulation track of an electric power system, which comprises the following steps: Acquiring current noise data containing current node fault characteristics in an electric power system, inputting the current noise data into a pre-trained DiT network, and outputting predicted noise contained in the current noise data through the pre-trained DiT network, wherein the pre-trained DiT network is obtained by training an initial DiT network by combining a condition characterization matrix output by a condition characterization network based on the noisy node fault high-dimensional characteristics and the predicted noise corresponding to the noisy node fault high-dimensional characteristics as mapping samples; Performing reverse denoising processing on the current noise data according to the predicted noise to obtain denoised node failure high-dimensional characteristics; And decoding and dimension-reducing the denoised node failure high-dimension characteristic to obtain a reconstructed node failure characteristic, and forming a transient simulation track according to the reconstructed node failure characteristic. Preferably, the training process of the pre-trained DiT network includes: acquiring a denoised node failure high-dimensional characteristic sample and a prediction noise label corresponding to the denoised node failure high-dimensional characteristic sample, and forming a training sample set; Inputting the training sample set into an initial DiT network, carrying out feature extraction on the denoised node fault high-dimensional feature samples according to a condition characterization matrix and a forward prediction time step output by a condition characterization network, optimizing network parameters in the initial DiT network by minimizing a mean square error between prediction noise and real noise until the network parameters converge or the iteration number reaches a preset maximum iteration number, and obtaining a DiT network after training as the pretrained DiT network. Preferably, the method further comprises: Acquiring an original node fault feature matrix of the power system after faults, masking the original node fault feature matrix, and performing high-dimensional representation coding on the original node fault feature matrix after masking to obtain a high-dimensional feature matrix; and carrying out forward diffusion noise adding processing on the high-dimensional feature matrix, combining preset controllable noise to obtain a noise-added node failure high-dimensional feature sample, and marking the noise-added node failure high-dimensional feature sample based on the preset controllable noise to obtain a predic