Search

CN-122021908-A - Power grid admittance parameter reverse inference method based on physical constraint deep learning

CN122021908ACN 122021908 ACN122021908 ACN 122021908ACN-122021908-A

Abstract

The invention discloses a power grid admittance parameter reverse inference method based on physical constraint deep learning, which comprises the following steps of 1, data acquisition and theoretical screening; step 2, extracting and constructing characteristics; step 3, constructing PhyG-Mamba inference models; the method comprises the steps of (1) physical constraint optimization training, (5) reverse inference, clear data lower bound and realize small sample efficient inference, wherein the minimum data volume required by the maximum inference threat is strictly proved to be L-N+1 (N is the number of nodes and L is the number of lines) from a mathematical level based on the theory of the residual tree (Cotree) in the graph theory, the data lower bound of the reverse inference of the power grid admittance parameters is defined for the first time, the minimum complete data set when the residual tree lines are respectively congested is screened, massive historical data is not needed, the pain point of large data demand in the traditional mathematical optimization method is solved, the data acquisition cost and the storage cost are reduced, and the efficient inference under the small sample scene is realized.

Inventors

  • JIANG HUAWEN
  • CHEN PENGBIN

Assignees

  • 上海辰文豪科技有限公司

Dates

Publication Date
20260512
Application Date
20260202

Claims (6)

  1. 1. A power grid admittance parameter reverse inference method based on physical constraint deep learning comprises the following steps: step 1, data acquisition and theoretical screening Constructing a homogeneous linear equation set based on a direct current power flow model, screening minimum electricity price data capable of enabling the coefficient matrix rank of the equation set to reach the maximum from historical data by utilizing a residual tree Cotree theory in a graph theory, and constructing a training data set; Step2, feature extraction and construction Extracting node characteristics and edge characteristics from the screened data set, wherein the node characteristics comprise congestion component MCC and topology information in LMP, and the edge characteristics comprise line shadow prices; step 3, constructing PhyG-Mamba inference model The method comprises the steps of constructing a deep neural network model comprising an improved encoder and Mamba modules, wherein the improved encoder adopts a mixed head attention MDR mechanism to dynamically route and encode node characteristics, and the Mamba modules process sequence data fused with node coding characteristics and edge characteristics based on a selective state space model to capture global dependency relations; Step 4, physical constraint optimization training Substituting the prediction admittance matrix output by the model into a direct current power flow equation, calculating physical constraint loss, constructing a total loss function by combining the prediction numerical error, and carrying out iterative update on model parameters; Step 5, reverse direction inference And (3) inputting public power data of a period to be tested by using the trained PhyG-Mamba model, and outputting admittance matrix parameters of the target power grid.
  2. 2. The method for reverse inference of power grid admittance parameters based on physical constraint deep learning of claim 1, wherein the homogeneous linear equation set constructed based on the DC power flow model in the step 1 is as follows: wherein A is a node-line incidence matrix, Is the congestion component vector at time t, And X is the line reactance reciprocal vector to be inferred, which is the shadow price vector at the time t.
  3. 3. The method for reverse inference of power grid admittance parameters based on physical constraint deep learning according to claim 1, wherein the specific method for filtering data by using the residual tree theory in the step 1 comprises the following steps: Determining a residual tree corresponding to any one spanning tree aiming at the power grid topology comprising N nodes and L lines, wherein the residual tree comprises L-N+1 lines; and selecting electricity price data of L-N+1 lines in the residual tree when congestion occurs in different time periods respectively as a minimum complete data set for enabling the equation set coefficient matrix to be full.
  4. 4. The method for reverse inference of power grid admittance parameters based on physical constraint deep learning according to claim 1, wherein the hybrid head attention MDR mechanism adopted by the improved encoder in the step 3 specifically comprises the following steps: the mixed-head attention MDR mechanism comprises h attention heads and a router, wherein the router activates Top-k attention heads for each Token, and the h attention heads are divided into Shared header A plurality of routing heads; The router calculates a route score for each input Token ; The weighted sum of the shared header and the activated routing header is calculated as output, and the output calculation formula is as follows: Wherein the method comprises the steps of As a projection matrix, route scores only when the ith head is activated And is not 0.
  5. 5. The reverse power grid admittance parameter inference method based on physical constraint deep learning according to claim 1, wherein the Mamba module in the step 3 is based on a selective state space model, and the specific characteristics of the method include: Introducing a selection mechanism to dynamically change parameters delta and B, C of a state space model along with input X (t), wherein the parameters are specifically calculated as follows: And discretizing continuous parameters by using a zero-order retention method, and adopting a parallel association scanning algorithm to realize the sequence modeling of linear time complexity so as to process the long-sequence power grid characteristic data.
  6. 6. The method for reverse inference of admittance parameters of a power grid based on physical constraint deep learning according to claim 1, wherein the total loss function L in the step 4 is constructed as follows: Wherein, the For the admittance matrix of the model predictions, As a matrix of true admittances, For line flow calculated based on the predicted admittance, For a line flow calculated based on the true admittance, Is a weight coefficient.

Description

Power grid admittance parameter reverse inference method based on physical constraint deep learning Technical Field The invention belongs to the technical field of intelligent power grid network safety, and particularly relates to a power grid admittance parameter reverse inference method based on physical constraint deep learning. Background With the development of transparency of the electric power market, system operators (ISO) generally issue market data such as node marginal electricity prices (LMPs) and shadow prices to the public, and the purpose of the system operators is to ensure fairness and transparency of electric power market transactions. However, related studies have shown that it is possible for an attacker to deduce the admittance matrix, a vital sensitive physical parameter in the power grid, in reverse, from these published data. The admittance matrix covers the detailed topological structure of the power grid, line parameters and other core information, and once the matrix information is revealed, an attacker can use the matrix information to initiate False Data Injection Attack (FDIA) or cascading failure attack, which can form serious threat to the safe and stable operation of the power grid. The existing reverse inference method mainly has two types of remarkable defects. One type is a traditional mathematical optimization method, such as a least square method, which has large data demand and extremely high computational complexity when dealing with large-scale power grid related problems, and often suffers from the problem of dimension disaster, so that the application of the method in actual large-scale power grid scenes is extremely limited. Another type is a common data driving method, such as a Convolutional Neural Network (CNN), a long-short-term memory network (LSTM), etc., and although such methods have relatively high calculation speed, they lack compliance with the physical rules of the power system, and parameters deduced by these methods often violate kirchhoff's law, for example, a negative resistance or the like, which does not conform to the actual physical characteristics. In addition, these methods also do not explicitly "reach the minimum data volume required for the maximum inferred threat" from the theoretical level, which makes it difficult for the grid operators to accurately evaluate the costs of related attacks and threat boundaries faced by the grid data privacy, and is not beneficial to the formulation and implementation of the grid data privacy protection policy. Disclosure of Invention Aiming at the situation, in order to overcome the defects of the prior art, the invention provides a power grid admittance parameter reverse inference method based on physical constraint deep learning, which effectively solves the problems proposed by the background art. In order to achieve the purpose, the invention provides a power grid admittance parameter reverse inference method based on physical constraint deep learning, which comprises the following steps: step 1, data acquisition and theoretical screening Constructing a homogeneous linear equation set based on a direct current power flow model, screening minimum electricity price data capable of enabling the coefficient matrix rank of the equation set to reach the maximum from historical data by utilizing a residual tree Cotree theory in a graph theory, and constructing a training data set; Step2, feature extraction and construction Extracting node characteristics and edge characteristics from the screened data set, wherein the node characteristics comprise congestion component MCC and topology information in LMP, and the edge characteristics comprise line shadow prices; step 3, constructing PhyG-Mamba inference model The method comprises the steps of constructing a deep neural network model comprising an improved encoder and Mamba modules, wherein the improved encoder adopts a mixed head attention MDR mechanism to dynamically route and encode node characteristics, and the Mamba modules process sequence data fused with node coding characteristics and edge characteristics based on a selective state space model to capture global dependency relations; Step 4, physical constraint optimization training Substituting the prediction admittance matrix output by the model into a direct current power flow equation, calculating physical constraint loss, constructing a total loss function by combining the prediction numerical error, and carrying out iterative update on model parameters; Step 5, reverse direction inference And (3) inputting public power data of a period to be tested by using the trained PhyG-Mamba model, and outputting admittance matrix parameters of the target power grid. Preferably, the homogeneous linear equation set constructed based on the dc power flow model in the step 1 is: wherein A is a node-line incidence matrix, Is the congestion component vector at time t,And X is the line reactance reciprocal vector