CN-121997325-A - Double-target attack resistance method for wind power prediction model

CN121997325ACN 121997325 ACN121997325 ACN 121997325ACN-121997325-A

Abstract

The invention discloses a double-target anti-attack method for a wind power prediction model, which belongs to the technical field of safety and artificial intelligence of power systems and comprises the following steps of S1, collecting original data and determining parameters, S2, training a timing diagram self-encoder detection model, S3, constructing and training the wind power prediction model, S4, constructing a double-target optimization function, S5, generating an anti-sample in an iteration mode, and S6, injecting and detecting the anti-sample. The wind power prediction model double-target anti-attack method is used for realizing the combined adjustment of attack destructiveness and concealment, reducing the detection probability and supporting deflection selection through the double-target weighting function and the reconstruction loss term coefficient adjustment, and improving the GAE structure capture time sequence dependence, avoiding artificial characteristic design and improving the accuracy of the detection time sequence anti-attack sample.

Inventors

LIANG GAOQI
LIANG CHENGHAO

Assignees

哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院)

Dates

Publication Date: 20260508
Application Date: 20260126

Claims (10)

1. The double-target attack resistance method of the wind power prediction model is characterized by comprising the following steps of: s1, collecting multi-period meteorological factor data, and determining rated power of a wind power plant, the number of predicted periods, the upper limit value of tamper proportion and the upper limit of iteration times; S2, training a time sequence diagram self-encoder detection model, wherein the time sequence diagram self-encoder takes samples of each time period as nodes, time sequence association of adjacent time periods as edges, node characteristics are weather factor values of corresponding time periods, the encoder is a two-layer diagram convolution network, the decoder is a layer of linear transformation, and the training target is to minimize mean square error of original and reconstructed weather data; s3, constructing and training a wind power prediction model, wherein the model is a fully connected network, inputs the multi-period meteorological factor data in S1, and outputs the multi-period meteorological factor data as wind power prediction power in a corresponding period; S4, constructing a double-target weighted summation objective function, wherein the double targets comprise attack destructiveness and attack disguise, the attack destructiveness is realized by maximizing wind power deviation, and the attack disguise is realized by minimizing reconstruction loss of a timing diagram self-encoder; S5, iteratively updating the countermeasure sample based on the multi-period meteorological factor data of S1, the wind power prediction model of S3, the time sequence diagram self-encoder detection model of S2 and the objective function of S4, wherein the deviation between each meteorological factor and the original value does not exceed the upper limit of the tamper proportion, and the optimal countermeasure sample with the minimum objective function value is output after the iteration is carried out for the upper limit times in the updating process; s6, injecting the optimal countermeasure sample into a meteorological database of the power system, and enabling a defender to calculate reconstruction loss of the input sample from the encoder model through the time sequence diagram of S2, and identifying the countermeasure sample according to the difference of the reconstruction loss.
2. The method for dual-objective attack resistance according to claim 1, wherein in S1, the meteorological factors include wind speed, temperature, relative humidity, ground air pressure and wind direction angle at 10m of the ground.
3. The method for dual-objective attack resistance of wind power prediction model according to claim 2, wherein in S2, the encoder mapping formula of the timing diagram self-encoder is specifically: the encoder first layer, denoted: ; in the formula, For the entered value of the weather factor, For the total number of predicted time periods, The number of meteorological factors; in order to embed the representation in the hidden layer, Is a hidden layer dimension; Is a weight matrix which can be learned; Is a normalized adjacency matrix; Is an activation function; The encoder second layer, denoted as: ; in the formula, A representation is embedded for the potential layer, Is a potential layer dimension; Is a weight matrix which can be learned; The decoder mapping formula is expressed as: ; in the formula, Decoding the weight matrix; Is the output result of the decoder, i.e. the reconstructed value of the meteorological factors.
4. The method for dual-objective attack resistance of wind power prediction model according to claim 3, wherein in S2, the mean square error loss function formula is: ; in the formula, Reconstruction loss; Is that Time period of Original values of individual meteorological factors; Is that Time period of And (5) reconstructing values of the meteorological factors.
5. The method for dual-objective attack resistance of a wind power prediction model according to claim 4, wherein in S3, the linear fully-connected network has a 6-layer structure, and comprises 1 input layer, 4 hidden layers and 1 output layer, and Relu activation functions are adopted between the hidden layers and the output layer.
6. The method for dual-objective attack resistance of wind power prediction model according to claim 5, wherein in S4, the dual-objective function formula is: ; in the formula, Is the attack direction; Respectively is The power after moment attack and the original predicted power; Rated power for the wind farm; The loss term coefficients are reconstructed from the encoder for the timing diagram.
7. The method for double-objective attack resistance of wind power prediction model according to claim 6, wherein, Is a forward attack, namely an attack of increasing wind power; is a negative attack, namely an attack for reducing wind power.
8. The method for twin object challenge according to claim 7, wherein in S5, the process of iteratively updating challenge samples comprises the steps of: S501, inputting multi-period meteorological factor data into a wind power prediction model of S3 to obtain original predicted power And will be the first Inputting the countermeasures sample of the secondary iteration into a wind power prediction model to obtain the countermeasures power And calculating a power deviation, expressed as: ; in the formula, Is that Power deviation at time; S502, will be The timing diagram of the counter-sample input S2 for the next iteration is calculated from the encoder model to obtain the reconstruction loss ; S503, will 、、 Substitution into S4 Obtaining a current objective function value; s504, updating the countermeasure sample based on the gradient of the objective function to generate the first The iteration is conducted on the countermeasure sample, and meanwhile, the deviation between each meteorological factor and the normal sample in all time periods is required to be met, and the deviation does not exceed the tamper proportion upper limit value determined by S1; When the iteration number reaches the upper limit, outputting the objective function value The smallest corresponding challenge sample is taken as the optimal challenge sample.
9.A computer device comprising a processor for coupling with a memory, reading and executing instructions and/or program code in the memory to perform the method of any of claims 1-8.
10. A computer readable medium, characterized in that the computer readable medium stores computer program code which, when run on a computer, causes the computer to perform the method according to any of claims 1-8.

Description

Double-target attack resistance method for wind power prediction model Technical Field The invention belongs to the technical field of electric power system safety and artificial intelligence intersection, and particularly relates to a wind power prediction model double-target attack resistance method. Background The permeability of wind power generation is continuously improved, and the accuracy of wind power prediction is increasingly important for the safety scheduling of an electric power information physical system. The meteorological factors are used as the input of a wind power prediction model, and are easy to be attack targets, so that the wind power prediction power is deviated, and the normal scheduling of the power system is affected. An attacker usually adopts a counterattack mode to launch attack on meteorological factors, namely, a counterattack sample (namely, meteorological factors) containing tiny disturbance is designed and is injected into a power system, so that the output of a wind power prediction model generates significant deviation. However, the prior art has the following defects that the prior art (FGSM and PGD) only focuses on 'attack destructiveness' (maximizing power deviation), ignores 'attack concealment', and leads to that an anti-sample is easy to be detected by a defender, wherein the maximum destructiveness is difficult to realize by single iteration of the FGSM, the variable tampering proportion is not controlled, the PGD is limited by the tampering proportion and has no concealed design, the SVM model depends on the characteristic engineering difficulty when adapting to high-dimensional meteorological data, the traditional self-encoder can reconstruct the sample only at a single moment and cannot capture abnormal time sequence association, and the traditional self-encoder focuses on node space association, lacks time sequence dependency capturing capability and cannot detect the time sequence anti-sample. Thus, a new method is needed. The related technical means of the attack and defense game of the power system are only used for constructing and verifying a self-safety protection system of the power network, aim to identify and resist external malicious attacks in advance, ensure the stable operation of the power system, and are never used for implementing malicious invasion and damage to power facilities or power grids. Disclosure of Invention The invention aims to provide a wind power prediction model double-target anti-attack method, which realizes attack destructiveness and concealment combined optimization, reduces detection probability and supports deflection selection through double-target weighting function and reconstruction loss term coefficient adjustment, and improves the accuracy of detection time sequence anti-sample through improving the dependence of a GAE structure capturing time sequence without artificial feature design. In order to achieve the above purpose, the invention provides a wind power prediction model double-target attack resistance method, which comprises the following steps: s1, collecting multi-period meteorological factor data, and determining rated power of a wind power plant, the number of predicted periods, the upper limit value of tamper proportion and the upper limit of iteration times; S2, training a time sequence diagram self-encoder detection model, wherein the time sequence diagram self-encoder takes samples of each time period as nodes, time sequence association of adjacent time periods as edges, node characteristics are weather factor values of corresponding time periods, the encoder is a two-layer diagram convolution network, the decoder is a layer of linear transformation, and the training target is to minimize mean square error of original and reconstructed weather data; s3, constructing and training a wind power prediction model, wherein the model is a fully connected network, inputs the multi-period meteorological factor data in S1, and outputs the multi-period meteorological factor data as wind power prediction power in a corresponding period; S4, constructing a double-target weighted summation objective function, wherein the double targets comprise attack destructiveness and attack disguise, the attack destructiveness is realized by maximizing wind power deviation, and the attack disguise is realized by minimizing reconstruction loss of a timing diagram self-encoder; S5, iteratively updating the countermeasure sample based on the multi-period meteorological factor data of S1, the wind power prediction model of S3, the time sequence diagram self-encoder detection model of S2 and the objective function of S4, wherein the deviation between each meteorological factor and the original value does not exceed the upper limit of the tamper proportion, and the optimal countermeasure sample with the minimum objective function value is output after the iteration is carried out for the upper limit times in the updating process; s6