Power load multi-objective robust deep forecasting method and system against adversarial attacks

Abstract

This invention discloses a robust multi-objective deep power load prediction method and system for defending against adversarial attacks. The method simulates adversarial examples for power load prediction based on the fast gradient sign method and projected gradient descent attacks. The CNN-BiLSTM model architecture parameters and training hyperparameters are used as optimization variables. During optimization, models with a power load prediction MSE below a set threshold undergo adversarial training. The three optimization objective functions are minimizing the model's MSE on clean samples, minimizing the model's MSE after adversarial training, and minimizing the number of model parameters. A multi-objective variable-length discrete extremum optimization technique is designed for offline multi-objective iterative optimization to obtain a compromise power load prediction model that achieves high prediction accuracy, strong adversarial robustness, and lightweight deployment. This model is then redeployed into the power load prediction system. This invention achieves proactive adversarial defense and lightweight deployment of the deep power load prediction model from the automated design stage of the adversarial robust model.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction technology in the field of power system load forecasting, specifically to a robust deep prediction method and system for multi-objective power load with defense against attacks. Background Technology

[0002] With the rapid evolution of new power systems, the integration of high proportions of renewable energy and the deep interaction between power generation, grid, load, and storage have significantly exacerbated the spatiotemporal volatility and uncertainty of power load. Against this backdrop, accurate power load forecasting has become a core supporting technology for ensuring the safe and stable operation of the power grid, optimizing resource allocation, and improving economic efficiency. In recent years, deep learning (DL), with its powerful data processing and feature extraction capabilities, has been widely applied in the field of power load forecasting, achieving more accurate predictions than traditional methods. However, while DL models bring performance improvements, their high dependence on smart terminals and communication networks introduces new security risks. The models are highly vulnerable to network attacks during data transmission and deployment, particularly adversarial attacks such as the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). These attacks are characterized by small disturbance amplitudes, strong concealment, and difficulty in being detected by traditional monitoring methods. They can induce the DL model to output prediction results that deviate significantly from the true value by injecting carefully designed malicious disturbances into the input data. This can mislead power grid dispatch decisions and, in severe cases, even cause major safety accidents such as large-scale power outages or system paralysis, posing a serious challenge to the power grid's security defenses.

[0003] Current research on robustness in power load forecasting remains significantly limited in addressing practical application needs. On one hand, while adversarial training-based defense strategies improve robustness, they often lead to redundant model parameters and neglect model complexity. On the other hand, research focusing on lightweight models typically fails to adequately consider stability under adversarial attacks. Existing optimization frameworks rarely incorporate prediction accuracy, adversarial robustness, and lightweight models into a coordinated optimization system, resulting in models that are difficult to deploy efficiently and robustly on power edge terminals with limited computing power and resources. To date, existing technologies lack multi-objective coordinated optimization that considers power load forecasting accuracy, adversarial robustness, and lightweight deep prediction models during the power load forecasting model design phase. Therefore, there is an urgent need to research power load forecasting technologies that balance high accuracy, strong adversarial robustness, and lightweight model optimization to achieve proactive adversarial defense for power load forecasting models. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a robust multi-objective depth prediction method and system for power load defense against attacks.

[0005] The objective of this invention is achieved through the following technical solution: The first aspect of this invention provides a robust depth prediction method for multi-objective power load defense against counter-attacks, comprising the following steps: S1. Construct normal training and normal test datasets based on the historical power system load database, and construct a real-time dataset based on the real-time power system load database. For the existing deep learning prediction model CM for power load, simulate and generate power load prediction adversarial attack training datasets using the fast gradient sign method and projective gradient descent attack, respectively. D adv_tr and adversarial attack test dataset D adv_te ; S2. Set the first parameter and construct the multi-objective optimization objective function of the hybrid prediction model CNN-BiLSTM, which combines a multi-objective adversarial robust convolutional neural network and a bidirectional long short-term memory network for power load. The multi-objective optimization objective function includes the mean squared error of the prediction of the power load adversarial robust CNN-BiLSTM prediction model on normal samples. f 1. Prediction mean square error on adversarial examples f 2 and the number of model parameters f 3; S3. Set the second parameter, construct a multi-objective adversarial robust CNN-BiLSTM prediction code for power load, and generate individuals to be mutated. p g And set the current iteration number. g =0; S4. Dealing with mutant individuals p g Perform mutation operations to generate a population. P g Assess the population P g medium-sized individuals optimization objectives f 1 and f 3, and determine f Is 1 less than or equal to the adversarial training threshold? T If not, set f 2=1, interrupting the individual If so, proceed to step S6 for individual adversarial training; Perform adversarial training and calculate the optimal objective. f 2; S5, for population P g All individuals according to the optimization objectivef 1. f 2 and f 3. Perform non-dominated sorting and store non-dominated solutions in an external archive solution set. S In, and based on external archive update technology S Update and obtain the individuals to be mutated for the next optimization iteration. p g Number of update iterations g = g +1; S6, based on D adv_tr For individuals Conduct adversarial training, and then based on D adv_te Calculate the optimization objective f 2; S7, when the number of iterations g Less than the set maximum number of iterations G When, repeat steps S4~S6; when g Greater than or equal to G At that time, mark the external archive solution set. S All solutions are Pareto front solutions. CNN-BiLSTM, which balances high prediction accuracy, strong adversarial robustness, and lightweight design, is selected and marked as the optimal model for multi-objective adversarial robust deep prediction of power load (BM). S8. Deploy model BM online in the power system, targeting the real-time dataset in the power system load real-time database. D t Perform online power system load forecasting and test and evaluate its robustness performance indicators.

[0006] Further, step S1 specifically includes: Based on the historical power system load database, historical data related to power system load are extracted and parsed to obtain a normal dataset. This normal dataset is then divided into a normal training dataset. D tr and normal test dataset D te Based on the real-time power system load database, real-time data related to power system load are extracted and parsed to obtain a real-time dataset. D t ; Will D tr The elements in the matrix are labeled as a m,h , a m,h This represents the first [item] in the normal training dataset. m The first sample heigenvalues, of which m express D tr The row number of the matrix corresponds to the row number in the normal training dataset. m samples, 1≤ m ≤ M , M This represents the maximum number of samples in the normal training dataset. h express D tr , D te and D t The column indices of the matrix correspond to the first, second, and third columns of the normal training dataset, normal test dataset, and real-time dataset, respectively. h Features, 1≤ h ≤ H , H This represents the maximum number of features in the normal training dataset, normal test dataset, and real-time dataset; D te The elements in the matrix are labeled as a w,h , a w,h This represents the first [item] in the normal test dataset. w The first sample h eigenvalues, of which w express D te The row number of the matrix corresponds to the row number in the normal test dataset. w samples, 1≤ w ≤ W , W This represents the maximum number of samples in the normal test dataset; D t The elements in the matrix are labeled as a rt,h , a rt,h Represents the first in the real-time dataset rt The first sample h eigenvalues, of which rt express D t The row number in the matrix corresponds to the row number in the real-time dataset. rt samples, 1≤ r t ≤ n rt , n rt This represents the maximum number of samples in the real-time dataset; To address the existing deep learning-based power load prediction model CM, we simulate and generate adversarial attack training datasets for power load prediction using the fast gradient sign method and projective gradient descent attacks, respectively. D adv_tr and adversarial attack test dataset D adv_te The CM model architecture comprises: a Convolutional Neural Network (CNN) module, a global average pooling layer, a Bidirectional Long Short-Term Memory (BiLSTM) network layer, and a fully connected layer stacked together. The CNN module includes convolutional layers and pooling layers. The CM model parameters include { ppt 1, ppt 2, ppt 3 , ppt 4}, ppt 1 represents the set of parameters for the number of modules in the convolutional neural network and the number of layers in the bidirectional long short-term memory network. ppt 2 represents the parameter set of the convolutional neural network module. ppt 3 represents the parameter set of the bidirectional long short-term memory network layer. ppt 4 represents the set of configuration parameters for the model training hyperparameters; ppt 1={ C 1, B 1}, where C 1 represents the number of modules in the convolutional neural network. B 1 represents the number of layers in the bidirectional long short-term memory network; ,in k 0 indicates the first k 0 convolutional neural network modules For the first k The number of convolutional filters in the convolutional layer of a 0-convolutional neural network module. For the first k The kernel size of a convolutional layer with 0 convolutional neural network modules. For the first k The activation function type of the convolutional layer in a convolutional neural network module with 0 modules. For the first k The pooling type used after the convolutional layers of 0 convolutional neural network modules. For the first k The pooling size used after the convolutional layers of 0 convolutional neural network modules; ,in j 0 indicates the first j 0 bidirectional long short-term memory network layers, For the first j The number of neurons in 0 bidirectional long short-term memory network layers For the first j Activation function types for 0 bidirectional long short-term memory network layers; ppt 4={ btc , dr c , lr c , opt c , ret c},in bt c For CM training batch size, dr c This refers to the dropout rate during CM training. lr c For CM training learning rate, opt c The type of optimizer used for CM training. ret c The regularization type used for CM training; the adversarial attack training dataset. D adv_tr and adversarial attack test dataset D adv_te Specifically, it is obtained through the following sub-steps: S11. When the adversarial attack method is the fast gradient sign method adversarial attack, the adversarial attack training dataset... D adv_tr and adversarial attack test dataset D adv_te The simulation generation process is shown in S11.1 to S11.4: S11.1 Initialize the strength of the fast gradient sign method adversarial attack perturbation. ε 1. Generate fast gradient sign-based adversarial attack samples according to formula (1). X advFGSM : (1) In the formula, x Represents the original sample; d x Represents the original input sample x The total number of feature dimensions; Indicates the original sample x The partial derivatives; Indicates sample x China's power load i The true values ​​of 0 features; Indicates the first i 0 feature loss functions θ For the network parameters of the CM model; sign represents the gradient calculation symbol; This indicates that the generated adversarial perturbation is constrained within the range [0,1]. S11.2, When using the fast gradient sign method adversarial attack training dataset D adv_tr_FGSM When generating adversarial examples, set x = x adv_tr_FGSM,tx ,1≤ tx ≤ Tr ,in x adv_tr_FGSM,tx This represents the training dataset for fast gradient signified adversarial attacks. D adv_tr_FGSM The first in tx One original sample, Tr This indicates that from the normal training dataset D tr Repeat the fast gradient sign-based adversarial sample generation process in step S11.1 without repeating the randomly selected number of samples, until the fast gradient sign-based adversarial attack training dataset is reached. D adv_tr_FGSM All original samples x adv_tr_FGSM,tx All were replaced with the fast gradient sign-based adversarial examples calculated by formula (1). X adv_tr_FGSM,tx ,in X adv_tr_FGSM,tx express D adv_tr_FGSM The simulated generation of the first tx One fast gradient sign adversarial example; S11.3, When using the Fast Gradient Signed Method adversarial attack test dataset D adv_te_FGSM When generating adversarial examples, set x = x adv_te_FGSM,ty ,1≤ ty ≤ Te ,in, x adv_te_FGSM,ty This represents the test dataset D for fast gradient signified attack. adv_te_FGSM The first in ty One original sample, Te This indicates the normal test dataset. D te Repeat the fast gradient sign-method adversarial sample generation process in step S11.1 above, without repeating the randomly selected number of samples, until the fast gradient sign-method adversarial attack test dataset is obtained. D adv_te_FGSM All original samples x adv_te,ty All were replaced with adversarial samples calculated using formula (1). X adv_te_FGSM,ty ,in Xadv_te_FGSM,ty for D adv_te_FGSM The simulated generation of the first ty One adversarial example; S11.4 The generated fast gradient signified adversarial attack training dataset D adv_tr_FGSM For normal training dataset D tr Replace the sample at the corresponding position in the middle, and then complete the above replacement. D tr The dataset is labeled as an adversarial attack training dataset. D adv_tr The generated fast gradient symbolic method adversarial attack test dataset will be used. D adv_te_FGSM For normal test dataset D te Replace the sample at the corresponding position in the middle, and then complete the above replacement. D te Dataset labeling adversarial attack test dataset D adv_te ; S12. When the adversarial attack method is projective gradient descent adversarial attack, the adversarial attack training dataset... D adv_tr and adversarial attack test dataset D te The simulation generation process is shown in S12.1 to S12.7: S12.1 Initialize the strength of the projection gradient descent adversarial attack perturbation. ε 2. Adversarial attack step size of projected gradient descent α 1. Number of iterations for adversarial attack against projected gradient descent ζ 1. Initial values ​​for projected gradient descent to counteract disturbances based on CM , This represents the initial perturbation matrix with the same shape as the original sample, and records the number of iterations of the projected gradient descent adversarial generation. t =0; S12.2. According to formula (2), iteratively generate the first... t +1-order projected gradient descent to counteract perturbations And apply perturbation constraints: (2) In the formula, x Represents the original sample; This indicates that the generated adversarial perturbation is performed within the range of [- ε 2, ε Constraints of 2]; Indicates the iterative generation of the first... tSubprojected gradient descent perturbation, 0≤ t < ζ 1; Indicates resistance to disturbance The partial derivatives; y true This represents the actual value of the sample power load; L ( θ ( x + ), y true ) represents the loss function. θ For the network parameters of the CM model; sign represents the gradient calculation symbol; S12.3. According to formula (3), generate the iteration number. t +1 adversarial examples of projective gradient descent that satisfy the range constraint : (3) In the formula, This indicates that the generated projective gradient descent adversarial examples are constrained to have a range of [0,1]; and the number of iterations for generating the projective gradient descent adversarial examples is recorded. t = t +1; The number of iterations for generating adversarial examples using projective gradient descent reaches [number]. ζ The adversarial examples generated after step 1 are X advPGD In the training dataset for projective gradient descent adversarial attacks D adv_tr_PGD When generating initial adversarial examples, x = x adv_tr_PGD,tx ,1≤ tx ≤ Tr ,in x adv_tr_PGD,tx This represents the training dataset for projective gradient descent adversarial attacks. D adv_tr_PGD The first in tx One original sample, Tr This represents the training dataset for projective gradient descent adversarial attacks. D adv_tr_PGD The total number of all samples; in the test dataset for projected gradient descent adversarial attacks. D adv_te_PGD When generating initial adversarial examples, x = x adv_te_PGD,ty ,1≤ ty ≤ Te ,in ty This represents the dataset used for adversarial attacks against projective gradient descent. Dadv_te_PGD The first in ty One original sample, Te This represents the dataset used for adversarial attacks against projective gradient descent. D adv_te_PGD The total number of all samples; S12.4, The maximum number of iterations for generating adversarial examples using projected gradient descent is reached. ζ The adversarial examples generated after step 1 are labeled as follows: X advPGD ; S12.5, On the training dataset for projected gradient descent adversarial attacks D adv_tr_PGD When generating adversarial examples for projected gradient descent, set x = x adv_tr_PGD,tx ,1≤ tx ≤ Tr ,in x adv_tr_PGD,tx This represents the training dataset for projective gradient descent adversarial attacks. D adv_tr_PGD The first in tx For each original sample, repeat steps S12.1 to S12.4 of the adversarial example generation process until the training dataset for projected gradient descent adversarial attacks is obtained. D adv_tr_PGD All original samples x adv_tr_PGD,tx All were replaced with projected gradient descent adversarial examples. X adv_tr_PGD,tx ,in X adv_tr_PGD,tx express D adv_tr_PGD The Middle tx One projected gradient descent adversarial example; S12.6, On the test dataset for adversarial attacks on projected gradient descent. D adv_te_PGD When generating adversarial examples for projected gradient descent, set x = x adv_te_PGD,ty ,1≤ ty ≤ Te ,in, ty This represents the dataset used for adversarial attacks against projective gradient descent. D adv_te_PGD The first in ty For each original sample, repeat steps S12.1 to S12.4 of the adversarial sample generation process until an adversarial attack test dataset is generated. D adv_te_PGD All original samples x adv_te,tyAll were replaced with projected gradient descent adversarial examples X adv_te_PGD,ty ,in X adv_te_PGD,ty express D adv_te_PGD The Middle ty One projected gradient descent adversarial example; S12.7 The generated projection gradient descent adversarial attack training dataset D adv_tr_PGD For normal training dataset D tr Replace the sample at the corresponding position in the middle, and then complete the replacement. D tr The dataset is labeled as an adversarial attack training dataset. D adv_tr The generated projection gradient descent adversarial attack test dataset D adv_te_PGD For normal test dataset D te Replace the sample at the corresponding position in the middle, and then complete the replacement. D te The dataset is labeled as an adversarial attack test dataset. D adv_te .

[0007] Furthermore, the first parameter includes the maximum number of offline iterations. G Maximum number of solutions in the archive A max Adversarial training threshold T ; The mean squared error of the prediction model for the power load anti-robust CNN-BiLSTM on normal samples f 1. Prediction mean square error on adversarial examples f 2 and the number of model parameters f 3. As shown in formulas (4) to (6): (4) (5) (6) In the formula, y w This represents the actual value of the power load. This represents the predicted value of a normal sample of power load. This represents the predicted value of the power load countermeasure sample; Indicates the first l The number of parameters in a CNN module C Indicates the number of CNN modules; Indicates the firstk The number of parameters in a BiLSTM layer. B Indicates the number of BiLSTM layers; P Dense This indicates the number of parameters in a fully connected network layer.

[0008] Furthermore, the second parameter includes a preset maximum number of CNN modules. C max and minimum value C min Maximum number of BiLSTM layers B max and minimum value B min Maximum number of convolutional kernels in a CNN module fn max and minimum value fn min , No. l Minimum number of convolutional kernels in a CNN module fn min and maximum value fn max , No. l Minimum kernel size of a CNN module kn min and maximum value kn max , No. l The activation function of each CNN module encodes the minimum value. an min and maximum value an max , No. l The minimum value is encoded by pooling type of each CNN module. pt min and maximum value pt max , No. l The minimum value is encoded by pooling type of each CNN module. ps min and maximum value ps max , No. k Minimum number of neurons in a layer BiLSTM nb min and maximum value nb max , No. k Layer BiLSTM activation function encodes minimum value ab min and maximum value ab max Minimum batch size bt min and maximum value bt maxMinimum value of optimizer encoding opt min and maximum value opt max Minimum learning rate lr min and maximum value lr max The minimum value encoded by regularization ret min and maximum value ret max Minimum discard rate dr min and maximum value dr max ; The construction of a multi-objective adversarial robust CNN-BiLSTM predictive code for power load generates individuals to be mutated. p g Specifically, this includes: encoding the parameters of the CNN-BiLSTM model to be optimized based on the second parameter. The encoding of each CNN-BiLSTM model to be optimized includes the following four parts: pt 1={ C , B}, pt 2={ pt 2,l = ( fn l , kn l , an l , pt l , ps l ), l =1, 2, …, C}, pt 3={ pt 3,k =( nb k , ab k ), k =1, 2, …, B}, pt 4=( bt , opt , lr , ret , dr ), pt 1 represents the set of numbers of CNN modules and BiLSTM layers. pt 2 represents the set of neural network parameters for the CNN module. pt 3 represents the parameter set of the BiLSTM layer neural network. pt4 represents the set of hyperparameters used in model training; C This indicates the number of CNN modules in the CNN-BiLSTM model to be optimized. B This indicates the number of BiLSTM layers in the CNN-BiLSTM model to be optimized; pt 2,l Indicates the first l CNN module parameter encoding fn l Indicates the first l Number of convolutional kernels per CNN module kn l Indicates the first l The size of the convolutional kernel in each CNN module. an l Indicates the first l Activation functions for each CNN module pt l Indicates the first l Each CNN module pooling layer type ps l Indicates the first l The pooling kernel size for each CNN module; pt 3,k Indicates the first k Layer BiLSTM parameter encoding, nb k Indicates the first k Number of BiLSTM neurons per layer ab k Indicates the first k Layer BiLSTM activation function; bt This indicates the batch size for training the CNN-BiLSTM model. opt This represents the training optimizer for the CNN-BiLSTM model. lr This represents the learning rate used to train the CNN-BiLSTM model. ret This indicates the regularization method used during the training of the CNN-BiLSTM model. dr This represents the dropout rate during the training of the CNN-BiLSTM model; in, pt 1. pt 2,l , pt 3,k , pt The specific values ​​of 4 are shown in formulas (7) to (10): (7) In the formula, Randint represents the random integer selection function. C min This represents the minimum value of the number of CNN modules. Cmax This represents the maximum value that can be taken for the number of CNN modules. C min ≤ C ≤ C max ; B min This represents the minimum value for the number of BiLSTM layers. B max This represents the maximum value of the BiLSTM layer number. B min ≤ B ≤ B max ; (8) In the formula, fn min Indicates the first l Minimum number of convolutional kernels for a CNN module fn max Indicates the first l The maximum number of convolutional kernels in a CNN module. fn min ≤ fn l ≤ fn max ; kn min Indicates the first l The minimum kernel size for each CNN module. kn max Indicates the first l The maximum value of the convolutional kernel size of each CNN module. kn min ≤ kn l ≤ kn max ; an min Indicates the first l The minimum value encoded by the activation function of each CNN module. an max Indicates the first l The maximum value encoded by the activation function of each CNN module. an min ≤ an l ≤ an max ; pt min Indicates the first l The minimum value of pooling type encoding for each CNN module. pt max Indicates the first lThe maximum value of the pooling type encoding of each CNN module. pt min ≤ pt l ≤ pt max ; ps min Indicates the first l The minimum value of pooling type encoding for each CNN module. ps max Indicates the first l The maximum value of the pooling type encoding of each CNN module. ps min ≤ ps l ≤ ps max ; (9) In the formula, nb min Indicates the first k Minimum number of neurons in a layer of BiLSTM nb max Indicates the first k The maximum number of neurons in a layer of BiLSTM. nb min ≤ nb k ≤ nb max ; ab min Indicates the first k The minimum value encoded by the activation function of the BiLSTM layer. ab max Indicates the first k The maximum value encoded by the activation function of the BiLSTM layer. ab min ≤ [[ID=......]] k ≤ ab max ; (10) In the formula, Random represents the function for selecting random decimals; ab min This represents the minimum batch size. bt max This represents the maximum value of the batch size. bt min ≤ bt ≤ bt max ; bt min This represents the minimum value encoded by the optimizer. optmax This represents the maximum value encoded by the optimizer. opt min ≤ opt ≤ opt max ; lr min This represents the minimum learning rate. lr max This represents the maximum value of the learning rate. lr min ≤ lr ≤ lr max ; opt min This represents the minimum value encoded using regularization. ret max This represents the maximum value encoded using regularization. ret min ≤ ret ≤ ret max ; dr min This represents the minimum discard rate. dr max This represents the maximum discard rate. dr min ≤ dr ≤ dr max ; Based on formulas (7) to (10), individuals to be mutated are randomly generated. p g Its code is ,in g This represents the current iteration number. ret 2,1 This indicates the parameter encoding of the first CNN module. Indicates the first C CNN module parameter encoding pt 3,1 This indicates the encoding of the first layer BiLSTM parameters. Indicates the first B Layer BiLSTM parameter encoding.

[0009] Furthermore, step S4 specifically includes the following sub-steps: S41. Treat the variant individuals according to formulas (7) to (10). p g Each gene locus is mutated sequentially to generate a scale of N g population ,in It is a population P gThe first individual, Indicates population P g The i Individual, Indicates population P g The N g Individual, , C g yes The number of CNN modules, B g yes The number of BiLSTM layers; ,in Individual The corresponding [[ID=......]] 1, Represents an individual The corresponding pt 2, Represents an individual The corresponding pt 3, Represents an individual The corresponding pt 4, express The parameters of the first dimension, express No. d g Dimensional parameters, express No. D g Dimensional parameters, D g express Maximum dimension ; During initialization, i.e., when the number of iterations... g When =0, initialize the individual according to formulas (7) to (10). p 0 Each gene locus is mutated sequentially to generate a scale of N Initial population of 0 ,in It is the first individual in the initial population. Denotes the initial population's... i Individual, Denotes the initial population's... N 0 individuals , C 0 is The number of CNN modules, B 0 is The number of BiLSTM layers; ,in Individual of pt 1, Represents an individual of pt 2, Represents an individual of pt 3, Represents an individual of pt 4, express The parameters of the first dimension, express No. d 0-dimensional parameters express No. D 0-dimensional parameters D 0 represents Maximum dimension ; S42, Based on normal training dataset D tr right P g medium-sized individuals The CNN-BiLSTM prediction model was trained based on the normal test dataset. D te Calculate individuals The corresponding optimization objective of the prediction model after the above training f 1 and f 3. Evaluation f Is 1 less than or equal to the adversarial training threshold? T If not, set optimization goals. f 2=1, interrupting the individual The confrontation training; if so, proceed to the next step. S 6. For individuals Perform adversarial training and calculate the optimal objective. f 2.

[0010] Furthermore, step S5 specifically includes the following sub-steps: S51, Based on the optimization objective f 1. f 2 and f 3 pairs of populations P g Perform a non-dominated ranking on all individuals, and check each individual one by one whether it is completely dominated by other individuals. That is, if there is an individual... exist f 1.f 2 and f All 3 are smaller than the individual This indicates that the individual Complete domination of individuals Then delete the controlled individual. Individuals who are not completely dominated retain their individuality. Mark all retained individuals as the retained solution set. U ,in, express P g China differs from individuals, that is i ≠ j ; S52, Solution set U External archive deset S Merging is marked as a solution set Q According to the optimization objective f 1. f 2 and f 3 pairs Q All solutions are sorted in a non-dominated order, and each solution is evaluated to determine whether it is completely dominated by other solutions. That is, if a solution exists... exist f 1. f 2 and f All 3 are less than the solution This explains the solution. Complete Domination Solution Then delete the dominated solution. If not fully dominated, the solution is retained. Mark all retained solutions as the new external archive solution set. S That is, updating the external archive solution set. S ,in, and Both represent solution sets Q The solution in, and the solution Different from the solution ,Right now i 1≠ j 1; S53, Judgment Is it greater than the number of external archives? A max If not, proceed to step S54; if yes, then deset the external archive. S Crowding distance is calculated and sorted for each non-dominated solution, and redundant solutions are removed; Represents external archive solution set S The number of dominant solutions in the Middle East; S54, External Archive Decomposition SRandomly select a non-dominated solution and mark it as the individual to be mutated in the next optimization iteration. p g Number of update iterations g = g +1.

[0011] Further, in step S53, the calculation step of the congestion distance includes: S53.1, Decomposing external archives S The solutions are sorted in ascending order according to a certain optimization objective value, and the minimum value of the sorted solutions in terms of that objective value is marked. f min and maximum value f max Set the crowding distance of the target that is the minimum and maximum solution after sorting to infinity; S53.2, For each optimization objective, sort a non-dominated solution in ascending order. , Represents external archive solution set S Sort in ascending order i Two solutions are available. The congestion distance is calculated according to formula (11). d s Then, based on the distance of the crowding level d s External archive deset S Sort all non-dominated solutions in descending order: (11) In the formula, and These represent external archive solution sets. S Non-dominated solutions sorted in ascending order In optimizing objectives The optimization objective value of the adjacent solutions above, and They represent the optimization objectives for ascending order sorting, respectively. The maximum and minimum values ​​on; S53.3, Complete the external archive decomposition. S After calculating the crowding distance of all non-dominated solutions and sorting them in descending order, delete them. t A solution with a small crowding distance t = – T .

[0012] Further, step S6 specifically includes: Based on adversarial attack training dataset D adv_tr For individuals The CNN-BiLSTM neural network model was subjected to adversarial training and tested on an adversarial attack dataset. D adv_te Calculate the individual according to formula (5) The corresponding optimization objective of the prediction model after adversarial training f 2.

[0013] Furthermore, step S8 specifically includes: Deploying BM online in the power system for real-time datasets D t Online power load forecasting is performed, and its robustness performance is evaluated. The performance evaluation indicators include root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), and their calculation formulas are shown in formulas (12) to (14), respectively. (12) (13) (14) In the formula, express D t No. rt The actual power load value of each test sample. Indicates BM's target D t The Middle rt The predicted power load for each test sample.

[0014] A second aspect of this invention provides a robust multi-objective depth prediction system for power load defense against counter-attacks, used to implement the aforementioned robust multi-objective depth prediction method for power load defense against counter-attacks. The system includes: The module for constructing normal and adversarial datasets is used to build normal training and test datasets based on historical power system load databases, and real-time datasets based on real-time power system load databases. Furthermore, it simulates adversarial attack training datasets for existing deep learning power load prediction models (CM) using fast gradient sign-matrix and projective gradient descent attacks, respectively. D adv_tr and adversarial attack test dataset D adv_t ; The parameter setting and objective function construction module is used to set the first parameter and construct the multi-objective optimization objective function of the multi-objective adversarial robust CNN-BiLSTM prediction model for power load. The multi-objective optimization objective function includes the mean square error of the CNN-BiLSTM prediction model on normal samples. f1. Prediction mean square error on adversarial examples f 2 and the number of model parameters f 3; The optimized model encoding definition module is used to set the second parameter, construct a multi-objective adversarial robust CNN-BiLSTM predictive encoding for power load, and generate individuals to be mutated. p g And set the current iteration number. g =0; A population generation module based on mutation operations and a module for determining whether to enter adversarial training are used to treat mutated individuals. p g Perform mutation operations to generate a population. P g Assess the population P g medium-sized individuals optimization objectives f 1 and f 3, and determine f Is 1 less than or equal to the adversarial training threshold? T If not, set f 2=1, interrupting the individual If so, proceed to the adversarial training module for individual training. Conduct adversarial training and calculate the optimal objective. f 2; A multi-objective non-dominated solution evaluation and external archive update module is used for population analysis. P g All individuals according to the optimization objective f 1. f 2 and f 3. Perform non-dominated sorting and store non-dominated solutions in an external archive solution set. S In, and based on external archive update technology S Update and obtain the individuals to be mutated for the next optimization iteration. p g Number of update iterations g = g +1; Adversarial training module, used for training based on D adv_tr For individuals Conduct adversarial training, and based on D adv_te Calculate the optimization objective of an individual f 2; The iteration judgment module is used to determine the iteration count. g Less than the set maximum number of iterations GAt that time, the system repeatedly enters the population generation and adversarial training decision module based on mutation operations, the multi-objective non-dominated solution evaluation and external archive update module, and the adversarial training module; when g Greater than or equal to G At that time, mark the external archive solution set. S All solutions are Pareto front solutions. CNN-BiLSTM, which balances high prediction accuracy, strong adversarial robustness, and lightweight design, is selected and marked as the optimal model for multi-objective adversarial robust deep prediction of power load (BM). The online deployment module is used to deploy the selected model BM online in the power system, targeting the real-time dataset of the power system load database. D t Perform online power system load forecasting and test and evaluate its robustness performance indicators.

[0015] Compared with the prior art, the beneficial effects of the present invention are: (1) In the model design stage, the CNN-BiLSTM model architecture parameters and training hyperparameters to be designed are used as optimization variables. The three optimization objective functions are minimizing the model MSE on clean samples, minimizing the model MSE after adversarial training, and minimizing the number of model parameters. A multi-objective adversarial robust automatic deep load forecasting method and system that takes into account high-precision prediction performance on clean samples, strong adversarial robust performance and lightweight model is developed. This avoids the passive embarrassment of existing power load forecasting models having to perform post-attack defense after being deployed online and subjected to adversarial attacks.

[0016] (2) When facing different types of adversarial attacks such as FGSM and PGD, the present invention demonstrates significant adversarial robustness and can control the prediction error to an extremely low level, thereby providing an efficient solution with high prediction accuracy, resistance to adversarial attacks and lightweight online deployment to ensure the safe and stable operation of the power load prediction system. Attached Figure Description

[0017] pt The flowchart shows the multi-objective robust depth prediction method for power load defense against attacks according to the present invention. pt This is a schematic diagram of the structure and parameters of a deployed power load forecasting model CM in one example of the present invention; Figure 1 This is a schematic diagram of the adversarial sample generation and prediction process in the multi-objective robust depth prediction method for power load defense against adversarial attacks of the present invention. Figure 2 The individual to be mutated in this invention p g A diagram illustrating the correspondence between codes and their actual meanings; Figure 3This invention provides an example of Pareto front diagrams obtained from different adversarial attacks on an example dataset; wherein, Figure 4 (a) in the figure is the Pareto front obtained against FGSM adversarial attacks; Figure 5 (b) in the figure is the Pareto front obtained against PGD adversarial attacks; Figure 5 This is an example of the present invention, showing a comparison curve of the predicted and actual load values ​​of the prior art CM and the present invention BM under the adversarial attack described in the example dataset; wherein, Figure 5 (a) in the figure is a curve comparing the predicted and actual load values ​​under FGSM counter-attacks; Figure 6 (b) in the figure is a curve comparing the predicted and actual load values ​​under PGD counter-attacks; Figure 6 This is an example of the invention comparing the robustness indicators of the January dataset of the power system; wherein, Figure 6 (a) in the figure is a comparison chart of load prediction MAE under FGSM adversarial attacks; Figure 7 (b) in the figure is a comparison chart of load prediction RMSE under FGSM adversarial attacks; Figure 7 (c) in the figure is a comparison chart of MAPE load prediction under FGSM counterattacks; Figure 7 This is an example of the invention comparing the robustness indicators of the January dataset of the power system; wherein, Figure 7 (a) in the figure is a comparison chart of the predicted MAE under PGD counterattacks; Figure 8 (b) in the figure is a comparison chart of load prediction RMSE under PGD adversarial attacks; Figure 8 (c) in the figure is a comparison chart of MAPE load prediction under PGD counterattack; Figure 8 This is a schematic diagram of a multi-objective robust depth prediction system for power load defense against attacks, as described in this invention. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0019] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0020] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0021] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0022] See Figure 8 The present invention provides a robust multi-objective depth prediction method for power load defense against attacks, which specifically includes the following steps: S1. Construct normal training and normal test datasets based on the historical power system load database, and construct a real-time dataset based on the real-time power system load database. For the existing deep learning prediction model for power load (CM), simulate adversarial attack training datasets for power load prediction using the Fast Gradient Sign Method (FGSM) and Projection Gradient Descent (PGD) attacks respectively. D adv_tr and adversarial attack test dataset D adv_te .

[0023] Specifically, based on the historical power system load database, historical data related to power system load are extracted and parsed to obtain a normal dataset, which is then divided into a normal training dataset. D tr and normal test dataset D te Based on the real-time power system load database, real-time data related to power system load are extracted and parsed to obtain a real-time dataset. D t .

[0024] Will Dtr The elements in the matrix are labeled as a m,h , a m,h This represents the first [item] in the normal training dataset. m The first sample h eigenvalues, of which m express D tr The row number of the matrix corresponds to the row number in the normal training dataset. m samples, 1≤ m ≤ M , M This represents the maximum number of samples in the normal training dataset. h express D tr , D te and D t The column indices of the matrix correspond to the first, second, and third columns of the normal training dataset, normal test dataset, and real-time dataset, respectively. h Features, 1≤ h ≤ H , H This represents the maximum number of features in the normal training dataset, normal test dataset, and real-time dataset; D te The elements in the matrix are labeled as a w,h , a w,h This represents the first [item] in the normal test dataset. w The first sample h eigenvalues, of which w express D te The row number of the matrix corresponds to the row number in the normal test dataset. w samples, 1≤ w ≤ W , W This represents the maximum number of samples in the normal test dataset; D t The elements in the matrix are labeled as a rt,h , a rt,h Represents the first in the real-time dataset rt The first sample h eigenvalues, of which rt express D t The row number in the matrix corresponds to the row number in the real-time dataset. rt samples, 1≤ r t ≤n rt , n rt This represents the maximum number of samples in the real-time dataset.

[0025] To address the existing deep learning-based power load prediction model CM, we simulate and generate adversarial attack training datasets for power load prediction using the fast gradient sign method and projective gradient descent attacks, respectively. D adv_tr and adversarial attack test dataset D adv_te .

[0026] The CM model architecture consists of a stacked structure of a Convolutional Neural Network (CNN) module, a global average pooling layer, a Bidirectional Long Short-Term Memory Network (BiLSTM) layer, and fully connected layers. The CNN module includes convolutional layers and pooling layers. The CM model parameters include { Figure 9 1, Figure 1 2, ppt 3 ppt 4}, ppt 1 represents the set of parameters for the number of modules in the convolutional neural network and the number of layers in the bidirectional long short-term memory network. , ppt 2 represents the parameter set of the convolutional neural network module. ppt 3 represents the parameter set of the bidirectional long short-term memory network layer. ppt 4 represents the set of configuration parameters for the model training hyperparameters; ppt 1={ C 1, B 1}, where C 1 represents the number of modules in the convolutional neural network. B 1 represents the number of layers in the bidirectional long short-term memory network; ,in k 0 indicates the first k 0 convolutional neural network modules For the first k The number of convolutional filters in the convolutional layer of a 0-convolutional neural network module. For the first k The kernel size of a convolutional layer with 0 convolutional neural network modules. For the first k The activation function type of the convolutional layer in a convolutional neural network module with 0 modules. For the first k The pooling type used after the convolutional layers of 0 convolutional neural network modules. For the first kThe pooling size used after the convolutional layers of 0 convolutional neural network modules; ,in j 0 indicates the first j 0 bidirectional long short-term memory network layers, For the first j The number of neurons in 0 bidirectional long short-term memory network layers For the first j Activation function types for 0 bidirectional long short-term memory network layers; ppt 4={ ppt c , dr c , lr c , ppt c , bt c},in opt c For CM training batch size, dr c This refers to the dropout rate during CM training. lr c For CM training learning rate, ret c The type of optimizer used for CM training. bt c The regularization type used for CM training.

[0027] For example, this embodiment describes the structure and parameters of the power load forecasting model CM deployed on the example dataset as follows: opt As shown, its model parameters specifically include { ret 1={2,1}, Figure 2 2={14, 7, sigmoid, MaxPooling, 5; 8,3, relu, MaxPooling, 5}, ppt 3 = {30, tanh} pptThe expression `4={92, 0.2, 0.0154, adamax,None}` indicates that CM has two convolutional neural network modules and one bidirectional long short-term memory network layer. The first convolutional neural network module has 14 convolutional filters, a kernel size of 7, and uses a sigmoid activation function. It then employs max pooling with a pooling size of 5. The second convolutional neural network module has 8 convolutional filters and a kernel size of 3. The activation function type of the first convolutional neural network module is ReLU (Rectified Linear Unit Activation). The pooling type used after the second convolutional neural network module is MaxPooling, and the pooling size used after the second convolutional neural network module is 5. The number of neurons in the first bidirectional long short-term memory network layer is 30, and the activation function type of the first bidirectional long short-term memory network layer is tanh (Hyperbolic Tangent Activation). The dropout rate during training is 0.2, the batch size is 92, the training learning rate is 0.0154, the optimizer type used for training is Adamax (Adaptive Moment Estimator with Infinite Norm), and None is used during training to indicate that no regularization is applied.

[0028] Adversarial attack training dataset D adv_tr and adversarial attack test dataset D adv_te Specifically, it is obtained through the following sub-steps: S11. When the adversarial attack method is FGSM adversarial attack, the adversarial attack training dataset is... D adv_tr and adversarial attack test dataset D adv_te The simulation generation process is shown in S11.1 to S11.4: S11.1 Initialize FGSM counter-attack perturbation strength ppt 1. Generate FGSM adversarial attack samples according to formula (1). X advFGSM : (1) In the formula, x Represents the original sample; d x Represents the original input sample x The total number of feature dimensions; Indicates the original sample x The partial derivatives; Indicates sample x China's power loadi The true values ​​of 0 features; Indicates the first i 0 feature loss functions , ppt For the network parameters of the CM model; sign represents the gradient calculation symbol; This indicates that the generated adversarial perturbation is constrained within the range [0,1].

[0029] S11.2, When the FGSM adversarial attack training dataset D adv_tr_FGSM When generating adversarial examples, set x = x adv_tr_FGSM,tx ,1≤ ε ≤ Tr ,in x adv_tr_FGSM,tx This represents the FGSM adversarial attack training dataset. D adv_tr_FGSM The first in θ One original sample, Tr This indicates that from the normal training dataset D tr Repeat the FGSM adversarial example generation process in step S11.1 above, without repeating the randomly selected number of samples, until the FGSM adversarial attack training dataset is obtained. D adv_tr_FGSM All original samples x adv_tr_FGSM,tx All were replaced with FGSM adversarial samples calculated by formula (1). X adv_tr_FGSM,tx ,in X adv_tr_FGSM,tx express D adv_tr_FGSM The simulated generation of the first tx One FGSM adversarial sample.

[0030] S11.3, When using the FGSM adversarial attack test dataset D adv_te_FGSM When generating adversarial examples, set x = x adv_te_FGSM,ty ,1≤ tx ≤ tx ,in, x adv_te_FGSM,ty This represents the FGSM adversarial attack test dataset D. adv_te_FGSM The first in ty One original sample, Te This indicates the normal test dataset. D teRepeat the FGSM adversarial example generation process in step S11.1 above, without repeating the randomly selected number of samples, until the FGSM adversarial attack test dataset is obtained. D adv_te_FGSM All original samples x adv_te,ty All were replaced with adversarial samples calculated using formula (1). X adv_te_FGSM,ty ,in X adv_te_FGSM,ty for D adv_te_FGSM The simulated generation of the first ty An adversarial example.

[0031] S11.4 The generated FGSM adversarial attack training dataset D adv_tr_FGSM For normal training dataset D tr Replace the sample at the corresponding position in the middle, and then complete the above replacement. D tr The dataset is labeled as an adversarial attack training dataset. D adv_tr The generated FGSM adversarial attack test dataset D adv_te_FGSM For normal test dataset D te Replace the sample at the corresponding position in the middle, and then complete the above replacement. D te Dataset labeling adversarial attack test dataset D adv_te .

[0032] S12. When the adversarial attack method is PGD adversarial attack, the adversarial attack training dataset is... D adv_tr and adversarial attack test dataset D te The simulation generation process is shown in S12.1 to S12.7: S12.1 Initialize PGD counter-attack perturbation strength Te 2. PGD Counterattack Step Size α 1. Number of iterations against PGD counterattacks ty 1. Generate initial values ​​for PGD to counteract disturbances based on CM , This represents the initial perturbation matrix with the same shape as the original sample, and records the number of PGD adversarial generation iterations. t =0.

[0033] S12.2. According to formula (2), iteratively generate the first... t+1 PGD counter-perturbation And apply perturbation constraints: (2) In the formula, x Represents the original sample; This indicates that the generated adversarial perturbation is performed within the range of [- ε 2, ζ Constraints of 2]; Indicates the iterative generation of the first... t Secondary PGD perturbation, 0≤ t < ε 1; Indicates resistance to disturbance The partial derivatives; y true This represents the actual value of the sample power load; L ( ε ( x + ), y true ) represents the loss function. ζ For the network parameters of the CM model; sign represents the gradient calculation symbol.

[0034] S12.3. According to formula (3), generate the iteration number. t +1 PGD adversarial examples that satisfy the range constraint : (3) In the formula, This indicates that the generated PGD adversarial examples are constrained to have a value range of [0,1]. The number of PGD adversarial generation iterations is also recorded. t = t +1.

[0035] The number of iterations for generating adversarial examples using PGD reached [number]. θ The adversarial examples generated after step 1 are X advPGD In the PGD adversarial attack training dataset D adv_tr_PGD When generating initial adversarial examples, x = x adv_tr_PGD,tx ,1≤ θ ≤ Tr ,in x adv_tr_PGD,tx This represents the training dataset for PGD adversarial attacks. D adv_tr_PGD The first in ζ One original sample, Tr This represents the training dataset for PGD adversarial attacks. Dadv_tr_PGD The total number of all samples in the PGD adversarial attack test dataset; D adv_te_PGD When generating initial adversarial examples, x = x adv_te_PGD,ty ,1≤ tx ≤ tx ,in ty This represents the PGD adversarial attack test dataset. D adv_te_PGD The first in Te One original sample, ty This represents the PGD adversarial attack test dataset. D adv_te_PGD The total number of all samples.

[0036] S12.4, The maximum number of iterations for generating PGD adversarial examples reaches [a certain number]. ty The adversarial examples generated after step 1 are labeled as follows: X advPGD .

[0037] S12.5, In the PGD adversarial attack training dataset D adv_tr_PGD When generating PGD adversarial examples, set x = x adv_tr_PGD,tx ,1≤ Te ≤ Tr ,in x adv_tr_PGD,tx This represents the training dataset for PGD adversarial attacks. D adv_tr_PGD The first in ζ For each original sample, repeat steps S12.1 to S12.4 of the adversarial sample generation process until the PGD adversarial attack training dataset is obtained. D adv_tr_PGD All original samples x adv_tr_PGD,tx All were replaced with PGD adversarial examples. X adv_tr_PGD,tx ,in X adv_tr_PGD,tx express D adv_tr_PGD The Middle tx One PGD adversarial sample.

[0038] S12.6, On the PGD adversarial attack test dataset D adv_te_PGD When generating PGD adversarial examples, set x = x adv_te_PGD,ty ,1≤ tx ≤tx ,in, ty This represents the PGD adversarial attack test dataset. D adv_te_PGD The first in Te For each original sample, repeat steps S12.1 to S12.4 of the adversarial sample generation process until an adversarial attack test dataset is generated. D adv_te_PGD All original samples x adv_te,ty All PGD adversarial samples were replaced. X adv_te_PGD,ty ,in X adv_te_PGD,ty express D adv_te_PGD The Middle ty One PGD adversarial sample.

[0039] S12.7 The generated PGD adversarial attack training dataset D adv_tr_PGD For normal training dataset D tr Replace the sample at the corresponding position in the middle, and then complete the replacement. D tr The dataset is labeled as an adversarial attack training dataset. D adv_tr The generated PGD adversarial attack test dataset D adv_te_PGD For normal test dataset D te Replace the sample at the corresponding position in the middle, and then complete the replacement. D te The dataset is labeled as an adversarial attack test dataset. D adv_te .

[0040] It should be understood that, M , ty rt and H This can be determined based on the actual dataset. For example, in this embodiment, the electricity load demand dataset from the Australian Energy Market Operator (AEMO) is used. The implementation example is based on the 2018 New South Wales (N...) data. S The dataset used in W) is the January dataset. A day consists of 24 hours, and one data point is collected every half hour, so each day in the dataset contains 48 data points; on the other hand, the data for each month in the dataset used in this embodiment is treated as an independent dataset. Each dataset is divided into two separate sets, including a training set and a test set. Therefore, for a 30-day month, there are There are 100 data points, with 75% of the early data points in each dataset serving as the training set and the remaining data points serving as the test set. H =48, M = , ty rt = .

[0041] Exemplarily, in this embodiment, Tr =192, W, n =48, W = n 1 = 0.1 Te 2 = 0.05, α 1 = 0.03 ε 1=50. It should be understood that the generated adversarial training and test datasets are injected with adversarial sample data by replacing the original datasets, such as... ε The diagram shown is a schematic of the adversarial sample generation and prediction process in the multi-objective robust depth prediction method for power load defense against adversarial attacks of the present invention.

[0042] S2. Set the first parameter and construct the multi-objective optimization objective function of the hybrid prediction model of power load multi-objective adversarial robust convolutional neural network (CNN) and bidirectional long short-term memory network (BiLSTM) (CNN-BiLSTM). The multi-objective optimization objective function includes the prediction mean square error of the power load adversarial robust CNN-BiLSTM prediction model on normal samples. f 1. Prediction mean square error on adversarial examples f 2 and the number of model parameters f 3.

[0043] Furthermore, the first parameter includes the maximum number of offline iterations. G Maximum number of solutions in the archive A max Adversarial training threshold T .

[0044] Furthermore, the mean squared error of the power load robust CNN-BiLSTM prediction model on normal samples f 1. Prediction mean square error on adversarial examples f 2 and the number of model parameters f 3. As shown in formulas (4) to (6): (4) (5) (6) In the formula, y w This represents the actual value of the power load. This represents the predicted value of a normal sample of power load. This represents the predicted value of the power load countermeasure sample; Indicates the first l The number of parameters in a CNN module C Indicates the number of CNN modules; Indicates the first k The number of parameters in a BiLSTM layer. B Indicates the number of BiLSTM layers; P Dense This indicates the number of parameters in a fully connected network layer.

[0045] It should be noted that the first parameter can be set according to the actual situation. For example, in this embodiment, the first parameter is set to... G =30、 A max =50、 T =0.01.

[0046] S3. Set the second parameter, construct a multi-objective adversarial robust CNN-BiLSTM prediction code for power load, and generate individuals to be mutated. p g And set the current iteration number. g =0.

[0047] Furthermore, the second parameter includes a preset maximum number of CNN modules. C max and minimum value C min Maximum number of BiLSTM layers B max and minimum value B min Maximum number of convolutional kernels in a CNN module ζ max and minimum value Figure 3 min , No. l Minimum number of convolutional kernels in a CNN module fn min and maximum value fn max , No. l Minimum kernel size of a CNN module fn min and maximum value fn max , No. l The activation function of each CNN module encodes the minimum value. kn min and maximum value kn max , No. lThe minimum value is encoded by pooling type of each CNN module. an min and maximum value an max , No. l The minimum value is encoded by pooling type of each CNN module. pt min and maximum value pt max , No. k Minimum number of neurons in a layer BiLSTM ps min and maximum value ps max , No. k Layer BiLSTM activation function encodes minimum value nb min and maximum value nb max Minimum batch size ab min and maximum value ab max Minimum value of optimizer encoding bt min and maximum value bt max Minimum learning rate lr min and maximum value lr max The minimum value encoded by regularization opt min and maximum value opt max Minimum discard rate dr min and maximum value dr max .

[0048] Furthermore, a multi-objective adversarial robust CNN-BiLSTM predictive encoding for power load is constructed to generate individuals to be mutated. p g Specifically, this includes: encoding the parameters of the CNN-BiLSTM model to be optimized based on the second parameter. The encoding of each CNN-BiLSTM model to be optimized includes the following four parts: ret 1={ C , B}, ret 2={ pt 2,l = ( pt l , pt l , fn l , knl , an l ), l =1, 2, …, C}, pt 3={ ps 3,k =( pt k , pt k ), k =1, 2, …, B}, nb 4=( ab , pt , lr , bt , dr ), opt 1 represents the set of numbers of CNN modules and BiLSTM layers. ret 2 represents the set of neural network parameters for the CNN module. pt 3 represents the parameter set of the BiLSTM layer neural network. pt 4 represents the set of hyperparameters used in model training; C This indicates the number of CNN modules in the CNN-BiLSTM model to be optimized. B This indicates the number of BiLSTM layers in the CNN-BiLSTM model to be optimized; pt 2,l Indicates the first l CNN module parameter encoding pt l Indicates the first l Number of convolutional kernels per CNN module pt l Indicates the first l The size of the convolutional kernel in each CNN module. fn l Indicates the first l Activation functions for each CNN module kn l Indicates the first l Each CNN module pooling layer type an l Indicates the first l The pooling kernel size for each CNN module; pt 3,k Indicates the first k Layer BiLSTM parameter encoding, ps k Indicates the first k Number of BiLSTM neurons per layer pt k Indicates the first k Layer BiLSTM activation function;nb This indicates the batch size for training the CNN-BiLSTM model. ab This represents the training optimizer for the CNN-BiLSTM model. lr This represents the learning rate used to train the CNN-BiLSTM model. bt This indicates the regularization method used during the training of the CNN-BiLSTM model. dr This represents the dropout rate during the training of the CNN-BiLSTM model.

[0049] in, opt 1. ret 2,l , pt 3,k , pt The specific values ​​of 4 are shown in formulas (7) to (10): (7) In the formula, Randint represents the random integer selection function. C min This represents the minimum value of the number of CNN modules. C max This represents the maximum value that can be taken for the number of CNN modules. C min ≤ C ≤ C max ; B min This represents the minimum value for the number of BiLSTM layers. B max This represents the maximum value of the BiLSTM layer number. B min ≤ B ≤ B max .

[0050] (8) In the formula, pt min Indicates the first l Minimum number of convolutional kernels for a CNN module pt max Indicates the first l The maximum number of convolutional kernels in a CNN module. fn min ≤ fn l ≤ fn max ; fn min Indicates the first l The minimum kernel size for each CNN module. fnmax Indicates the first l The maximum value of the convolutional kernel size of each CNN module. kn min ≤ kn l ≤ kn max ; kn min Indicates the first l The minimum value encoded by the activation function of each CNN module. kn max Indicates the first l The maximum value encoded by the activation function of each CNN module. an min ≤ an l ≤ an max ; [[ID=32......]] min Indicates the first l The minimum value of pooling type encoding for each CNN module. an max Indicates the first l The maximum value of the pooling type encoding of each CNN module. an min ≤ pt l ≤ pt max ; pt min Indicates the first l The minimum value of pooling type encoding for each CNN module. pt max Indicates the first l The maximum value of the pooling type encoding of each CNN module. pt min ≤ ps l ≤ ps max .

[0051] (9) In the formula, ps min Indicates the first k Minimum number of neurons in a layer of BiLSTM ps max Indicates the first k The maximum number of neurons in a layer of BiLSTM. ps min ≤ nb k ≤ nb max ; nbmin Indicates the first k The minimum value encoded by the activation function of the BiLSTM layer. nb max Indicates the first k The maximum value encoded by the activation function of the BiLSTM layer. nb min ≤ ab k ≤ [[ID=......]] max .

[0052] (10) In the formula, Random represents the function for selecting random decimals; ab min This represents the minimum batch size. ab max This represents the maximum value of the batch size. ab min ≤ ab ≤ bt max ; [[ID min This represents the minimum value encoded by the optimizer. ​ max This represents the maximum value encoded by the optimizer. ​ min ≤ ​ ≤ ​ max ; lr min This represents the minimum learning rate. lr max This represents the maximum value of the learning rate. lr min ≤ lr ≤ lr max ; ​ min This represents the minimum value encoded using regularization. ​ max This represents the maximum value encoded using regularization. ​ min ≤ ​ ≤ ​ max ; dr min This represents the minimum discard rate. dr max This represents the maximum discard rate. dr min ≤ dr ≤ dr max .

[0053] Based on formulas (7) to (10), individuals to be mutated are randomly generated. p g Its code is ,in g This represents the current iteration number. ​ 2,1 This indicates the parameter encoding of the first CNN module. Indicates the first C CNN module parameter encoding ​ 3,1 This indicates the encoding of the first layer BiLSTM parameters. Indicates the first B Layer BiLSTM parameter encoding.

[0054] For example, in this implementation, the ranges of the various encoding parameters are set as follows: C min =1, C max =5; B min =1, B max =3; ​ min =8, ​ max =96; ​ min =0, ​ max =3; ​ min =0, ​ max =2; ​ min =0, ​ max =1; ​ min =1, ​ max =3; ​ min =30, ​ max =110; ​ min =0, ​ max =1; ​ min =40, ​ max =100; ​ min =0, ​ max =3; lrmin =0.001, lr max =0.04; ​ min =0, ​ max =3; dr min =0.2, dr max =0.5. Wherein, ​ min =0, ​ max =3, that is ​ The four optimizers corresponding to the values ​​0, 1, 2 and 3 are: Adaptive Moment Estimator (adam), Root Mean Square Propagation Optimizer (rmsprop), Infinite Norm Adaptive Moment Estimator (adamax), and Adaptive Incremental Optimizer (adadelta). ​ min =0, ​ max =3, that is ​ The four regularization methods corresponding to the values ​​0, 1, 2, and 3 are: None, L1 regularization, L2 regularization, and L1L2 regularization, respectively. ​ min =0, ​ max =2, that is ​ l The values ​​corresponding to the first, second, and third times respectively when they are 0, 1, and 2 l The three activation functions for each CNN module are: ReLU (rectified linear unit activation function), Leaky ReLU (rectified linear unit activation function with leakage), and Swish (self-gated activation function). ​ min =0, ​ max =1, that is ​ k The values ​​corresponding to the first and second digits when they are 0 and 1 respectively k The two activation functions for the BiLSTM layer are the hyperbolic tangent activation function tanh and the scaling exponential linear unit activation function selu.

[0055] For example, the individual to be mutated p g = The solution {1,2,12,1,0,1,1,40,0,60,1,50,1,0.02,1,0.35} is encoded as follows: ​ As shown.

[0056] S4. Dealing with mutant individuals pg Perform mutation operations to generate a population. P g Assess the population P g medium-sized individuals optimization objectives f 1 and f 3, and determine f Is 1 less than or equal to the adversarial training threshold? T If not, set f 2=1, interrupting the individual If so, proceed to step S6 for individual adversarial training; Perform adversarial training and calculate the optimal objective. f 2.

[0057] S41. Treat the variant individuals according to formulas (7) to (10). p g Each gene locus is mutated sequentially to generate a scale of N g population ,in, It is a population P g The first individual, Indicates population P g The i Individual, Indicates population P g The N g Individual, , C g yes The number of CNN modules, B g yes The number of BiLSTM layers; ,in, Individual The corresponding ​ 1, Represents an individual The corresponding ​ 2, Represents an individual The corresponding ​ 3, Represents an individual The corresponding ​ 4, express The parameters of the first dimension, express No. dg Dimensional parameters, express No. D g Dimensional parameters, D g express Maximum dimension .

[0058] During initialization, i.e., when the number of iterations... g When =0, initialize the individual according to formulas (7) to (10). p 0 Each gene locus is mutated sequentially to generate a scale of N Initial population of 0 ,in It is the first individual in the initial population. Denotes the initial population's... i Individual, Denotes the initial population's... N 0 individuals , C 0 is The number of CNN modules, B 0 is The number of BiLSTM layers; ,in Individual of ​ 1, Represents an individual of ​ 2, Represents an individual of ​ 3, Represents an individual of ​ 4, express The parameters of the first dimension, express No. d 0-dimensional parameters express No. D 0-dimensional parameters D 0 represents Maximum dimension .

[0059] S42, Based on normal training dataset D tr right P g medium-sized individuals The CNN-BiLSTM prediction model was trained based on the normal test dataset. D te Calculate individuals The corresponding optimization objective of the prediction model after the above training f 1 and f 3. Evaluation f Is 1 less than or equal to the adversarial training threshold? T If not, set optimization goals. f 2=1, interrupting the individual The confrontation training; if so, proceed to the next step. S 6. For individuals Perform adversarial training and calculate the optimal objective. f 2.

[0060] For example, in this embodiment, the individual to be mutated p g = The mutation {1,2,12,1,0,1,1,40,0,60,1,50,1,0.02,1,0.35} will generate a population of 16 individuals. It should be understood that each mutation only changes one dimension of the individual to be mutated. If the mutation occurs at the layer number of the individual to be mutated, the parameter settings of each layer of the neural network will be increased or truncated to ensure that the number of layers corresponds to the parameter settings of each layer.

[0061] S5, for population P g All individuals according to the optimization objective f 1. f 2 and f 3. Perform non-dominated sorting and store non-dominated solutions in an external archive solution set. S In, and based on external archive update technology S Update and obtain the individuals to be mutated for the next optimization iteration. p g Number of update iterations g = g +1.

[0062] S51, Based on the optimization objective f 1. f 2 and f 3 pairs of populations P g Perform a non-dominated ranking on all individuals, and check each individual one by one whether it is completely dominated by other individuals. That is, if there is an individual... exist f 1. f 2 and f All 3 are smaller than the individual This indicates that the individual Complete domination of individuals Then delete the controlled individual. Individuals who are not completely dominated retain their individuality. Mark all retained individuals as the retained solution set. U ,in, express P g China differs from individuals, that is i ≠ j .

[0063] S52, Solution set U External archive deset S Merging is marked as a solution set Q According to the optimization objective f 1. f 2 and f 3 pairs Q All solutions are sorted in a non-dominated order, and each solution is evaluated to determine whether it is completely dominated by other solutions. That is, if a solution exists... exist f 1. f 2 and f All 3 are less than the solution This explains the solution. Complete Domination Solution Then delete the dominated solution. If not fully dominated, the solution is retained. Mark all retained solutions as the new external archive solution set. S That is, updating the external archive solution set. S ,in, and Both represent solution sets Q The solution in, and the solution Different from the solution ,Right now i 1≠ j 1.

[0064] S53, Judgment Is it greater than the number of external archives? A max If not, proceed to step S54; if yes, then deset the external archive. S Crowding distance is calculated and sorted for each non-dominated solution, and redundant solutions are removed; Represents external archive solution set S The number of solutions dominated by non-dominated solutions.

[0065] Furthermore, the steps for calculating the congestion distance include: S53.1, Decomposing external archives SThe solutions are sorted in ascending order according to a certain optimization objective value, and the minimum value of the sorted solutions in terms of that objective value is marked. f min and maximum value f max Set the crowding distance of the target that is the minimum and maximum solution after sorting to infinity.

[0066] S53.2, For each optimization objective, sort a non-dominated solution in ascending order. , Represents external archive solution set S Sort in ascending order i Two solutions are available. The congestion distance is calculated according to formula (11). d s Then, based on the distance of the crowding level d s External archive deset S Sort all non-dominated solutions in descending order: (11) In the formula, and These represent external archive solution sets. S Non-dominated solutions sorted in ascending order In optimizing objectives The optimization objective value of the adjacent solutions above, and They represent the optimization objectives for ascending order sorting, respectively. The maximum and minimum values ​​on.

[0067] S53.3, Complete the external archive decomposition. S After calculating the crowding distance of all non-dominated solutions and sorting them in descending order, delete them. t A solution with a small crowding distance t = – T .

[0068] S54, External Archive Decomposition S Randomly select a non-dominated solution and mark it as the individual to be mutated in the next optimization iteration. p g Number of update iterations g = g +1.

[0069] S6, based on D adv_tr For individuals Conduct adversarial training, and then based on D adv_te Calculate the optimization objective f 2.

[0070] Specifically, based on the adversarial attack training dataset D adv_tr For individuals The CNN-BiLSTM neural network model was subjected to adversarial training and tested on an adversarial attack dataset. D adv_te Calculate the individual according to formula (5) The corresponding optimization objective of the prediction model after adversarial training f 2.

[0071] S7, when the number of iterations g Less than the set maximum number of iterations G When, repeat steps S4~S6; when g Greater than or equal to G At that time, mark the external archive solution set. S All solutions are Pareto front solutions. CNN-BiLSTM, which balances high prediction accuracy, strong adversarial robustness, and lightweight design, is selected and marked as the optimal model for multi-objective adversarial robust deep prediction of power load (BM).

[0072] For example, Pareto fronts obtained for different adversarial attacks on the example dataset are as follows: ​ As shown; where, ​ (a) in the figure is the Pareto front obtained against FGSM adversarial attacks; ​ (b) in the figure shows the Pareto front obtained against PGD adversarial attacks. The optimal CNN-BiLSTM power load forecasting model BM, which balances adversarial defense and lightweight design, was selected from the Pareto front.

[0073] S8. Deploy model BM online in the power system, targeting the real-time dataset in the power system load real-time database. D t Perform online power system load forecasting and test and evaluate its robustness performance indicators.

[0074] Specifically, BM will be deployed online in the power system, targeting real-time datasets. D t Online power load forecasting is performed, and its robustness performance is evaluated. The performance evaluation indicators include root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), and their calculation formulas are shown in formulas (12) to (14), respectively. (12) (13) (14) In the formula, express D t No. rt The actual power load value of each test sample. Indicates BM's target D t The Middle rt The predicted power load for each test sample.

[0075] For example, in this embodiment, the comparison curves of the predicted and actual load values ​​of the existing technology CM and the present invention technology BM under the adversarial attack on the example dataset are shown below. ​ As shown; where, ​ (a) in the figure is a curve comparing the predicted and actual load values ​​under FGSM counter-attacks; ​ (b) in the figure is a curve comparing the predicted and actual load values ​​under PGD adversarial attacks. A comparison chart of adversarial robustness indices for the January power system dataset is shown below. ​ and ​ As shown; where, ​ (a) in the figure is a comparison chart of load prediction MAE under FGSM adversarial attacks; ​ (b) in the figure is a comparison chart of load prediction RMSE under FGSM adversarial attacks; ​ (c) in the figure is a comparison chart of MAPE load prediction under FGSM counterattacks; ​ (a) in the figure is a comparison chart of the predicted MAE under PGD counterattacks; ​ (b) in the figure is a comparison chart of load prediction RMSE under PGD adversarial attacks; ​ (c) in the figure is a comparison chart of MAPE load prediction under PGD counterattack. ​ This fully demonstrates that the BM, trained for adversarial defense, exhibits higher robustness under various adversarial attacks, effectively controlling prediction errors to an extremely low level. For example, ​ In (a), the deployed CM has a MAE of 0.014 when not under attack, but the MAE spikes to 0.598 when under PGD attack. After countermeasures, the MAE is 0.030, which still shows a large error compared to the unattacked case. In contrast, the BM obtained by the technology of this invention has a MAE of 0.019 after countermeasures, which controls the error to an extremely low level. Other robustness indicators and robustness indicators under FGSM attack also show the same pattern.

[0076] In summary, through the above technical solutions, this invention proposes a multi-objective adversarial robust automatic deep load forecasting method that balances high-precision prediction performance on clean samples, strong adversarial robustness, and lightweight model design. It simulates and generates adversarial samples for power load forecasting based on PGD and FGSM adversarial attacks, using the parameters of the CNN-BiLSTM model architecture to be designed and the training hyperparameters as optimization variables. During optimization, models with power load forecast MSE below a set threshold undergo adversarial training. The three optimization objective functions are minimizing the model MSE on clean samples, minimizing the MSE after adversarial training, and minimizing the number of model parameters. A multi-objective variable-length discrete extremum optimization technique is designed for offline multi-objective iterative optimization, thereby obtaining a compromise power load forecasting model that achieves "high prediction accuracy, strong adversarial robustness, and lightweight design." Furthermore, this invention expands the application of multi-objective automatic optimization technology in power systems, providing a power load forecasting solution that balances high accuracy, high efficiency, and multiple adversarial defenses. Through steps S1-S8 and specific parameter settings and adjustments during the optimization process, the optimal balance is ultimately achieved.

[0077] It is worth mentioning that the present invention also provides a robust depth prediction system for multi-objective power load to defend against counterattacks, which is used to implement the robust depth prediction method for multi-objective power load to defend against counterattacks in the above embodiments.

[0078] See ​ The system includes a normal dataset and adversarial sample dataset construction module 10, a parameter setting and objective function optimization construction module 20, an optimization model encoding definition module 30, a population generation based on mutation operation and a judgment module for whether to enter adversarial training 40, a multi-objective non-dominated solution evaluation and external archive update module 50, an adversarial training module 60, an iteration judgment module 70, and an online deployment module 80.

[0079] In this embodiment, the normal dataset and adversarial example dataset construction module 10 is used to construct a normal training dataset and a normal test dataset based on the historical power system load database, and to construct a real-time dataset based on the real-time power system load database. Furthermore, for the existing deep learning prediction model CM for power load, the module simulates and generates power load prediction adversarial attack training datasets using the fast gradient sign method and projective gradient descent attack, respectively. D adv_tr and adversarial attack test dataset D adv_te .

[0080] In this embodiment, the parameter setting and objective function construction module 20 is used to set the first parameter and construct a multi-objective optimization objective function for the multi-objective adversarial robust CNN-BiLSTM prediction model for power load. The multi-objective optimization objective function includes the mean square error of the CNN-BiLSTM prediction model on normal samples. f 1. Prediction mean square error on adversarial examples f 2 and the number of model parameters f 3.

[0081] In this embodiment, the optimized model encoding definition module 30 is used to set the second parameter, construct a multi-objective adversarial robust CNN-BiLSTM prediction encoding for power load, and generate individuals to be mutated. p g And set the current iteration number. g =0.

[0082] In this embodiment, the population generation and adversarial training determination module 40 based on mutation operations is used to treat mutated individuals. p g Perform mutation operations to generate a population. P g Assess the population P g medium-sized individuals optimization objectives f 1 and f 3, and determine f Is 1 less than or equal to the adversarial training threshold? T If not, set f 2=1, interrupting the individual If so, proceed to the adversarial training module for individual training. Conduct adversarial training and calculate the optimal objective. f 2.

[0083] In this embodiment, the multi-objective non-dominated solution evaluation and external archive update module 50 is used for population... P g All individuals according to the optimization objective f 1. f 2 and f 3. Perform non-dominated sorting and store non-dominated solutions in an external archive solution set. S In, and based on external archive update technology S Update and obtain the individuals to be mutated for the next optimization iteration. p g Number of update iterations g = g +1.

[0084] In this embodiment, the adversarial training module 60 is used for training based on...D adv_tr For individuals Conduct adversarial training, and based on D adv_te Calculate the optimization objective of an individual f 2.

[0085] In this embodiment, the iteration judgment module 70 is used to determine the iteration number. g Less than the set maximum number of iterations G At that time, the system repeatedly enters the population generation and adversarial training decision module based on mutation operations, the multi-objective non-dominated solution evaluation and external archive update module, and the adversarial training module; when g Greater than or equal to G At that time, mark the external archive solution set. S All solutions are Pareto front solutions. CNN-BiLSTM, which balances high prediction accuracy, strong adversarial robustness, and lightweight design, is selected and marked as the optimal model for multi-objective adversarial robust deep prediction of power load (BM).

[0086] In this embodiment, the online deployment module 80 is used to deploy the selected model BM online in the power system, targeting the real-time dataset of the power system load real-time database. D t Perform online power system load forecasting and test and evaluate its robustness performance indicators.

[0087] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes: ROM. 、 RAM, magnetic disks, optical disks, and other media that can store program code.

[0088] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of multi-objective robust deep forecasting of power loads defending against adversarial attacks, characterized in that, Includes the following steps: S1. Construct normal training and normal test datasets based on the historical power system load database, and construct a real-time dataset based on the real-time power system load database. For the existing deep learning prediction model CM for power load, simulate and generate power load prediction adversarial attack training datasets using the fast gradient sign method and projective gradient descent attack, respectively. D adv_tr and adversarial attack test dataset D adv_te ; S2. Set the first parameter and construct the multi-objective optimization objective function of the hybrid prediction model CNN-BiLSTM, which combines a multi-objective adversarial robust convolutional neural network and a bidirectional long short-term memory network for power load. The multi-objective optimization objective function includes the mean squared error of the prediction of the power load adversarial robust CNN-BiLSTM prediction model on normal samples. f 1. Mean squared error of prediction on adversarial examples f 2 and the number of model parameters f 3; S3, set the second parameter, construct power load multi-objective confrontation robust CNN-BiLSTM prediction code, generate individual to be mutated p g And set the current iteration number g =0; S4. Dealing with mutant individuals p g Perform mutation operations to generate a population. P g Assess the population P g medium-sized individuals optimization objectives f 1 and f 3, and determine f Is 1 less than or equal to the adversarial training threshold? T If not, set f 2=1, interrupting the individual If so, proceed to step S6 for individual adversarial training; Perform adversarial training and calculate the optimal objective. f 2; S5, for population P g All individuals according to the optimization objective f 1. f 2 and f 3. Perform non-dominated sorting and store non-dominated solutions in an external archive solution set. S In, and based on external archive update technology S Update and obtain the individuals to be mutated for the next optimization iteration. p g Number of update iterations g = g +1; S6、based on D adv_tr for the individual performing adversarial training, and then based on D adv_te computing an optimization objective f 2; S7, when the number of iterations g Less than the set maximum number of iterations G When, repeat steps S4~S6; when g Greater than or equal to G At that time, mark the external archive solution set. S All solutions are Pareto front solutions. CNN-BiLSTM, which balances high prediction accuracy, strong adversarial robustness, and lightweight design, is selected and marked as the optimal model for multi-objective adversarial robust deep prediction of power load (BM). S8. Deploy model BM online in the power system, targeting the real-time dataset in the power system load real-time database. D t Perform online power system load forecasting and test and evaluate its robustness performance indicators.

2. The power load multi-objective robust deep forecasting method against adversarial attacks according to claim 1, wherein, Step S1 specifically includes: Based on the historical power system load database, historical data related to power system load are extracted and parsed to obtain a normal dataset. This normal dataset is then divided into a normal training dataset. D tr and normal test dataset D te Based on the real-time power system load database, real-time data related to power system load are extracted and parsed to obtain a real-time dataset. D t ; Will D tr The elements in the matrix are labeled as a m,h , a m,h This represents the first [item] in the normal training dataset. m The first sample h eigenvalues, of which m express D tr The row number of the matrix corresponds to the row number in the normal training dataset. m samples, 1≤ m ≤ M , M This represents the maximum number of samples in the normal training dataset. h express D tr , D te and D t The column indices of the matrix correspond to the first, second, and third columns of the normal training dataset, normal test dataset, and real-time dataset, respectively. h Features, 1≤ h ≤ H , H This represents the maximum number of features in the normal training dataset, normal test dataset, and real-time dataset; D te The elements in the matrix are labeled as a w,h , a w,h This represents the first [item] in the normal test dataset. w The first sample h eigenvalues, of which w express D te The row number of the matrix corresponds to the row number in the normal test dataset. w samples, 1≤ w ≤ W , W This represents the maximum number of samples in the normal test dataset; D t The elements in the matrix are labeled as a rt,h , a rt,h Represents the first in the real-time dataset rt The first sample h eigenvalues, of which rt express D t The row number in the matrix corresponds to the row number in the real-time dataset. rt samples, 1≤ r t ≤ n rt , n rt denotes the maximum number of samples of the real-time dataset; To address the existing deep learning-based power load prediction model CM, we simulate and generate adversarial attack training datasets for power load prediction using the fast gradient sign method and projective gradient descent attacks, respectively. D adv_tr and adversarial attack test dataset D adv_te The CM model architecture comprises: a Convolutional Neural Network (CNN) module, a global average pooling layer, a Bidirectional Long Short-Term Memory (BiLSTM) network layer, and a fully connected layer stacked together. The CNN module includes convolutional layers and pooling layers. The CM model parameters include { ppt 1, ppt 2, ppt 3 , ppt 4}, ppt 1 represents the set of parameters for the number of modules in the convolutional neural network and the number of layers in the bidirectional long short-term memory network. ppt 2 represents the parameter set of the convolutional neural network module. ppt 3 represents the parameter set of the bidirectional long short-term memory network layer. ppt 4 represents the set of configuration parameters for the model training hyperparameters; ppt 1={ C 1, B 1}, where C 1 represents the number of modules in the convolutional neural network. B 1 represents the number of layers in the bidirectional long short-term memory network; ,in k 0 indicates the first k 0 convolutional neural network modules For the first k The number of convolutional filters in the convolutional layer of a 0-convolutional neural network module. For the first k The kernel size of a convolutional layer with 0 convolutional neural network modules. For the first k The activation function type of the convolutional layer in a convolutional neural network module with 0 modules. For the first k The pooling type used after the convolutional layers of 0 convolutional neural network modules. For the first k The pooling size used after the convolutional layers of 0 convolutional neural network modules; ,in j 0 indicates the first j 0 bidirectional long short-term memory network layers, For the first j The number of neurons in 0 bidirectional long short-term memory network layers For the first j Activation function types for 0 bidirectional long short-term memory network layers; ppt 4={ bt c , dr c , lr c ,opt c , ret c },in bt c For CM training batch size, dr c This refers to the dropout rate during CM training. lr c For CM training learning rate, opt c The type of optimizer used for CM training. ret c The regularization type used for CM training; the adversarial attack training dataset. D adv_tr and adversarial attack test dataset D adv_te Specifically, it is obtained through the following sub-steps: S11. When the adversarial attack method is the fast gradient sign method adversarial attack, the adversarial attack training dataset... D adv_tr and adversarial attack test dataset D adv_te The simulation generation process is shown in S11.1 to S11.4: S11.1, initialize the fast gradient sign method attack perturbation strength ε 1. generate a fast gradient sign method attack sample according to formula (1) X advFGSM : (1) In the formula, x Represents the original sample; d x Represents the original input sample x The total number of feature dimensions; Indicates the original sample x The partial derivatives; Indicates sample x China's power load i The true values ​​of 0 features; Indicates the first i 0 feature loss functions θ These are the network parameters of the CM model; sign represents the gradient calculation symbol. This indicates that the generated adversarial perturbation is constrained within the range [0,1]. S11.2, When using the fast gradient sign method adversarial attack training dataset D adv_tr_FGSM When generating adversarial examples, set x = x adv_tr_FGSM,tx ,1≤ tx ≤ Tr ,in x adv_tr_FGSM,tx This represents the training dataset for fast gradient signified adversarial attacks. D adv_tr_FGSM The first in tx One original sample, Tr This indicates that from the normal training dataset D tr Repeat the fast gradient sign-based adversarial sample generation process in step S11.1 without repeating the randomly selected number of samples, until the fast gradient sign-based adversarial attack training dataset is reached. D adv_tr_FGSM All original samples x adv_tr_FGSM,tx All were replaced with the fast gradient sign-based adversarial examples calculated by formula (1). X adv_tr_FGSM,tx ,in X adv_tr_FGSM,tx express D adv_tr_FGSM The simulated generation of the first tx One fast gradient sign adversarial example; S11.3, When using the Fast Gradient Signed Method adversarial attack test dataset D adv_te_FGSM When generating adversarial examples, set x = x adv_te_FGSM,ty ,1≤ ty ≤ Te ,in, x adv_te_FGSM,ty This represents the test dataset D for fast gradient signified attack. adv_te_FGSM The first in ty One original sample, Te This indicates the normal test dataset. D te Repeat the fast gradient sign-method adversarial sample generation process in step S11.1 above, without repeating the randomly selected number of samples, until the fast gradient sign-method adversarial attack test dataset is obtained. D adv_te_FGSM All original samples x adv_te,ty All were replaced with adversarial samples calculated using formula (1). X adv_te_FGSM,ty ,in X adv_te_FGSM,ty for D adv_te_FGSM The simulated generation of the first ty One adversarial example; S11.4 The generated fast gradient signified adversarial attack training dataset D adv_tr_FGSM For normal training dataset D tr Replace the sample at the corresponding position in the middle, and then complete the above replacement. D tr The dataset is labeled as an adversarial attack training dataset. D adv_tr The generated fast gradient symbolic method adversarial attack test dataset will be used. D adv_te_FGSM For normal test dataset D te Replace the sample at the corresponding position in the middle, and then complete the above replacement. D te Dataset labeling adversarial attack test dataset D adv_te ; S12. When the adversarial attack method is projective gradient descent adversarial attack, the adversarial attack training dataset... D adv_tr and adversarial attack test dataset D te The simulation generation process is shown in S12.1 to S12.7: S12.1 Initialize the strength of the projection gradient descent adversarial attack perturbation. ε 2. Adversarial attack step size of projected gradient descent α 1. Number of iterations for adversarial attack against projected gradient descent ζ 1. Initial values ​​for projected gradient descent to counteract disturbances based on CM , This represents the initial perturbation matrix with the same shape as the original sample, and records the number of iterations of the projected gradient descent adversarial generation. t =0; S12.2, iteratively generate the first t +1 projection gradient descent against perturbation and perturbation constraint: (2) In the formula, x Represents the original sample; This indicates that the generated adversarial perturbation is performed within the range of [- ε 2, ε Constraints of 2]; Indicates the iterative generation of the first... t Subprojected gradient descent perturbation, 0≤ t < ζ 1; Indicates resistance to disturbance The partial derivatives; y true This represents the actual value of the sample power load; L ( θ ( x + ), y true ) represents the loss function. θ For the network parameters of the CM model; sign represents the gradient calculation symbol; S12.

3. According to formula (3), generate the iteration number. t +1 adversarial examples of projective gradient descent that satisfy the range constraint : (3) In the formula, This indicates that the generated projective gradient descent adversarial examples are constrained to have a range of [0,1]; and the number of iterations for generating the projective gradient descent adversarial examples is recorded. t = t +1; The number of iterations for generating adversarial examples using labeled projection gradient descent reaches [number]. ζ The adversarial examples generated after step 1 are X advPGD In the training dataset for projective gradient descent adversarial attacks D adv_tr_PGD When generating initial adversarial examples, x = x adv_tr_PGD,tx ,1≤ tx ≤ Tr ,in x adv_tr_PGD,tx This represents the training dataset for projective gradient descent adversarial attacks. D adv_tr_PGD The first in tx One original sample, Tr This represents the training dataset for projective gradient descent adversarial attacks. D adv_tr_PGD The total number of all samples; in the test dataset for projected gradient descent adversarial attacks. D adv_te_PGD When generating initial adversarial examples, x = x adv_te_PGD,ty ,1≤ ty ≤ Te ,in ty This represents the dataset used for adversarial attacks against projective gradient descent. D adv_te_PGD The first in ty One original sample, Te This represents the dataset used for adversarial attacks against projective gradient descent. D adv_te_PGD The total number of all samples; S12.4, The maximum number of iterations for generating adversarial examples using projected gradient descent is reached. ζ The adversarial examples generated after step 1 are labeled as follows: X advPGD ; S12.5, On the training dataset for projected gradient descent adversarial attacks D adv_tr_PGD When generating adversarial examples for projected gradient descent, set x = x adv_tr_PGD,tx ,1≤ tx ≤ Tr ,in x adv_tr_PGD,tx This represents the training dataset for projective gradient descent adversarial attacks. D adv_tr_PGD The first in tx For each original sample, repeat steps S12.1 to S12.4 of the adversarial example generation process until the training dataset for projected gradient descent adversarial attacks is obtained. D adv_tr_PGD All original samples x adv_tr_PGD,tx All were replaced with projected gradient descent adversarial examples. X adv_tr_PGD,tx ,in X adv_tr_PGD,tx express D adv_tr_PGD The Middle tx One projected gradient descent adversarial example; S12.6, On the test dataset for adversarial attacks on projected gradient descent. D adv_te_PGD When generating adversarial examples for projected gradient descent, set x = x adv_te_PGD,ty ,1≤ ty ≤ Te ,in, ty This represents the dataset used for adversarial attacks against projective gradient descent. D adv_te_PGD The first in ty For each original sample, repeat steps S12.1 to S12.4 of the adversarial sample generation process until an adversarial attack test dataset is generated. D adv_te_PGD All original samples x adv_te,ty All were replaced with projected gradient descent adversarial examples X adv_te_PGD,ty ,in X adv_te_PGD,ty express D adv_te_PGD The Middle ty One projected gradient descent adversarial example; S12.7 The generated projection gradient descent adversarial attack training dataset D adv_tr_PGD For normal training dataset D tr Replace the sample at the corresponding position in the middle, and then complete the replacement. D tr The dataset is labeled as an adversarial attack training dataset. D adv_tr The generated projection gradient descent adversarial attack test dataset D adv_te_PGD For normal test dataset D te Replace the sample at the corresponding position in the middle, and then complete the replacement. D te The dataset is labeled as an adversarial attack test dataset. D adv_te .

3. The power load multi-objective robust deep forecasting method against adversarial attacks according to claim 1, wherein, The first parameter includes the maximum number of offline iterations. G Maximum number of solutions in the archive A max Adversarial training threshold T ; The mean squared error of the prediction model for the power load anti-robust CNN-BiLSTM on normal samples f 1. Prediction mean square error on adversarial examples f 2 and the number of model parameters f 3. As shown in formulas (4) to (6): (4) (5) (6) In the formula, y w This represents the actual value of the power load. This represents the predicted value of a normal sample of power load. This represents the predicted value of the power load countermeasure sample; Indicates the first l The number of parameters in a CNN module C Indicates the number of CNN modules; Indicates the first k The number of parameters in a BiLSTM layer. B Indicates the number of BiLSTM layers; P Dense This indicates the number of parameters in a fully connected network layer.

4. The power load multi-objective robust deep forecasting method against adversarial attacks of claim 1, wherein, The second parameter includes the preset maximum number of CNN modules. C max and minimum value C min Maximum number of BiLSTM layers B max and minimum value B min Maximum number of convolutional kernels in a CNN module fn max and minimum value fn min , No. l Minimum number of convolutional kernels in a CNN module fn min and maximum value fn max , No. l Minimum kernel size of a CNN module kn min and maximum value kn max , No. l The activation function of each CNN module encodes the minimum value. an min and maximum value an max , No. l The minimum value is encoded by pooling type of each CNN module. pt min and maximum value pt max , No. l The minimum value is encoded by pooling type of each CNN module. ps min and maximum value ps max , No. k Minimum number of neurons in a layer BiLSTM nb min and maximum value nb max , No. k Layer BiLSTM activation function encodes minimum value ab min and maximum value ab max Minimum batch size bt min and maximum value bt max Minimum value of optimizer encoding opt min and maximum value opt max Minimum learning rate lr min and maximum value lr max The minimum value encoded by regularization ret min and maximum value ret max Minimum discard rate dr min and maximum value dr max ; The construction of a multi-objective adversarial robust CNN-BiLSTM predictive code for power load generates individuals to be mutated. p g Specifically, this includes: encoding the parameters of the CNN-BiLSTM model to be optimized based on the second parameter. The encoding of each CNN-BiLSTM model to be optimized includes the following four parts: pt 1={ C , B }, pt 2={ pt 2,l = ( fn l , kn l , an l , pt l , ps l ), l =1,2, …, C }, pt 3={ pt 3,k =( nb k , ab k ), k =1, 2, …, B }, pt 4=( bt , opt , lr , ret , dr ), pt 1 represents the set of numbers of CNN modules and BiLSTM layers. pt 2 represents the set of neural network parameters for the CNN module. pt 3 represents the parameter set of the BiLSTM layer neural network. pt 4 represents the set of hyperparameters used in model training; C This indicates the number of CNN modules in the CNN-BiLSTM model to be optimized. B This indicates the number of BiLSTM layers in the CNN-BiLSTM model to be optimized; pt 2,l Indicates the first l CNN module parameter encoding fn l Indicates the first l Number of convolutional kernels per CNN module kn l Indicates the first l The size of the convolutional kernel in each CNN module. an l Indicates the first l Activation functions for each CNN module pt l Indicates the first l Each CNN module pooling layer type ps l Indicates the first l The pooling kernel size for each CNN module; pt 3,k Indicates the first k Layer BiLSTM parameter encoding, nb k Indicates the first k Number of BiLSTM neurons per layer ab k Indicates the first k Layer BiLSTM activation function; bt This indicates the batch size for training the CNN-BiLSTM model. opt This represents the training optimizer for the CNN-BiLSTM model. lr This represents the learning rate used to train the CNN-BiLSTM model. ret This indicates the regularization method used during the training of the CNN-BiLSTM model. dr This represents the dropout rate during the training of the CNN-BiLSTM model; in, pt 1. pt 2,l , pt 3,k , pt The specific values ​​of 4 are shown in formulas (7) to (10): (7) In the formula, Randint represents the random integer selection function. C min This represents the minimum value of the number of CNN modules. C max This represents the maximum value that can be taken for the number of CNN modules. C min ≤ C ≤ C max ; B min This represents the minimum value for the number of BiLSTM layers. B max This represents the maximum value of the BiLSTM layer number. B min ≤ B ≤ B max ; (8) In the formula, fn min Indicates the first l Minimum number of convolutional kernels for a CNN module fn max Indicates the first l The maximum number of convolutional kernels in a CNN module. fn min ≤ fn l ≤ fn max ; kn min Indicates the first l The minimum kernel size for each CNN module. kn max Indicates the first l The maximum value of the convolutional kernel size of each CNN module. kn min ≤ kn l ≤ kn max ; an min Indicates the first l The minimum value encoded by the activation function of each CNN module. an max Indicates the first l The maximum value encoded by the activation function of each CNN module. an min ≤ an l ≤ an max ; pt min Indicates the first l The minimum value of pooling type encoding for each CNN module. pt max Indicates the first l The maximum value of the pooling type encoding of each CNN module. pt min ≤ pt l ≤ pt max ; ps min Indicates the first l The minimum value of pooling type encoding for each CNN module. ps max Indicates the first l The maximum value of the pooling type encoding of each CNN module. ps min ≤ ps l ≤ ps max ; (9) In the formula, nb min Indicates the first k Minimum number of neurons in a layer of BiLSTM nb max Indicates the first k The maximum number of neurons in a layer of BiLSTM. nb min ≤ nb k ≤ nb max ; ab min Indicates the first k The minimum value encoded by the activation function of the BiLSTM layer. ab max Indicates the first k The maximum value encoded by the activation function of the BiLSTM layer. ab min ≤ ab k ≤ ab max ; (10) In the formula, Random represents the function for selecting random decimals; bt min This represents the minimum batch size. bt max This represents the maximum value of the batch size. bt min ≤ bt ≤ bt max ; opt min This represents the minimum value encoded by the optimizer. opt max This represents the maximum value encoded by the optimizer. opt min ≤ opt ≤ opt max ; lr min This represents the minimum learning rate. lr max This represents the maximum value of the learning rate. lr min ≤ lr ≤ lr max ; ret min This represents the minimum value encoded using regularization. ret max This represents the maximum value encoded using regularization. ret min ≤ ret ≤ ret max ; dr min This represents the minimum discard rate. dr max This represents the maximum discard rate. dr min ≤ dr ≤ dr max ; Based on formulas (7) to (10), individuals to be mutated are randomly generated. p g Its encoding is ,in g This represents the current iteration number. pt 2,1 This indicates the parameter encoding of the first CNN module. Indicates the first C CNN module parameter encoding pt 3,1 This indicates the encoding of the first layer BiLSTM parameters. Indicates the first B Layer BiLSTM parameter encoding.

5. The multi-objective robust depth prediction method for power load defense against attacks according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S41. Treat the variant individuals according to formulas (7) to (10). p g Each gene locus is mutated sequentially to generate a scale of N g population ,in It is a population P g The first individual, Indicates population P g The i Individual, Indicates population P g The N g Individual, , C g yes The number of CNN modules, B g yes The number of BiLSTM layers; ,in Individual The corresponding pt 1, Represents an individual The corresponding pt 2, Represents an individual The corresponding pt 3, Represents an individual The corresponding pt 4, express The parameters of the first dimension, express No. d g Dimensional parameters, express No. D g Dimensional parameters, D g express Maximum dimension ; During initialization, i.e., when the number of iterations... g When =0, initialize the individual according to formulas (7) to (10). p 0 Each gene locus is mutated sequentially to generate a scale of N Initial population of 0 ,in It is the first individual in the initial population. Denotes the initial population's... i Individual, Denotes the initial population's... N 0 individuals , C 0 is The number of CNN modules, B 0 is The number of BiLSTM layers; ,in Individual of pt 1, Represents an individual of pt 2, Represents an individual of pt 3, Represents an individual of pt 4, express The parameters of the first dimension, express No. d 0-dimensional parameters express No. D 0-dimensional parameters D 0 represents Maximum dimension ; S42, Based on normal training dataset D tr right P g medium-sized individuals The CNN-BiLSTM prediction model was trained based on the normal test dataset. D te Calculate individuals The corresponding optimization objective of the prediction model after the above training f 1 and f 3. Evaluation f Is 1 less than or equal to the adversarial training threshold? T If not, set optimization goals. f 2=1, interrupting the individual The confrontation training; if so, proceed to the next step. S 6. For individuals Perform adversarial training and calculate the optimal objective. f 2.

6. The multi-objective robust depth prediction method for power load defense against attacks according to claim 1, characterized in that, Step S5 specifically includes the following sub-steps: S51, Based on the optimization objective f 1. f 2 and f 3 pairs of populations P g Perform a non-dominated ranking on all individuals, and check each individual one by one whether it is completely dominated by other individuals. That is, if there is an individual... exist f 1. f 2 and f All 3 are smaller than the individual This indicates that the individual Complete domination of individuals Then delete the controlled individual. Individuals who are not completely dominated retain their individuality. Mark all retained individuals as the retained solution set. U ,in, express P g China differs from individuals, that is i ≠ j ; S52, Solution set U External archive deset S Merging is marked as a solution set Q According to the optimization objective f 1. f 2 and f 3 pairs Q All solutions are sorted in a non-dominated order, and each solution is evaluated to determine whether it is completely dominated by other solutions. That is, if a solution exists... exist f 1. f 2 and f All 3 are less than the solution This explains the solution. Complete Domination Solution Then delete the dominated solution. If not fully dominated, the solution is retained. Mark all retained solutions as the new external archive solution set. S That is, updating the external archive solution set. S ,in, and Both represent solution sets Q The solution in, and the solution Different from the solution ,Right now i 1≠ j 1; S53, Judgment Is it greater than the number of external archives? A max If not, proceed to step S54; if yes, then deset the external archive. S Crowding distance is calculated and sorted for each non-dominated solution, and redundant solutions are removed; Represents external archive solution set S The number of dominant solutions in the Middle East; S54, External Archive Decomposition S Randomly select a non-dominated solution and mark it as the individual to be mutated in the next optimization iteration. p g Number of update iterations g = g +1.

7. The multi-objective robust depth prediction method for power load defense against attacks according to claim 6, characterized in that, In step S53, the calculation of the congestion distance includes: S53.1, Decomposing external archives S The solutions are sorted in ascending order according to a certain optimization objective value, and the minimum value of the sorted solutions in terms of that objective value is marked. f min and maximum value f max Set the crowding distance of the target that is the minimum and maximum solution after sorting to infinity; S53.2, For each optimization objective, sort a non-dominated solution in ascending order. , Represents external archive solution set S Sort in ascending order i Two solutions are available. The congestion distance is calculated according to formula (11). d s Then, based on the distance of the crowding level d s External archive deset S Sort all non-dominated solutions in descending order: (11) In the formula, and These represent external archive solution sets. S Non-dominated solutions sorted in ascending order In optimizing objectives The optimization objective value of the adjacent solutions above, and They represent the optimization objectives for ascending order sorting, respectively. The maximum and minimum values ​​on; S53.3, Complete the external archive decomposition. S After calculating the crowding distance of all non-dominated solutions and sorting them in descending order, delete them. t A solution with a small crowding distance t = – T .

8. The multi-objective robust depth prediction method for power load defense against attacks according to claim 1, characterized in that, Step S6 specifically includes: Based on adversarial attack training dataset D adv_tr For individuals The CNN-BiLSTM neural network model was subjected to adversarial training and tested on an adversarial attack dataset. D adv_te Calculate the individual according to formula (5) The corresponding optimization objective of the prediction model after adversarial training f 2.

9. The multi-objective robust depth prediction method for power load defense against attacks according to claim 1, characterized in that, Step S8 specifically includes: Deploying BM online in the power system for real-time datasets D t Online power load forecasting is performed, and its robustness performance is evaluated. The performance evaluation indicators include root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), and their calculation formulas are shown in formulas (12) to (14), respectively. (12) (13) (14) In the formula, express D t No. rt The actual power load value of each test sample. Indicates BM's target D t The Middle rt The predicted power load for each test sample.

10. A robust depth prediction system for multi-objective power load defense against attacks, used to implement the robust depth prediction method for multi-objective power load defense against attacks as described in any one of claims 1-9, characterized in that, The system includes: The module for constructing normal and adversarial datasets is used to build normal training and test datasets based on historical power system load databases, and real-time datasets based on real-time power system load databases. Furthermore, it simulates adversarial attack training datasets for existing deep learning power load prediction models (CM) using fast gradient sign-matrix and projective gradient descent attacks, respectively. D adv_tr and adversarial attack test dataset D adv_t ; The parameter setting and objective function construction module is used to set the first parameter and construct the multi-objective optimization objective function of the multi-objective adversarial robust CNN-BiLSTM prediction model for power load. The multi-objective optimization objective function includes the mean square error of the CNN-BiLSTM prediction model on normal samples. f 1. Mean squared error of prediction on adversarial examples f 2 and the number of model parameters f 3; The optimized model encoding definition module is used to set the second parameter, construct a multi-objective adversarial robust CNN-BiLSTM predictive encoding for power load, and generate individuals to be mutated. p g And set the current iteration number. g =0; A population generation module based on mutation operations and a module for determining whether to enter adversarial training are used to treat mutated individuals. p g Perform mutation operations to generate a population. P g Assess the population P g medium-sized individuals optimization objectives f 1 and f 3, and determine f Is 1 less than or equal to the adversarial training threshold? T If not, set f 2=1, interrupting the individual If so, proceed to the adversarial training module for individual training. Conduct adversarial training and calculate the optimal objective. f 2; A multi-objective non-dominated solution evaluation and external archive update module is used for population analysis. P g All individuals according to the optimization objective f 1. f 2 and f 3. Perform non-dominated sorting and store non-dominated solutions in an external archive solution set. S In, and based on external archive update technology S Update and obtain the individuals to be mutated for the next optimization iteration. p g Number of update iterations g = g +1; a contrast training module for training an individual based on D adv_tr performing contrast training on an individual and based on D adv_te computing an optimization objective for the individual f 2; The iteration judgment module is used to determine the iteration count. g Less than the set maximum number of iterations G At that time, the system repeatedly enters the population generation and adversarial training decision module based on mutation operations, the multi-objective non-dominated solution evaluation and external archive update module, and the adversarial training module; when g Greater than or equal to G At that time, mark the external archive solution set. S All solutions are Pareto front solutions. CNN-BiLSTM, which balances high prediction accuracy, strong adversarial robustness, and lightweight design, is selected and marked as the optimal model for multi-objective adversarial robust deep prediction of power load (BM). The online deployment module is used to deploy the selected model BM online in the power system, targeting the real-time dataset of the power system load database. D t Perform online power system load forecasting and test and evaluate its robustness performance indicators.

Images