Robust Photovoltaic Power Prediction Method and System to Resist Latent Multimodal Injection Attacks

Abstract

This invention discloses a robust photovoltaic (PV) power prediction method and system to resist latent multimodal injection attacks. The method includes: modeling the coupling relationships between multimodal data such as PV power measurements, weather forecasts, and cloud imagery using attention moments to construct a latent injection attack applied to the multimodal data; proposing a multimodal learning framework as the environment for robust PV power prediction; proposing an agent based on the actor-critic algorithm, and pre-training the actor module using a clear sky index membership loss function; and using the agent to iteratively search for multi-step robust PV power prediction results. This invention can be used in PV power plants, providing high-precision PV power prediction results within 0.5 hours that are resistant to external data attacks, effectively improving the accuracy and robustness of PV power prediction.

Description

Technical Field

[0001] This invention belongs to the field of renewable energy development and utilization technology, specifically relating to a robust photovoltaic power prediction method and system for resisting latent multimodal injection attacks. Background Technology

[0002] With the massive influx of photovoltaic (PV) power into the grid, PV power prediction is crucial for ensuring the safe and stable operation of the power grid. Currently, multimodal data, including PV measurements, ground-based cloud maps, and weather forecasts, are widely used in PV power prediction. However, as multimodal data access increases and the nonlinearity of prediction models improves, the vulnerability of these models to latent data attacks also increases. Therefore, it is necessary to balance the robustness and accuracy of PV power prediction, effectively resisting latent multimodal data injection attacks, and using data support to ensure the consumption of new energy sources and the security of the power grid.

[0003] Existing research only examines the impact of single-modal data injection attacks on model robustness, failing to consider the coupling effects of multimodal data, making it difficult to adapt to commonly used multimodal photovoltaic power prediction models. Therefore, simulating injection-type data attacks that take into account the coupling characteristics of multimodal data and providing efficient methods to improve the robustness of prediction models are the current challenges and key areas of research in robust photovoltaic power prediction. Summary of the Invention

[0004] Purpose of the invention: To address the aforementioned difficulties in robust photovoltaic power prediction, this invention proposes a robust photovoltaic power prediction method and system that resists latent multimodal injection attacks, thereby synergistically improving the accuracy and robustness of photovoltaic power prediction.

[0005] Technical Solution: To achieve the above-mentioned objectives, the first aspect of this invention proposes a robust photovoltaic power prediction method to resist latent multimodal injection attacks, comprising the following steps:

[0006] Photovoltaic power measurements, weather forecasts, and ground-based cloud images are used as multimodal data to drive photovoltaic power prediction tasks. Attention moments are constructed using self-attention of each mode and cross-attention of different modes to build an injection attack strategy. The optimal injection attack strategy applied to each mode is obtained by solving for the photovoltaic power prediction loss.

[0007] Based on the characteristics of multimodal data, a robust photovoltaic power prediction environment based on multimodal learning is constructed. For photovoltaic power measurement values, ground-based cloud maps, and weather forecasts, sparse self-attention, spatial pyramid attention, and multilayer perceptron are used to extract corresponding features, and photovoltaic power prediction results are obtained through integrated weighting.

[0008] Based on the constructed robust photovoltaic power prediction environment, the state space and action space are defined, and the injection attack strategy obtained from the solution is utilized. To simulate a latent injection attack, a smart agent based on the actor-critic algorithm is constructed by tampering with state variables. The smart agent includes an actor model and a critic model. The actor model interacts with the environment through a policy function and selects action variables based on the state variables. The critic model performs a global value evaluation of the action variables of the actor model. The actor model is pre-trained using the clear sky index membership loss function. The optimal action function is obtained by interacting with the environment and iteratively searching, and the multi-step robust prediction result of photovoltaic power is obtained by solving the problem.

[0009] According to the first aspect, in some specific implementations, photovoltaic power measurements, weather forecasts, and ground-based cloud maps are used as multimodal data to drive the photovoltaic power prediction task. Attention moments are constructed using the self-attention of each mode and the cross-attention of different modes in order to build an injection attack strategy. The optimal injection attack strategy applied to each mode is obtained by solving for the photovoltaic power prediction loss. include:

[0010] Multimodal data, including photovoltaic power measurement, weather forecast, and ground-based cloud map, are represented as X = {X...} (1) ,X (2) ,X (3) Attention moments are constructed using modal self-attention and cross-attention between different modalities to build an injection attack strategy. The attack strategies applied to each mode are obtained by using the moment estimation method.

[0011]

[0012]

[0013] Where, ρ i SA is a weighting factor applied to various types of attention mechanisms. (i) For the self-attention calculation result of mode i, CA (j)→(i) The result of cross-attention calculation for mode j and mode i;

[0014] Injection attack strategy Obtain the injection attack modulus applied to each mode. Based on photovoltaic power prediction model A loss function is constructed to train the model, ensuring that the predicted result after the attack deviates as much as possible from the photovoltaic power measurement value Y at the point to be predicted:

[0015]

[0016] Where l is the root mean square error loss function, ||·|| ∞ ξ is an infinite norm, used to constrain the magnitude of the injection attack. The optimal injection attack strategy.

[0017] According to the first aspect, in some specific embodiments, spatial pyramid attention is used to extract latent features from cloud imagery, including:

[0018] With X (3) Representing ground-based cloud map stream data, spatial pyramid attention is used to extract features H from each level of the cloud map. (3) :

[0019] H (3) =f Att (X (3) ,SR(X (3) )W,SR(X (3) W)

[0020]

[0021] Where SR represents the spatial dimensionality reduction operation, and W and W S This indicates that a linear regression layer has been applied; This indicates the attenuation of the input x in terms of spatial size. times, f Norm This represents the normalization function.

[0022] According to the first aspect, in some specific embodiments, the photovoltaic power prediction result obtained through integrated weighting is expressed as follows:

[0023]

[0024] Where, ω m Let be the weighting factor for the m-th modality data, which will be solved by iterative optimization by the agent, and satisfy ∑ m ω m =1; H m This represents the feature extracted from the m-th modality data; This represents a robust photovoltaic power prediction environment driven by multimodal data.

[0025] According to the first aspect, in some specific embodiments, in the agent based on the actor-critic algorithm, the actor model uses the policy function a = μ(s|θ) μ It interacts with the environment, selecting action variable a from state variable s, with parameter θ. μ Once the action variable a is selected, the next state s′ and immediate reward r are obtained from the predefined environment; the critic model evaluates the action variable a proposed by the actor model for the global value Q(s,a|θ). QThe gain of ) has a parameter θ. Q ;

[0026] Selected state variable s t Includes: a sequence of photovoltaic power measurements with a time window length of d. and cloud map sequence And a weather forecast sequence with a forecast time window length of p.

[0027] Selected action variable a t Weighting factors for each modal feature

[0028] Choose reward function r t for l represents the loss function.

[0029] According to the first aspect, in some specific implementations, the actor model is pre-trained using the clear sky index membership loss function, including:

[0030] For state variable s t Select a cloud map with a time window length p to be predicted. Calculated clear sky index Φ t+1:t+p As the target value, a clear sky index membership loss function is constructed using cross-entropy, and the actor model is pre-trained: Among them, D KL This is the KL divergence calculation method.

[0031] According to the first aspect, in some specific implementations, the optimal action function is obtained by utilizing the interaction between the intelligent agent and the environment and iterative search, and the multi-step robust prediction result of photovoltaic power is obtained by solving the problem, including:

[0032] The ultimate goal of solving the actor model is to maximize the cumulative reward function J: Where γ is the decay factor and n is the number of steps to calculate the cumulative reward function;

[0033] Solving the critic model requires maximizing the strategy evaluation function μ(s) = argmax. a Q(s,a) is represented as:

[0034] The actor model parameters θ obtained through online learning Q With the critic model parameters θ μ The target actor model θ is updated using an asymptotic parameter update method. Q′ and the target critic model θ μ′ To achieve stable updates of model parameter information through iterative searching by the intelligent agent: Where τ << 1 is the parameter update equilibrium factor, from which the deterministic action policy function μ(s|θ) can be obtained. μ Through environmental interaction, multi-step robust prediction results of photovoltaic power are obtained.

[0035] A second aspect of the present invention provides a robust photovoltaic power prediction system resistant to latent multimodal injection attacks, comprising:

[0036] The injection attack strategy construction module uses photovoltaic power measurements, weather forecasts, and ground-based cloud images as multimodal data to drive photovoltaic power prediction tasks. It constructs attention moments using self-attention of each mode and cross-attention of different modes to build injection attack strategies. The optimal injection attack strategy applied to each mode is obtained by solving for the photovoltaic power prediction loss.

[0037] The photovoltaic power robust prediction environment construction module is used to construct a photovoltaic power robust prediction environment based on multimodal learning according to the characteristics of multimodal data. Specifically, for photovoltaic power measurement values, ground-based cloud maps, and weather forecasts, sparse self-attention, spatial pyramid attention, and multilayer perceptron are used to extract corresponding features, and photovoltaic power prediction results are obtained through integrated weighting.

[0038] The photovoltaic power robust prediction intelligent agent solution module is used to define the state space and action space based on the constructed photovoltaic power robust prediction environment, and to utilize the injection attack strategy obtained from the solution. To simulate a latent injection attack, a smart agent based on the actor-critic algorithm is constructed by tampering with state variables. The smart agent includes an actor model and a critic model. The actor model interacts with the environment through a policy function and selects action variables based on the state variables. The critic model performs a global value evaluation of the action variables of the actor model. The actor model is pre-trained using the clear sky index membership loss function. The optimal action function is obtained by interacting with the environment and iteratively searching, and the multi-step robust prediction result of photovoltaic power is obtained by solving the problem.

[0039] A third aspect of the invention also provides a computer device comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when executed by the processors, the programs implement the steps of the photovoltaic power robust prediction method against latent multimodal injection attacks as described in the first aspect of the invention.

[0040] A fourth aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the photovoltaic power robust prediction method against latent multimodal injection attacks as described in the first aspect of the present invention.

[0041] Beneficial effects: (1) This invention simulates the optimal latent injection attack strategy applied to each modal input data through attention moment simulation, and quantitatively analyzes the vulnerability of the multimodal prediction model, providing data support for solving the bi-layer optimization problem of robust photovoltaic power prediction; (2) The robust photovoltaic power prediction method based on multimodal learning proposed in this invention can efficiently extract features for multimodal input data features, ensuring the accuracy of photovoltaic power prediction; (3) This invention realizes the efficient response to latent data attacks by realizing the interaction between the agent and the environment to solve the actor-critic model, eliminating the misleading effect of latent data attacks on photovoltaic power prediction output, and improving the robustness of photovoltaic power prediction; (4) This invention fits the actual situation of photovoltaic power plant production and grid economic and security dispatch, and can be applied to photovoltaic power plants that widely use multimodal deep learning, meeting the photovoltaic power robust prediction task within 0.5 hours, reducing the misleading effect of data attacks on photovoltaic grid-connected power generation plans on grid economic and security dispatch, promoting the consumption of new energy and the security of power supply, and has good social and economic benefits. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the latent multimodal injection attack proposed in this invention;

[0043] Figure 2 This is a schematic diagram of the photovoltaic power robust prediction intelligent agent solution method proposed in this invention;

[0044] Figure 3 This is a schematic diagram of the overall framework for robust photovoltaic power prediction against latent multimodal injection attacks proposed in this invention.

[0045] Figure 4 The results of the hidden multimodal injection attack on photovoltaic power prediction in the embodiments of the present invention;

[0046] Figure 5 These are robust prediction results of photovoltaic power under injection attacks of different intensities in the embodiments of the present invention;

[0047] Figure 6 This is a multi-step prediction curve of photovoltaic power in an embodiment of the present invention. Detailed Implementation

[0048] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0049] Reference Figures 1-3 This invention proposes a robust photovoltaic power prediction method to resist latent multimodal injection attacks, comprising the following steps:

[0050] Step S1: Use attention moments to model the coupling relationships between multimodal data and construct a latent injection attack that is applied to the multimodal data.

[0051] Multimodal data, including photovoltaic power measurement, weather forecast, and ground-based cloud map, are represented as X = {X...} (1) ,X (2) ,X (3) To fully model the coupling relationships between multimodal data, attention moments are constructed using self-attention of each modality and cross-attention of different modalities to build an injection attack strategy. The attack strategies applied to each mode are obtained by using the moment estimation method.

[0052]

[0053]

[0054] Where ρ1, ρ2, and ρ3 are weighting factors applied to each type of attention mechanism. (1) The self-attention calculation result for mode (1) is CA. (2)→(1) The calculation results for the cross-attention of mode (2) and mode (1) are as follows:

[0055] SA (1) =f Att (X (1) ,X (1) ,X (1) )

[0056] CA (2)→(1) =f Att (X (2) ,X (1) ,X (1) )

[0057] Among them, f Att This is for the attention mechanism. The meanings of the other parameters are similar and will not be repeated.

[0058] Injection attack strategy applied by external attackers The solution method is as follows: based on injection attack strategy The injection attack modulus applied to each mode can be obtained. Next, since the internal core prediction model is usually strictly protected, external attackers need to steal the prediction model based on the mapping relationship between multimodal data input and measured photovoltaic power values. Then, a loss function is constructed to solve for the optimal injection attack strategy, so that the prediction result after the attack deviates as much as possible from the photovoltaic power measurement value Y of the point to be predicted:

[0059]

[0060] Where l is the root mean square error loss function, ||·|| ∞ ξ is an infinite norm used to constrain the magnitude of the injection attack. Generally, ξ needs to be controlled within 5% to ensure that such injection attacks are latent and difficult to detect by existing injection attack identification methods. The optimal injection attack strategy.

[0061] Step S2 involves constructing a multimodal learning framework. This framework utilizes temporal autocorrelation attention and spatial pyramid attention to extract latent features from photovoltaic measurements and cloud imagery. These features are then integrated and weighted to obtain the photovoltaic power prediction result, serving as the environment for robust photovoltaic power prediction. Specifically, this includes:

[0062] First, spatial pyramid attention is used to extract rich features H from each level of the ground-based cloud image stream data. (3) :

[0063] H (3) =f Att (X (3) ,SR(X (3) )W,SR(X (3) W)

[0064]

[0065] Where SR represents the spatial dimensionality reduction operation, and W and W S This indicates that a linear regression layer was applied to ensure that the feature vectors have a uniform shape. This indicates the attenuation of the input x in terms of spatial size. times, f Norm This represents the normalization function to ensure that the data distribution of the feature vector remains unchanged after spatial scale decay.

[0066] Secondly, for photovoltaic power sequence data, sparse self-attention is used to extract the temporal features H of the photovoltaic power sequence. (1) Furthermore, a multilayer perceptron was used to obtain the weather forecast features H from the weather forecast data. (2) Finally, the photovoltaic power prediction results were calculated using the integrated weighted method.

[0067]

[0068] Where, ω m Let be the weighting factor, and satisfy ∑ m ω m =1. This allows for the construction of a photovoltaic power prediction environment that integrates multimodal data.

[0069] In step S3, an agent based on the actor-critic algorithm is constructed. The actor model is pre-trained using the clear sky index membership loss function. The agent interacts with the environment and iteratively searches to obtain the optimal action function, thereby obtaining the multi-step robust prediction result of photovoltaic power.

[0070] The actor-critic algorithm-based intelligent agent includes an actor (shown as a performer in the diagram) model and a critic model. (See reference...) Figure 2 The actor model uses the policy function a = μ(s|θ) μ The interaction with the environment is determined by the state variable s, which selects the action variable a, with parameter θ. μ Once the action variable 'a' is selected, both the next state 's' and the immediate reward 'r' can be obtained from the predefined environment. Simultaneously, the critic model evaluates the action variable 'a' proposed by the actor model for its contribution to the global value Q(s,a|θ). Q The gain of ) has a parameter θ. Q .

[0071] Selected state variable s t Includes: a sequence of photovoltaic power measurements with a time window length of d. and cloud map sequence And a weather forecast sequence with a forecast time window length of p. Reference Figure 3 To account for the effect of injection attacks on the modification of input state variables, the injection attack strategy obtained from the solution is used. Simulate the manipulation of state variables;

[0072] Selected action variable a t The weighting factor ω is used to characterize the modal characteristics of photovoltaic measurements, ground-based cloud maps, and meteorological data.

[0073] Choose reward function r t To balance the accuracy of photovoltaic power prediction with the ability to resist latent multimodal injection attacks, the following settings are adopted:

[0074]

[0075] According to an embodiment of the present invention, the actor model is pre-trained using the clear sky index membership loss function, specifically including:

[0076] To reduce the complexity of the action function search space, ensure fitting efficiency, and improve goodness of fit, for the state variable s t Select a cloud map with a time window length p to be predicted. Calculated clear sky index Φ t+1:t+p As the target value, a clear sky index membership loss function is constructed using cross-entropy, and the actor model is pre-trained:

[0077]

[0078] Among them, D KL This is the KL divergence calculation method.

[0079] According to an embodiment of the present invention, the optimal action function is obtained by utilizing the interaction between the intelligent agent and the environment and iterative search, thereby obtaining a multi-step robust prediction result for photovoltaic power, specifically including:

[0080] First, the ultimate goal of solving the agent actor model is to maximize the cumulative reward function J:

[0081]

[0082] Where γ is the decay factor and n is the number of steps to calculate the cumulative reward function. Therefore, the gradient descent rule for the cumulative reward function is:

[0083]

[0084] Where j is the index of the state variable, To achieve the desired operation.

[0085] Secondly, solving the critic model requires maximizing the policy evaluation function μ(s) = argmax. a Q(s,a) is represented as:

[0086]

[0087] Finally, the actor model parameters θ obtained through online learning are... Q With the critic model parameters θ μ The target actor model θ is updated using an asymptotic parameter update method. Q′ and the target critic model θ μ′ To achieve stable updates of model parameter information through iterative searching by the intelligent agent:

[0088]

[0089] Where τ << 1 is the parameter update equilibrium factor. The deterministic action policy function μ(s|θ) can then be obtained from this solution. μ Robust prediction results for photovoltaic power can be obtained through environmental interaction.

[0090] To verify the performance of the method proposed in this invention, the following experiments were conducted in this embodiment. Photovoltaic measurement and ground-based cloud map data were obtained from the SKIPP'D dataset developed by Stanford University. The photovoltaic installed capacity was 29kW, and the ground-based cloud map, after preprocessing, had a resolution of 80*80 pixels and a minimum time resolution of 10 minutes. Meteorological data were obtained from the ERA-5 meteorological reanalysis dataset developed by the European Centre for Weather Forecasts, selecting temperature and irradiance meteorological data for the area near Stanford University (37.2°N-37.6°N, 122.2°W-122.4°N). Based on the above dataset, this embodiment evaluates the robust prediction results of photovoltaic power using three error indices: Mean Error (MAE) and Root Mean Square Error (RMSE), defined as follows:

[0091]

[0092]

[0093] in, and Y t Let N represent the predicted and measured photovoltaic power at time t, respectively. t The number of samples.

[0094] Temporal synchronization matching was performed on the Stanford University photovoltaic-ground-based cloud map dataset and the ERA-5 meteorological reanalysis dataset. Data from January 2017 to December 2018 was selected as the training set, and data from the entire year of 2019 was used as the test sample to test the robust prediction method for ultra-short-term photovoltaic power within 0.5 hours. This embodiment uses cutting-edge temporal robust algorithms and stochastic smoothing algorithms as comparison methods. The robust prediction results of photovoltaic power obtained using this invention are shown in Table 1. Figure 4 The results of the latent multimodal injection attack on photovoltaic power prediction perturbation are shown, demonstrating that the latent multimodal injection attack proposed in this invention can exert a considerable misleading effect on photovoltaic power prediction results within a limited perturbation amplitude range. Figure 5 The robust prediction results of photovoltaic power under different injection attack intensities after applying the robust prediction method are shown. It can be seen that as the injection attack intensity increases, all prediction methods have a certain ability to eliminate the misleading effect of the attack. However, the present invention faces an injection attack amplitude of 5% and only has an error improvement of 8.3%, which is significantly smaller than the time-series robust algorithm (12.3%) and the stochastic smoothing algorithm (12.1%). This shows that the present invention can significantly improve the robustness of photovoltaic power driven by multimodal data. Figure 6 The multi-step photovoltaic power prediction curve obtained using the present invention is shown, demonstrating that the present invention can also be adapted to the task of multi-step photovoltaic power prediction.

[0095] Table 1. Sample error of photovoltaic power prediction test within 0.5 hours in advance based on Stanford University and ERA-5 meteorological reanalysis data.

[0096]

[0097] In summary, the robust photovoltaic power prediction method designed in this invention, which resists latent multimodal injection attacks, firstly proposes a latent multimodal injection attack method based on attention moments. This method utilizes attention moments to model the coupling relationships between multimodal data such as photovoltaic power measurement, weather forecasts, and cloud imagery, constructing a latent injection attack applied to the multimodal data. Secondly, a multimodal learning framework is proposed as the environment for robust photovoltaic power prediction. Finally, an agent based on the actor-critic algorithm is proposed, and the actor module is pre-trained using the clear sky index membership loss function. The agent iteratively searches to obtain multi-step robust photovoltaic power prediction results. This invention can be applied to photovoltaic power plants, providing high-precision photovoltaic power prediction results within 0.5 hours that are resistant to external data attacks, significantly improving the accuracy and robustness of photovoltaic power prediction.

[0098] Based on the same technical concept as the method embodiments, the present invention also provides a robust photovoltaic power prediction system resistant to latent multimodal injection attacks, comprising:

[0099] The injection attack strategy construction module uses photovoltaic power measurements, weather forecasts, and ground-based cloud images as multimodal data to drive photovoltaic power prediction tasks. It constructs attention moments using self-attention of each mode and cross-attention of different modes to build injection attack strategies. The optimal injection attack strategy applied to each mode is obtained by solving for the photovoltaic power prediction loss.

[0100] The photovoltaic power robust prediction environment construction module is used to construct a photovoltaic power robust prediction environment based on multimodal learning according to the characteristics of multimodal data. Specifically, for photovoltaic power measurement values, ground-based cloud maps, and weather forecasts, sparse self-attention, spatial pyramid attention, and multilayer perceptron are used to extract corresponding features, and photovoltaic power prediction results are obtained through integrated weighting.

[0101] The photovoltaic power robust prediction intelligent agent solution module is used to define the state space and action space based on the constructed photovoltaic power robust prediction environment, and to utilize the injection attack strategy obtained from the solution. To simulate a latent injection attack, a smart agent based on the actor-critic algorithm is constructed by tampering with state variables. The smart agent includes an actor model and a critic model. The actor model interacts with the environment through a policy function and selects action variables based on the state variables. The critic model performs a global value evaluation of the action variables of the actor model. The actor model is pre-trained using the clear sky index membership loss function. The optimal action function is obtained by interacting with the environment and iteratively searching, and the multi-step robust prediction result of photovoltaic power is obtained by solving the problem.

[0102] It should be understood that the photovoltaic power robust prediction system for resisting latent multimodal injection attacks in the embodiments of the present invention can realize all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.

[0103] The present invention also provides a computer device comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the photovoltaic power robust prediction method against latent multimodal injection attacks as described above.

[0104] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the photovoltaic power robust prediction method against latent multimodal injection attacks as described above.

[0105] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), computer devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

Claims

1. A robust photovoltaic power prediction method to resist latent multimodal injection attacks, characterized in that, Includes the following steps: Photovoltaic power measurements, weather forecasts, and ground-based cloud images are used as multimodal data to drive photovoltaic power prediction tasks. Attention moments are constructed using self-attention of each mode and cross-attention of different modes to build an injection attack strategy. The optimal injection attack strategy applied to each mode is obtained by solving for the photovoltaic power prediction loss. ; Based on the characteristics of multimodal data, a robust photovoltaic power prediction environment based on multimodal learning is constructed. For photovoltaic power measurement values, ground-based cloud maps, and weather forecasts, sparse self-attention, spatial pyramid attention, and multilayer perceptron are used to extract corresponding features, and photovoltaic power prediction results are obtained through integrated weighting. Based on the constructed robust photovoltaic power prediction environment, the state space and action space are defined, and the injection attack strategy obtained from the solution is utilized. To simulate a latent injection attack, a smart agent based on the actor-critic algorithm is constructed by tampering with state variables. The smart agent includes an actor model and a critic model. The actor model interacts with the environment through a policy function and selects action variables based on the state variables. The critic model performs a global value evaluation of the action variables of the actor model. The actor model is pre-trained using the clear sky index membership loss function. The optimal action function is obtained by interacting with the environment and iteratively searching, and the multi-step robust prediction result of photovoltaic power is obtained by solving the problem.

2. The method according to claim 1, characterized in that, Photovoltaic power measurements, weather forecasts, and ground-based cloud images are used as multimodal data to drive photovoltaic power prediction tasks. Attention moments are constructed using self-attention of each mode and cross-attention of different modes to build an injection attack strategy. The optimal injection attack strategy applied to each mode is obtained by solving for the photovoltaic power prediction loss. ,include: Multimodal data, including photovoltaic power measurement, weather forecast, and ground-based cloud maps, are characterized as follows: Attention moments are constructed using modal self-attention and cross-attention between different modalities to build an injection attack strategy. The moment estimation method is used to solve the problem and obtain the attack strategies applied to each mode. : in, As a weighting factor applied to various types of attention mechanisms, The result of self-attention calculation for mode i. The result of cross-attention calculation for mode j and mode i; Injection attack strategy The injection attack modulus applied to each modality is obtained. Based on photovoltaic power prediction model A loss function is constructed to train the model, ensuring that the predicted result after the attack deviates as much as possible from the photovoltaic power measurement value at the point to be predicted. : in, Let the root mean square error loss function be . It is an infinite norm. Used to limit the extent of injection attacks. The optimal injection attack strategy.

3. The method according to claim 2, characterized in that, Extracting latent features from cloud images using spatial pyramid attention includes: by To represent ground-based cloud image stream data, spatial pyramid attention is used to extract features from each level of the cloud image. : in, For spatial dimensionality reduction operations, and This indicates that a linear regression layer has been applied; Indicates input Attenuation in spatial dimensions times, This represents the normalization function.

4. The method according to claim 3, characterized in that, The photovoltaic power prediction result obtained through integrated weighting is expressed as follows: in, Let be the weighting factor for the m-th modality data, which will be solved iteratively by the agent and satisfy . ; This represents the feature extracted from the m-th modality data; This represents a robust photovoltaic power prediction environment driven by multimodal data.

5. The method according to claim 4, characterized in that, In the aforementioned agent based on the actor-critic algorithm, the actor model uses a policy function. Interacting with the environment, via state variables Select action variables Its parameters are When the action variable After selection, the next state and instant rewards Obtained from a predefined environment; action variables proposed by the critic model and the actor model. Overall value The gain, whose parameter is ; Selected state variables Including: time window length is Photovoltaic power measurement sequence and cloud map sequence And the length of the time window to be predicted is Weather forecast sequence ; Selected action variables Weighting factors for each modal feature ; Selecting a reward function for , This represents the loss function.

6. The method according to claim 5, characterized in that, The actor model was pre-trained using the clear sky index membership loss function, including: For state variables Select the length of the time window to be predicted. cloud map Estimated clear sky index As the target value, a clear sky index membership loss function is constructed using cross-entropy, and the actor model is pre-trained: ,in, This is the KL divergence calculation method.

7. The method according to claim 5, characterized in that, By utilizing the interaction between the intelligent agent and the environment and iteratively searching to obtain the optimal action function, the multi-step robust prediction results of photovoltaic power are obtained, including: The ultimate goal of solving the actor model is to maximize the cumulative reward function. : ,in, As the attenuation factor, To calculate the number of steps required for the cumulative reward function; Solving the critic model requires maximizing the strategy evaluation function. Represented as: ; Actor model parameters obtained from online learning With critic model parameters The target actor model is updated using a progressive parameter update method. and the target critic model To achieve stable updates of model parameter information through iterative searching by the intelligent agent: ,in, The equilibrium factor is updated for the parameters, thereby obtaining the deterministic action policy function. Multi-step robust prediction results of photovoltaic power are obtained through environmental interaction.

8. A robust photovoltaic power prediction system resistant to latent multimodal injection attacks, characterized in that, include: The injection attack strategy construction module uses photovoltaic power measurements, weather forecasts, and ground-based cloud images as multimodal data to drive photovoltaic power prediction tasks. It constructs attention moments using self-attention of each mode and cross-attention of different modes to build injection attack strategies. The optimal injection attack strategy applied to each mode is obtained by solving for the photovoltaic power prediction loss. ; The photovoltaic power robust prediction environment construction module is used to construct a photovoltaic power robust prediction environment based on multimodal learning according to the characteristics of multimodal data. Specifically, for photovoltaic power measurement values, ground-based cloud maps, and weather forecasts, sparse self-attention, spatial pyramid attention, and multilayer perceptron are used to extract corresponding features, and photovoltaic power prediction results are obtained through integrated weighting. The photovoltaic power robust prediction intelligent agent solution module is used to define the state space and action space based on the constructed photovoltaic power robust prediction environment, and to utilize the injection attack strategy obtained from the solution. To simulate a latent injection attack, a smart agent based on the actor-critic algorithm is constructed by tampering with state variables. The smart agent includes an actor model and a critic model. The actor model interacts with the environment through a policy function and selects action variables based on the state variables. The critic model performs a global value evaluation of the action variables of the actor model. The actor model is pre-trained using the clear sky index membership loss function. The optimal action function is obtained by interacting with the environment and iteratively searching, and the multi-step robust prediction result of photovoltaic power is obtained by solving the problem.

9. A computer device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the photovoltaic power robust prediction method against latent multimodal injection attacks as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the photovoltaic power robust prediction method against latent multimodal injection attacks as described in any one of claims 1-7.

Images