A transformer-based photovoltaic model parameter extraction multi-task optimization method

Abstract

A Transformer-based photovoltaic model parameter extraction multi-task optimization method belongs to the cross field of evolutionary computation and photovoltaic modeling. The method constructs a unified search space, and in the multi-task joint evolution process, a Transformer knowledge generation network based on a multi-head attention mechanism is used to learn the directional mapping relationship from the source task solution to the target task solution, and an adaptive candidate solution for the target task is generated. Through the quality evaluation and sample filtering mechanism, the generated candidate solution is distinguished and screened, and only the high-quality migration individual is injected into the target task population, thereby reducing the negative migration risk and improving the stability and effectiveness of knowledge migration. The method takes into account cross-task knowledge utilization and local search within the task, and can effectively improve the convergence performance and solution accuracy in complex heterogeneous multi-task optimization scenarios. It can be applied to parameter extraction problems of single-diode models, double-diode models and photovoltaic component single-diode models.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of evolutionary computation and photovoltaic modeling, and in particular relates to a multi-task optimization method for extracting photovoltaic model parameters based on Transformer. Background Technology

[0002] With the rapid development of photovoltaic (PV) power generation technology, PV systems have been widely applied in distributed energy, power plant operation, and smart grids. To improve the accuracy of PV system modeling, control, and operation optimization, it is usually necessary to accurately extract key parameters from the PV model. However, the internal parameters of PV models are often difficult to measure directly and usually need to be solved by numerical optimization methods based on measured current-voltage characteristic curves. Since the PV model parameter extraction problem typically has characteristics such as continuity, nonlinearity, multi-peaks, and strong coupling, it places high demands on the global search capability, solution accuracy, and robustness of the optimization algorithm. In existing research, single-diode models, dual-diode models, and single-diode models of PV modules are three commonly used PV models. These models have certain similarities in physical mechanisms and parameter semantics, but differ in parameter dimensions and the degree of nonlinear coupling, thus providing a basis for multi-task collaborative optimization. Traditional PV model parameter extraction methods often adopt a single-task optimization approach, solving different models independently, which makes it difficult to effectively utilize the potential shared information between models, resulting in problems such as high computational overhead, low search efficiency, and susceptibility to local optima. Based on this, evolutionary multitasking optimization has gradually attracted attention, as it improves overall search performance through knowledge transfer between tasks.

[0003] The performance of existing evolutionary multi-task optimization methods largely depends on the design of the knowledge transfer mechanism. Implicit knowledge transfer methods typically exchange information between different tasks through genetic operations such as crossover and mutation. Lim et al. used information entropy to measure search uncertainty and tuned the parameters of the SBX-based unary mutation operator. (TY Lim, CJTan, WP Wong, CP Lim, An information entropy-based evolutionary computation for multi-factorial optimization, Applied Soft Computing 114(2022) 108071.doi:https: / / doi.org / 10.1016 / j.asoc.2021.108071.). These methods are simple to implement and have low computational overhead, but knowledge transfer mainly relies on random genetic operations. When the task correlation is weak or the task heterogeneity is strong, the transfer effect is prone to instability and may even lead to negative transfer. To improve knowledge transfer modeling capabilities, Wang et al. combined variational autoencoders with contrastive learning, utilizing variational autoencoders to learn the mapping relationships between task domains (R.Wang, X. Feng, H. Yu, Contrastive variational autoencoder driven convergenceguidance in evolutionary multitasking, Applied Soft Computing 163 (2024)111883.doi:https: / / doi.org / 10.1016 / j.asoc.2024.111883.). This method can improve the representation alignment between tasks to some extent, but its optimization objective mainly revolves around reconstruction rather than directly promoting the fitness of the target task. Therefore, the generated transfer individuals are prone to problems such as over-smoothing, insufficient solution diversity, and limited ability to generate high-quality candidate solutions. Subsequently, Xue et al. proposed a knowledge transfer framework based on neural networks, which learns bidirectional mapping between task-specific solution spaces through fitness normalization pairing (Z.-F. Xue, Z.-J. Wang, Z.-H. Zhan, S.Kwong, J. Zhang, Neural network-based knowledge transfer for multitaskoptimization, IEEE Transactions on Cybernetics 54 (12) (2024) 7541–7554.).While this method enhances the ability to model task relationships, it focuses more on task-level mapping learning and lacks explicit quality control over the reliability of transfer results. When task heterogeneity is high, problems such as large fluctuations in transfer quality and the misinjection of low-quality candidate solutions into the target population may still occur. Although these studies have made some progress, they still face three challenges: First, implicit transfer methods based on genetic operations such as crossover and mutation are highly random and can easily lead to transfer instability when there are large differences in models. Second, explicit transfer methods based on autoencoders or neural networks still do not fully characterize complex nonlinear task relationships and are difficult to directly generate high-quality candidate solutions that are truly beneficial to the optimization of target model parameters. Third, existing methods lack effective quality control over transfer results, which can easily inject low-quality or misleading individuals into the target population, thereby affecting the accuracy of parameter extraction and the overall optimization stability.

[0004] To address the aforementioned issues, this invention proposes a multi-task optimization method for photovoltaic model parameter extraction based on Transformer. By constructing a knowledge generation network based on a multi-head attention mechanism to learn the directional mapping relationship between different photovoltaic model tasks, and combining quality assessment and sample filtering mechanisms to screen generated candidate solutions, the accuracy, efficiency, and stability of photovoltaic model parameter extraction are improved. Summary of the Invention

[0005] The purpose of this invention is to propose a Transformer-based multi-task optimization method for photovoltaic model parameter extraction, addressing issues such as insufficient modeling capability for complex task relationships, unstable quality of transfer candidate solutions, and susceptibility to negative transfer in existing photovoltaic model parameter extraction methods. This method introduces a knowledge generation network based on a multi-head attention mechanism to learn the directional mapping relationship between source task solutions and target task solutions. Combined with quality assessment and sample filtering mechanisms, the transfer quality is judged before injecting candidate solutions, thereby improving the effectiveness, stability, and robustness of multi-task collaborative search.

[0006] The technical solution of this invention: A multi-task optimization method for extracting photovoltaic model parameters based on Transformer, comprising the following steps: Step 1: Build a system containing A set of multi-task optimization problems for a set of tasks to be optimized is proposed. The objective function for each task is established, and the population of each task is initialized in a unified search space. Fitness evaluation is performed to obtain the current best individual for each task. Furthermore, the multi-task optimization problem can be expressed as: in, For the objective function The value of the decision variable that is minimized. For the first The search space for each task. For the first The objective function of each task For the first The optimal solution for each task. For the first Decision variables for each task.

[0007] Furthermore, the tasks to be optimized include optimizing the single-diode model, the dual-diode model, and the single-diode model of the photovoltaic module; Furthermore, individuals from different tasks are first mapped to a unified search space for encoding and evolution, and then decoded back to their respective original task spaces during fitness evaluation.

[0008] Step 2: During the evolution process, for the target task and the source task, a supervised training sample pair consisting of the elite solution set of the source task and the elite solution set of the target task is constructed according to the fitness ranking of individuals in the population. The Transformer-based knowledge generation network is trained to learn the directional mapping relationship from the source task solution to the target task solution. Furthermore, the elite solution sets of the source task and the elite solution sets of the target task are mapped to latent representations respectively. and And introduce the task embedding vectors of the source task and the target task. and This constitutes enhanced input: in, Indicates the number of training sample pairs; Indicates the feature embedding dimension of the solution; Indicates the task embedding dimension; and These represent the task embedding vectors respectively. and The copied matrix; Represents a set of real numbers 3D matrix space; This represents the matrix concatenation operation.

[0009] Furthermore, a linear mapping is performed on the enhanced input to obtain the query matrix, key matrix, and value matrix: in, Represents the query matrix; Represents the key matrix; Represents a value matrix; , and Let be the trainable parameter matrix. A multi-head attention mechanism is used for mapping, and its output is represented as: in, This represents a multi-head attention mapping; Indicates the number of heads of attention; This indicates that the outputs of multiple attention heads are concatenated column by column; Indicates the output mapping matrix; the first Each attention head is represented as: in, , and They represent the first The query matrix, key matrix, and value matrix corresponding to each attention head; The feature dimension of a single attention head; This represents the normalized exponential function.

[0010] After multi-head attention output, residual connections and layer normalization are introduced to obtain the hidden representation: in, Indicates hidden representation; Presentation layer normalization operation.

[0011] Then, a feedforward neural network is used to generate candidate solutions for the target task: in, , , , For the parameters of the feedforward neural network, Represents the linear rectified activation function. These are the candidate solutions for the generated target task.

[0012] Furthermore, the supervised training samples are constructed by pairing elite individuals from the source task population and the target task population according to their normalized fitness ranking, and the knowledge generation network is trained using a mean squared error loss function, which is: in, This represents the loss function of a knowledge generation network. Indicates the first Generate candidate solutions. This represents the elite solution to the corresponding target task. This represents the L2 norm.

[0013] Step 3: Using the trained knowledge generation network, the candidate solutions in the source task are mapped to generate candidate solutions for the target task; Step 4: Based on the fitness distribution of the current population for the target task, dynamically construct high-quality and low-quality samples, train a quality assessment and sample filtering model, and perform quality discrimination and filtering on the candidate solutions for the target task generated in Step 3 to screen out low-quality migrant individuals. Furthermore, based on the fitness ranking of the current population for the target task, the top-ranked population is selected. Individuals as positive sample set After selecting the ranking Individuals as negative sample set Construct a dynamic training set: in, and These represent the proportions of positive and negative samples, respectively. Indicates the current population size for the target task.

[0014] Furthermore, the quality assessment and sample filtering model employs a feedforward neural network, and its output is: in, The Sigmoid activation function is used. Represents an individual The probability of a high-quality solution. Training is performed using a binary cross-entropy loss function. When candidate solutions are generated satisfy If the candidate solution passes the filter, it is allowed to be injected into the target task population; otherwise, it is discarded. This is the quality discrimination threshold.

[0015] Step 5: Inject the filtered candidate solutions into the target task population, and perform differential evolution operations on each task, including mutation, crossover, and selection, to update the population of each task and the current best individual; Furthermore, when the preset migration generation interval is reached, a source task is selected from the remaining tasks other than the target task. The knowledge generation network is used to generate candidate solutions for the target task. Several high-quality candidate solutions are selected from the candidate solutions filtered in step 4 to replace the inferior individuals in the target task population, thereby completing the elite injection. Subsequently, mutation operations based on difference vectors, crossover operations based on crossover probability, and selection operations based on fitness comparison are performed on each task population to achieve the coordinated implementation of cross-task knowledge transfer and intra-task local search.

[0016] Step 6: Repeat steps 2 to 5 until the maximum number of function evaluations or the maximum number of iterations is reached, and output the optimal solution set for each task.

[0017] Furthermore, the Transformer-based knowledge generation network follows the training cycle. Update training, knowledge transfer, and sample filtering according to the transfer cycle. The knowledge generation process is decoupled from the quality assessment process, allowing the knowledge generation network to learn the directional mapping from the source task to the target task, while the quality assessment and sample filtering models are trained only based on the current population distribution of the target task, thereby reducing the risk of negative transfer and improving the stability of knowledge transfer.

[0018] Compared with the prior art, the present invention has the following advantages: The proposed Transformer-based multi-task optimization method for photovoltaic model parameter extraction enables effective knowledge transfer among multiple related photovoltaic models, improving the globality and synergy of parameter search. By modeling the complex nonlinear relationship between the source and target tasks, this invention can generate candidate solutions that better align with the target task's search direction, thereby improving parameter extraction accuracy. Through quality assessment and sample filtering mechanisms, this invention effectively filters out low-quality transfer individuals, reducing the risk of negative transfer and enhancing the stability and robustness of the optimization process. Since single-diode models, dual-diode models, and photovoltaic module single-diode models are correlated in terms of physical mechanisms and parameter semantics, this invention can fully utilize shared information between models, improving parameter extraction efficiency while providing a more reliable parameter foundation for photovoltaic system modeling, control, and operation optimization. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the present invention; Figure 2 A schematic diagram of a knowledge generation network structure based on Transformer provided by this invention; Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The present invention includes, but is not limited to, the following embodiments.

[0021] Figure 1 The overall flowchart of this invention is shown, specifically including the following steps: defining photovoltaic model parameters, population initialization, fitness evaluation, training of the Transformer-based knowledge generation network, quality assessment and sample filtering, elite injection, differential evolution update, and application examples. 1. Definition of Photovoltaic Model Parameter Extraction Problem In this embodiment, the photovoltaic model parameter extraction problem is taken as the application object of the method of the present invention. The single diode model, the dual diode model, and the single diode model of the photovoltaic module are respectively regarded as three related optimization tasks, denoted as... , and Let the number of experimental sampling points be... , No. The experimental voltage and experimental current corresponding to each sampling point are denoted as follows: and ,in .

[0022] For the task Its decision vector is defined as: in, Indicates photocurrent, Indicates the reverse saturation current of the diode. Indicates series resistance. Indicates parallel resistance. This represents the diode's ideality factor.

[0023] For the task Its decision vector is defined as: in, and These represent the reverse saturation currents of the two diodes, respectively. and These represent the ideal factors of the two diodes, respectively; the meanings of the remaining symbols are the same as their functions. same.

[0024] For the task Its decision vector is defined as: in, Indicates photocurrent, Indicates reverse saturation current. Indicates series resistance. Indicates parallel resistance. This represents the diode's ideality factor.

[0025] The three tasks each use the sum of squared errors between the model-predicted current and the measured current as the optimization objective. The residual function is defined as: in, Indicates thermal voltage; The corresponding objective function is: Task The residual function is defined as: The corresponding objective function is: Task The residual function is defined as: in, Indicates the number of solar cells connected in series; The corresponding objective function is: Furthermore, let the first The optimization objective for each task is expressed as: in, In the search space Internal objective function Decision variables that achieve their minimum value; Indicates the first The original search space for each task; Indicates the first The objective function for each task; Indicates the first The optimal solution for each task.

[0026] 2. Population initialization and fitness assessment For each task The initial size is Population: in, Indicates the first The task is performed in the initial generation of the population. Indicates the first The first task One initial individual; This represents the maximum value of the dimension of the decision variables across all tasks.

[0027] During initialization, the variables in each dimension satisfy: in, Representing an interval A uniform distribution on the surface.

[0028] Decode each individual and calculate the objective function value: in, Indicates the first Task No. The fitness value of the initial individual; For the first The objective function for each task; For the first The decoding function for each task is used to map individuals in the unified search space to the original search space of that task.

[0029] Therefore, we obtain the first... The optimal individual for each task in the initial generation: 3. Construction of supervised sample pairs based on fitness ranking Figure 2 This invention demonstrates a Transformer-based knowledge generation network structure. This network represents the task semantics by introducing task embedding vectors and combines multi-head attention, residual connections, layer normalization, and feedforward neural networks to learn the directional mapping relationship from source task solutions to target task solutions, thereby generating candidate solutions for the target task. In the knowledge transfer stage, let the current target task be... The source task is First, we will analyze the populations separately. and Sort the elite individuals from best to worst fitness, and denote the sorted sequence as: in, Let represent the number of elite samples used for training, and satisfy: Pair elite individuals with the same fitness ranking position in the source task and the target task to construct a set of supervised training sample pairs: in, Indicates the first substitute for the task To the mission The set of supervised training sample pairs.

[0030] 4. Training the Transformer Knowledge Generation Network To learn the mapping relationship between the source task solution and the target task solution, this embodiment employs a Transformer-based knowledge generation network. First, elite samples from both the source and target tasks are mapped to a common embedding space. Let the embedding dimension be... Then we have: in, and It is a linear projection matrix. and Let represent the original dimensions of the source task and the target task, respectively. This indicates the corresponding embedded representation.

[0031] To characterize the semantic differences between tasks, task embedding vectors for the source and target tasks are introduced. ,in This represents the task embedding dimension. The task embedding is copied to the same number of samples and then concatenated with the deembedding to obtain the augmented input: in, and They represent respectively by and The matrix obtained by copying.

[0032] A linear mapping is performed on the augmented input to obtain the query matrix, key matrix, and value matrix: in, , and This is a trainable parameter matrix.

[0033] Let the number of attention heads be The output of multi-head attention is then represented as: Among them, the Each attention head is represented as: in, This represents the feature dimension of a single attention head.

[0034] Following the multi-head attention module, residual connections and layer normalization are introduced to obtain the hidden representation: Then, the hidden representations are mapped back to the target task decision space using a feedforward neural network to obtain the generated candidate solutions: in, , , and For the parameters of the feedforward neural network, This represents the generated candidate solution matrix for the target task.

[0035] The knowledge generation network is trained using supervised learning, and its loss function is defined as mean squared error. in, Indicates the first Generate candidate solutions. This represents the elite solution to the corresponding target task.

[0036] The Adam optimizer is used to update the parameters of the knowledge generation network. The update process can be represented as follows: in, This represents all trainable parameters of the knowledge generation network. Indicates about parameters gradient, Indicates the learning rate; 5. Quality assessment and sample filtering model training After training is complete, the knowledge generation network is used to map the individual source tasks to obtain the set of candidate solutions for the target task: in, Indicates the first The set of candidate solutions for the target task. These represent the generated candidate solutions in the set. This indicates the number of candidate solutions generated.

[0037] To avoid injecting low-quality candidate solutions into the target task population, this embodiment introduces a quality assessment and sample filtering mechanism. Let the current population of the target task be: Their fitness, ranked from best to worst, is denoted as: Before selection Each individual is used as the positive sample set: After selection Individuals are used as a negative sample set: in, Indicates the proportion of positive samples. This represents the proportion of negative samples. Therefore, a dynamic training set for the quality evaluator is constructed: This embodiment uses a lightweight feedforward neural network to construct a target task quality evaluator, whose output is defined as: in, , , and For the evaluator network parameters, The Sigmoid activation function is used. Represents an individual The predicted probability of a high-quality solution.

[0038] The evaluator is trained using a binary cross-entropy loss function: After training, generate a set of candidate solutions. Each individual in the dataset is scored for quality, and then ranked according to a threshold. Filtering is performed to obtain the set of candidate solutions that pass the filter: in, Represents the set of candidate solutions that passed the filter. 6. Elite Injection and Differential Evolution Evaluate the true objective function for the filtered candidate solutions: Select the best from the worst fitness. The candidate solutions constitute the injection set: ,in, This indicates the number of individuals injected. Simultaneously, it selects the individual with the worst fitness from the current population for the target task. Individuals constitute the set to be replaced: The target task population will then be updated as follows after injection: After knowledge transfer and elite injection are completed, differential evolution updates are performed on each task population.

[0039] 7. Parameter setting example In a specific implementation example, the population size for each task is set to N=50, and the maximum number of function evaluations is set to: In differential evolution, the mutation factor and crossover probability are set as follows: The number of attention heads, solution feature embedding dimension, and task embedding dimension in the Transformer knowledge generation network are set as follows: The number of training rounds and the learning rate of the knowledge generation network are set as follows: The parameters in the quality assessment and sample filtering mechanism are set as follows: The training cycle and knowledge transfer cycle of the knowledge generation network are set as follows: The parameters described above are merely one embodiment of the present invention, used to illustrate the feasibility of the method of the present invention, and should not be construed as limiting the scope of protection of the present invention.

[0040] 8. Verify model performance This invention treats the single-diode model (SDM), dual-diode model (DDM), and single-diode model (SMM) of photovoltaic modules as three related optimization tasks, and collaboratively solves these three tasks within a unified search framework. Since the three models share certain similarities in physical mechanisms, objective function forms, and parameter semantics, but differ in parameter dimensions and nonlinear coupling, beneficial search experience can be shared through inter-task knowledge transfer while preserving the unique search characteristics of each task. For a given experimental voltage... and measuring current The parameter extraction process for each photovoltaic model can be described as minimizing the mismatch between the model-predicted current and the measured current. By using the Transformer-based knowledge generation network constructed in this invention to learn the mapping relationship between different photovoltaic model tasks, and combining quality assessment and sample filtering mechanisms to screen transfer candidate solutions, the accuracy and search efficiency of parameter extraction can be effectively improved. Experimental results show that the method of this invention can obtain better or more competitive objective function values ​​in this type of real-world engineering scenario, verifying its effectiveness and robustness in the photovoltaic model parameter extraction problem.

[0041] In summary, this embodiment details the complete implementation process of a multi-task optimization method for photovoltaic model parameter extraction based on Transformer proposed in this invention. This method utilizes Transformer to establish a nonlinear mapping relationship between the source task solution and the target task solution, and combines quality assessment and sample filtering mechanisms to screen candidate solutions for migration, thereby reducing the risk of negative migration while ensuring the effectiveness of migration. It is suitable for solving complex heterogeneous multi-task optimization problems.

[0042] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A multi-task optimization method for extracting photovoltaic model parameters based on Transformer, characterized in that, The steps are as follows: Step 1: Build a system containing A set of multi-task optimization problems for a set of tasks to be optimized is proposed. The objective function for each task is established, and the population of each task is initialized in a unified search space. Fitness evaluation is performed to obtain the current best individual for each task. Step 2: During the evolution process, for the target task and the source task, a supervised training sample pair consisting of the elite solution set of the source task and the elite solution set of the target task is constructed according to the fitness ranking of individuals in the population. The Transformer-based knowledge generation network is trained to learn the directional mapping relationship from the source task solution to the target task solution. Step 3: Using the trained knowledge generation network, the candidate solutions in the source task are mapped to generate candidate solutions for the target task; Step 4: Based on the fitness distribution of the current population for the target task, dynamically construct high-quality and low-quality samples, train a quality assessment and sample filtering model, and perform quality discrimination and filtering on the candidate solutions for the target task generated in Step 3 to screen out low-quality migrant individuals. Step 5: Inject the filtered candidate solutions into the target task population, and perform differential evolution operations on each task, including mutation, crossover, and selection, to update the population of each task and the current best individual; Step 6: Repeat steps 2 to 5 until the maximum number of function evaluations or the maximum number of iterations is reached, and output the optimal solution set for each task.

2. The multi-task optimization method for photovoltaic model parameter extraction based on Transformer according to claim 1, characterized in that, In step 1, the multi-task optimization problem is expressed as: in, For the objective function The value of the decision variable that is minimized. For the first The search space for each task. For the first The objective function of each task For the first The optimal solution for each task. For the first Decision variables for each task.

3. The multi-task optimization method for photovoltaic model parameter extraction based on Transformer according to claim 2, characterized in that, In step 1, the task to be optimized includes optimizing the single-diode model, the dual-diode model, and the single-diode model of the photovoltaic module; Individuals with different tasks are first mapped to a unified search space for encoding and evolution, and then decoded back to their respective original task spaces during fitness evaluation.

4. The multi-task optimization method for photovoltaic model parameter extraction based on Transformer according to claim 3, characterized in that, In step 2, Map the elite solution set of the source task and the elite solution set of the target task to latent representations respectively. and And introduce the task embedding vectors of the source task and the target task. and This constitutes enhanced input: in, Indicates the number of training sample pairs; Indicates the feature embedding dimension of the solution; Indicates the task embedding dimension; and These represent the task embedding vectors respectively. and The copied matrix; Represents a set of real numbers 3D matrix space; This represents the matrix concatenation operation.

5. The multi-task optimization method for photovoltaic model parameter extraction based on Transformer according to claim 4, characterized in that, In step 2, Perform a linear mapping on the enhanced input to obtain the query matrix, key matrix, and value matrix: in, Represents the query matrix; Represents the key matrix; Represents a value matrix; , and The parameter matrix is ​​trainable; a multi-head attention mechanism is used for mapping, and its output is represented as: in, This represents a multi-head attention mapping; Indicates the number of heads of attention; This indicates that the outputs of multiple attention heads are concatenated column by column; Indicates the output mapping matrix; the first Each attention head is represented as: in, , and They represent the first The query matrix, key matrix, and value matrix corresponding to each attention head; The feature dimension of a single attention head; This represents the normalized exponential function; After multi-head attention output, residual connections and layer normalization are introduced to obtain the hidden representation: in, Indicates hidden representation; Presentation layer normalization operation; Then, a feedforward neural network is used to generate candidate solutions for the target task: in, , , , For the parameters of the feedforward neural network, Represents the linear rectification activation function. These are the candidate solutions for the generated target task.

6. The multi-task optimization method for photovoltaic model parameter extraction based on Transformer according to claim 5, characterized in that, In step 2, The supervised training samples are constructed by pairing elite individuals from the source task population and the target task population one-to-one according to their normalized fitness ranking, and the knowledge generation network is trained using a mean squared error loss function. The loss function is: in, This represents the loss function of a knowledge generation network. Indicates the first Generate candidate solutions. This represents the elite solution to the corresponding target task. This represents the L2 norm.

7. The multi-task optimization method for photovoltaic model parameter extraction based on Transformer according to claim 6, characterized in that, In step 4, Based on the fitness ranking of the current population for the target task, select the top-ranked... Individuals as positive sample set After selecting the ranking Individuals as negative sample set Construct a dynamic training set: in, and These represent the proportions of positive and negative samples, respectively. Indicates the current population size for the target task.

8. The multi-task optimization method for photovoltaic model parameter extraction based on Transformer according to claim 7, characterized in that, In step 4, The quality assessment and sample filtering model uses a feedforward neural network, and its output is: in, For the Sigmoid activation function, Represents an individual The probability of a high-quality solution; training is performed using a binary cross-entropy loss function. When candidate solutions are generated satisfy If the candidate solution passes the filter, it is allowed to be injected into the target task population; otherwise, it is discarded. This is the quality discrimination threshold.

9. A multi-task optimization method for photovoltaic model parameter extraction based on Transformer according to claim 8, characterized in that, In step 5, When the preset migration interval is reached, a source task is selected from the remaining tasks other than the target task. The knowledge generation network is used to generate candidate solutions for the target task. Several high-quality candidate solutions are selected from the candidate solutions filtered in step 4 to replace the inferior individuals in the target task population, so as to complete the elite injection. Subsequently, mutation operation based on difference vector, crossover operation based on crossover probability, and selection operation based on fitness comparison are performed on each task population to realize the coordinated implementation of cross-task knowledge transfer and intra-task local search.

10. A multi-task optimization method for photovoltaic model parameter extraction based on Transformer according to claim 9, characterized in that, In step 6, the Transformer-based knowledge generation network is trained according to the training cycle. Update training, knowledge transfer, and sample filtering according to the transfer cycle. The knowledge generation process is decoupled from the quality assessment process, allowing the knowledge generation network to learn the directional mapping from the source task to the target task, while the quality assessment and sample filtering models are trained only based on the current population distribution of the target task, thereby reducing the risk of negative transfer and improving the stability of knowledge transfer.

Images