Multi-scene transfer learning method and system of electric power artificial intelligence algorithm model

Abstract

The invention discloses a multi-scene transfer learning method and system for an electric power artificial intelligence algorithm model, and the method comprises the steps: collecting the operation data of a plurality of electric power scenes, and carrying out the preprocessing of the operation data, and forming samples corresponding to different electric power scenes; generating an electric power artificial intelligence algorithm model used for representing the association relationship between the electric power scenes; initializing structural parameters of a pre-constructed memory network model by adopting a parrot optimization algorithm in which a memory feedback mechanism is introduced; simulating imitation behaviors of parrot individuals by utilizing an electric power artificial intelligence algorithm model, and iteratively optimizing structural parameters of a memory network model by taking migration precision, task execution stability and generalization ability of the memory network model in a source domain and a target domain under a plurality of electric power scenes as fitness; and selecting structural parameters of the memory network model with the highest fitness, applying the structural parameters to a target power scene, and completing knowledge migration and task adaptation through migration learning and meta learning.

Description

Technical Field

[0001] This invention belongs to the field of transfer learning technology, and in particular relates to a multi-scenario transfer learning method and system for an artificial intelligence algorithm model for power. Background Technology

[0002] The intelligentization and automation of power systems have become crucial means to improve power operation efficiency, optimize resource allocation, and ensure power system stability. With the rapid development of information technology and big data, artificial intelligence-based power algorithm models are increasingly being widely applied, particularly in areas such as power load forecasting, equipment fault detection, meteorological data analysis, and energy management. Traditional power AI algorithms largely rely on large-scale datasets for training to achieve model accuracy and robustness. However, in practical applications, especially in some power application scenarios, data scarcity and insufficient sample sizes frequently occur, directly impacting the performance of power AI algorithm models. This is particularly true when the data volume is small or the task types are diverse, significantly reducing the model's training effectiveness. To address this issue, researchers and engineers have proposed various techniques, among which transfer learning and meta-learning methods are widely considered effective solutions.

[0003] Transfer learning is a technique that uses knowledge gained in one domain to aid learning in another. In the power industry, the application of transfer learning can overcome the problems of insufficient data and poor model generalization ability across different scenarios. Specifically, transfer learning can improve the performance of the target domain model by transferring knowledge from one power scenario to another. Especially when the amount of data in the target power task scenario is very limited, transfer learning can effectively borrow knowledge gained from the source domain task to compensate for the lack of data. However, the practical application of transfer learning in power applications also faces some challenges. First, the differences between power tasks are significant, and the similarity between the source and target domains may not be significant, thus limiting the effectiveness of transfer learning. Second, because data in power systems has time-series characteristics, traditional transfer learning methods often fail to effectively capture the temporal features of the data, thereby affecting the accuracy of the model.

[0004] Meta-learning, complementing transfer learning, focuses on enabling models to quickly adapt to new tasks across multiple tasks. The core idea of ​​meta-learning is to summarize learning experiences from multiple tasks and optimize the model's learning strategy, allowing it to adapt more rapidly to new tasks. In the power industry, meta-learning can improve the model's generalization ability across different power task scenarios. Specifically, by training the model on multiple power tasks and gaining cross-task experience, meta-learning enables the model not only to adapt to the target power task scenario but also to transfer knowledge between multiple power task scenarios. The advantage of meta-learning lies in its ability to handle tasks with limited data samples; by sharing knowledge across multiple tasks, it significantly improves the model's training efficiency and accuracy. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a multi-scenario transfer learning method and system for power artificial intelligence algorithm models.

[0006] The technical solution of the present invention is as follows:

[0007] A multi-scenario transfer learning method for an artificial intelligence algorithm model in the power industry, comprising:

[0008] Collect operational data from multiple power scenarios and preprocess the operational data to form samples corresponding to different power scenarios;

[0009] Based on graph structure, each power scenario sample is transformed into a node in the graph structure, and connection edges are established through feature similarity between nodes to generate a power artificial intelligence algorithm model for representing the relationship between power scenarios.

[0010] The parrot optimization algorithm with a memory feedback mechanism is used to initialize the structural parameters of a pre-built memory network model. The memory network model is used to process temporal information and historical data. Each parrot represents a set of structural parameters of the memory network model.

[0011] Using an artificial intelligence algorithm model for power systems, the imitation behavior of an individual parrot is simulated. The migration accuracy, task execution stability, and generalization ability of the memory network model in the source and target domains under multiple power scenarios are used as the fitness to iteratively optimize the structural parameters of the memory network model.

[0012] The structural parameters of the memory network model with the highest fitness are selected and applied to the target power scenario. Knowledge transfer and task adaptation are completed through transfer learning and meta-learning.

[0013] Furthermore, the operational data includes load data, equipment status, meteorological data, and historical operational records.

[0014] Furthermore, the specific method for generating an artificial intelligence algorithm model for power scenarios, based on a graph structure, transforming samples of each power scenario into nodes in the graph structure, and establishing connection edges through feature similarity between nodes, includes:

[0015] Each power scenario sample is transformed into a node in a graph structure, and the node features are represented in the form of feature vectors; the feature vectors are composed of load data, equipment status and meteorological information from the operational data.

[0016] Calculate the similarity between nodes, use cosine similarity to measure the similarity between nodes, establish the relationship between nodes, and establish a connection edge between each pair of nodes;

[0017] Based on the similarity between nodes, an artificial intelligence algorithm model for the power industry is generated. Each power scenario sample is represented by nodes in a graph structure, and the relationships between power scenarios are established through connecting edges between nodes.

[0018] Based on the power artificial intelligence algorithm model, the set of neighboring nodes for each node is calculated, and the weighted average of the node feature representations corresponding to the neighboring nodes is calculated.

[0019] By combining the node feature representation of the current node with the weighted average of the neighboring nodes, an updated node feature representation is generated, thereby optimizing the power artificial intelligence algorithm model.

[0020] Furthermore, the specific method for initializing the structural parameters of the pre-built memory network model using the parrot optimization algorithm with an introduced memory feedback mechanism includes:

[0021] The parrot optimization algorithm is used to initialize the structural parameters of the pre-built memory network model. A set of structural parameters is assigned to each parrot individual, and each parrot individual represents a set of structural parameters of the memory network model.

[0022] In the parrot optimization algorithm, a memory feedback mechanism is introduced to optimize the structural parameters of the memory network model. The structural parameters of each parrot are adjusted according to samples corresponding to different power scenarios, as shown below:

[0023] ;

[0024] in, Let be the structural parameters of parrot individual i after the (t+1)th iteration; Let be the structural parameters of parrot individual i at the t-th iteration; i is the parrot individual index in the parrot optimization algorithm; This is the learning rate adjustment factor; Let i be the set of neighboring individuals of parrot individual i; For parrot individual i; The fitness of parrot individual k; This is the disturbance intensity coefficient; Let be the modulus of the fitness vector of parrot individual i; Let be the modulus of the fitness vector of parrot individual k; Let i be the set of non-neighbor individuals of parrot individual i; Let be the structural parameters of parrot individual j at the t-th iteration; k be the neighboring individuals of parrot individual i; j be the non-neighboring individuals of parrot individual i; t be the number of iterations used to adjust the structural parameters of each parrot individual. The difference in structural parameters between parrot individual i and parrot individual j at the t-th iteration, The task adaptability coefficient. Contribution coefficient;

[0025] By using the adjusted structural parameters of each individual parrot, the amount of information stored in the memory network model is adaptively adjusted.

[0026] By utilizing the adjusted structural parameters of each individual parrot, important historical operation records are dynamically selected for storage and retrieval based on changes in equipment status, thereby updating the memory network model.

[0027] Furthermore, the specific method for iteratively optimizing the structural parameters of the memory network model by using an electric power artificial intelligence algorithm model to simulate the imitation behavior of an individual parrot, and using the transfer accuracy, task execution stability, and generalization ability of the memory network model in multiple power scenarios as fitness, includes:

[0028] Using an electric power artificial intelligence algorithm model, the imitation behavior of individual parrots is simulated, and the structural parameters of the memory network model are optimized, so that low-fitness parrots can imitate the structural parameters of high-fitness parrots.

[0029] By measuring the transfer accuracy, task execution stability, and generalization ability of the memory network model in multiple power scenarios, the final fitness of each parrot individual in the source and target domains is calculated.

[0030] Furthermore, the method for using an electric power artificial intelligence algorithm model to simulate the imitation behavior of individual parrots and optimize the structural parameters of the memory network model, wherein the structural parameters of the low-fitness parrot individual imitating the high-fitness parrot individual include:

[0031] Using an artificial intelligence algorithm model for electricity, the parrot's imitation behavior is simulated in each round of optimization, and information is transmitted based on the fitness and structural parameters of its neighbors.

[0032] A mimicry factor is defined for the mimicry operation. The mimicry factor determines the degree to which a low-fitness parrot individual imitates a high-fitness parrot individual. The fitness of a low-fitness parrot individual is between 0 and 0.3, while the fitness of a high-fitness parrot individual is between 0.7 and 1. The low-fitness parrot individual will imitate the structural parameters based on the fitness differences of its neighbors.

[0033] Low-fit parrot individuals optimize the structural parameters of the memory network model by continuously imitating high-fit parrot individuals:

[0034] ;

[0035] in, Individuals with low fitness; l is Index of individual parrots; These are weighting coefficients; For the neighboring individual The fitness of; To Summing the fitness of all neighboring individuals; For the neighboring individual In the The structure parameters at the next iteration; For low-fitness parrot individuals In the The structure parameters at the next iteration;

[0036] When imitating, parrots with low fitness will choose individuals with higher fitness from their neighbors as their imitation targets.

[0037] Through multiple iterations, low-fit parrot individuals gradually approached the structural parameters of high-fit parrot individuals.

[0038] Furthermore, the specific method for calculating the final fitness of each parrot individual in the source and target domains by measuring the transfer accuracy, task execution stability, and generalization ability of the memory network model under multiple power scenarios includes:

[0039] The transfer accuracy of the memory network model in the source and target domains is calculated. The transfer accuracy is the ratio of the number of correctly executed samples in the target domain task to the total number of samples in the target domain task. The value of the transfer accuracy ranges from 0 to 1, where 1 indicates that all samples are correctly classified.

[0040] Calculate the task execution stability, which is the average absolute deviation of the performance fluctuation of the task over multiple executions. The task execution performance is the performance value at each execution, and the number of executions is the total number of times the task is executed.

[0041] The generalization ability of the memory network model is calculated. The generalization ability is the difference between the accuracy on the training set and the accuracy on the test set. The generalization ability value ranges from 0 to 1, where 0 means that the accuracy on the training set and the test set are exactly equal, and 1 means that the difference between the accuracy on the training set and the test set is the largest.

[0042] The fitness of each parrot is calculated by combining transfer accuracy, task execution stability and generalization ability. The fitness is composed of a weighted sum of transfer accuracy, task execution stability and generalization ability, and the weighting coefficient is used to balance the contribution of the three factors.

[0043] A transfer accuracy of 0.8 or higher indicates a high execution accuracy of the task in the target domain; a task execution stability of less than 0.1 indicates small fluctuations during execution; and a generalization ability of less than 0.05 indicates small differences in performance under different power scenarios.

[0044] Furthermore, the specific method for selecting the structural parameters of the memory network model with the highest fitness includes:

[0045] The final fitness of each parrot individual in the source and target domains is evaluated, and the final fitness of each parrot individual represents the final fitness of the parrot individual in the source and target domains:

[0046] ;

[0047] in, For the final fitness of an individual parrot, The total number of individual parrots. , , These are the 1st, 2nd, and 3rd weighting coefficients, respectively. Let be the migration accuracy of the i-th parrot individual. For the task execution stability of the i-th parrot individual, Let represent the generalization ability of the i-th parrot individual. , , The first, second, and third hyperparameters are used to adjust the impact of migration accuracy, task execution stability, and generalization ability on fitness, respectively.

[0048] All parrot individuals are sorted according to their final fitness, and the top five parrot individuals with the highest final fitness are selected as the population memory. The population memory stores the memory network structure parameters corresponding to the five selected parrot individuals, and the weights of the final memory network structure are obtained.

[0049] The memory network structure parameters of the top five parrot individuals with the highest fitness are compared with those of the remaining parrot individuals to select the optimal structure combination, ensuring that the structure performs optimally in both the source and target domains. The optimal memory network structure represents the weight of the optimal memory network structure.

[0050] The optimal memory network structure is output as the optimized memory network model and then output.

[0051] Furthermore, the specific methods for applying knowledge transfer and task adaptation to the target power scenario through transfer learning and meta-learning include:

[0052] The optimal memory network structure of the optimized memory network model is adjusted to adapt to the feature distribution of the target power task scenario;

[0053] The transfer learning method is used to transfer knowledge from the source domain to the target domain. By leveraging the similarity between the source and target domains, knowledge from the source domain is transferred to the target domain, thereby optimizing the task performance in the target domain and overcoming the small sample problem in the power industry through transfer learning.

[0054] Based on the requirements of the target power scenario, the optimized memory network model is fine-tuned. The structural parameters of the optimized memory network model are adjusted by minimizing the loss function of the target domain to make up for the shortcomings in the small sample power scenario.

[0055] The fine-tuned memory network model was further optimized using meta-learning methods to ensure its adaptability to different power task scenarios and knowledge sharing across multiple power tasks.

[0056] ;

[0057] in, The adjusted fitness; The number of samples for the target task; For the first The true labels of each target task sample; For the first The predicted output for each target task sample; For migration fitness weighting coefficients; To maintain the weighting coefficients of source domain knowledge; This refers to the number of structural parameters of the memory network model; For the first in the target domain Each memory network model structural parameters, For the first in the source domain Each memory network model structural parameters, For each sample in the target power task dataset, This serves as an index for the various structural parameters of the memory network model;

[0058] In the target power scenario, the optimized memory network model is used to perform the task, evaluate the performance of the task, and output the evaluation results.

[0059] Based on the evaluation results, the memory network model was further optimized, and its structural parameters were adjusted to complete knowledge transfer and task adaptation across multiple power scenarios.

[0060] A multi-scenario transfer learning system for an artificial intelligence algorithm model for power systems includes a data acquisition and preprocessing module, an artificial intelligence algorithm model generation module, an initialization module, an iterative optimization module, and a transfer learning and meta-learning module.

[0061] The data acquisition and preprocessing module is used to acquire operational data from multiple power scenarios and preprocess the operational data to form samples corresponding to different power scenarios.

[0062] The power artificial intelligence algorithm model generation module is used to transform the samples of each power scenario into nodes in the graph structure based on the graph structure, and to establish connecting edges through the feature similarity between nodes to generate a power artificial intelligence algorithm model that represents the relationship between power scenarios.

[0063] The initialization module is used to initialize the structural parameters of a pre-built memory network model using a parrot optimization algorithm that incorporates a memory feedback mechanism. The memory network model is used to process temporal information and historical data, and each parrot represents a set of structural parameters of the memory network model.

[0064] The iterative optimization module is used to simulate the imitation behavior of an individual parrot using an electric power artificial intelligence algorithm model. The structural parameters of the memory network model are iteratively optimized by using the migration accuracy, task execution stability and generalization ability of the memory network model in the source and target domains under multiple power scenarios as fitness.

[0065] The transfer learning and meta-learning module is used to select the structural parameters of the memory network model with the highest fitness and apply them to the target power scenario, thereby completing knowledge transfer and task adaptation through transfer learning and meta-learning.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] This invention proposes a multi-scenario transfer learning method and system for power artificial intelligence algorithm models. Combining the Parrot Optimization Algorithm, a memory network model, and the power artificial intelligence algorithm model, this invention provides users with accurate multi-scenario transfer learning. By integrating transfer learning, the Parrot Optimization Algorithm, the memory network model, and meta-learning, this invention addresses the adaptability and data scarcity issues of multi-scenario tasks in power systems, proposing an innovative multi-scenario transfer learning method that significantly improves the performance of power artificial intelligence algorithm models in multiple power tasks. In practical applications, tasks such as power load forecasting and equipment fault detection often face the problems of data scarcity and scenario differences, which directly leads to the shortcomings of traditional deep learning methods, such as low accuracy and long training time, during training and inference. By adopting the multi-scenario transfer learning method of this invention, these challenges can be effectively overcome.

[0068] This invention utilizes transfer learning technology to transfer knowledge from the source domain to the target domain in power systems, thereby reducing reliance on large amounts of labeled data. In tasks such as power load forecasting and equipment fault detection, especially in scenarios with limited sample sizes, the system demonstrates strong adaptability. Transfer learning allows knowledge accumulated in the source domain to be directly applied to the target domain, significantly improving the learning efficiency and accuracy of the target task and solving the problem of traditional methods requiring large amounts of sample data. Secondly, the introduction of meta-learning enables the system to share knowledge across multiple tasks, optimizing the learning process in power tasks and allowing the model to quickly adapt and achieve better performance when facing new tasks. In this way, the system's cross-task adaptability is greatly enhanced, especially for the numerous changes and uncertainties present in power systems, enabling rapid adjustments and ensuring stable task execution results.

[0069] This invention also incorporates the Parrot Optimization algorithm, which optimizes low-fitness individuals by mimicking the structural parameters of high-fitness individuals, thereby achieving rapid convergence of model parameters. The addition of the Parrot Optimization algorithm not only improves the optimization efficiency of the system but also avoids overfitting or underfitting problems that may occur in traditional methods, ensuring the robustness and stability of the model. Furthermore, this invention adaptively adjusts the structure and weights of the memory network, enabling the model to handle more complex and dynamic data in power systems.

[0070] This invention addresses the common problems of insufficient data and poor task adaptability in the power sector, significantly improving the performance of power artificial intelligence algorithm models. In power load forecasting tasks, applying the method of this invention substantially reduces prediction errors, improving the accuracy and stability of task execution. In equipment fault detection, the increased accuracy enables the system to perform more reliable real-time monitoring and fault early warning, reducing the incidence of equipment failures and ensuring the stable operation of the power system. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the multi-scenario transfer learning method of the power artificial intelligence algorithm model of the present invention. Detailed Implementation

[0072] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0073] Example 1:

[0074] This invention provides a multi-scenario transfer learning method for an artificial intelligence algorithm model in the power sector, such as... Figure 1 As shown, it includes:

[0075] S1. Collect operational data from multiple power scenarios and preprocess the operational data to form samples corresponding to different power scenarios;

[0076] S2. Based on the graph structure, the samples of each power scenario are transformed into nodes in the graph structure, and the connection edges are established through the feature similarity between nodes to generate a power artificial intelligence algorithm model for representing the relationship between power scenarios.

[0077] S3. The parrot optimization algorithm with a memory feedback mechanism is used to initialize the structural parameters of the pre-built memory network model. The memory network model is used to process temporal information and historical data. Each parrot represents a set of structural parameters of the memory network model.

[0078] Using an artificial intelligence algorithm model for the power industry, the imitation behavior of individual parrots is simulated. The migration accuracy, task execution stability, and generalization ability of the memory network model in the source and target domains under multiple power scenarios are used as fitness to iteratively optimize the structural parameters of the memory network model.

[0079] In this embodiment, an artificial intelligence algorithm model for the power industry is used to simulate the imitative behavior of an individual parrot. The specific method for iteratively optimizing the structural parameters of the memory network model, using the transfer accuracy, task execution stability, and generalization ability of the memory network model in multiple power scenarios as fitness, includes:

[0080] S4. Using an electric power artificial intelligence algorithm model, simulate the imitation behavior of individual parrots, optimize the structural parameters of the memory network model, and enable low-fitness parrots to imitate the structural parameters of high-fitness parrots.

[0081] S5. By measuring the transfer accuracy, task execution stability, and generalization ability of the memory network model in multiple power scenarios, calculate the final fitness of each parrot individual in the source and target domains.

[0082] S6. Select the structural parameters of the memory network model with the highest fitness and apply them to the target power scenario. Knowledge transfer and task adaptation are completed through transfer learning and meta-learning.

[0083] This invention achieves accurate task adaptation and knowledge transfer in the application of power artificial intelligence algorithm models through multi-scenario transfer learning. By collecting and preprocessing operational data from multiple power scenarios / application scenarios, diverse training data is provided to the model, thereby enhancing its generalization ability. The memory network model built based on this data generates a graph structure of the power system through node similarity, effectively capturing the correlation information between various power scenarios. The introduction of the parrot optimization algorithm ensures that the initial settings of the memory network structure can efficiently adapt to different scenarios. The model optimizes structural parameters by simulating the imitation behavior of individual parrots, effectively improving the performance of low-fitness individuals and gradually optimizing the network structure. The model comprehensively evaluates the fitness of each individual parrot based on task transfer accuracy, execution stability, and generalization ability, ensuring that the optimal network structure is used as the final output. By applying the optimized memory network model to the target power task scenario, combined with source domain knowledge transfer and meta-learning methods, the model can quickly adapt to new tasks and maintain efficient task execution capabilities, achieving efficient knowledge transfer and adaptation across task scenarios.

[0084] In this embodiment, the operational data includes load data, equipment status, meteorological data, and historical operational records.

[0085] This invention provides multi-dimensional information input for power system models by collecting operational data from various power application scenarios, including power load, equipment status, meteorological data, and historical operating records. Through comprehensive analysis of this data, the model can gain a complete understanding of the power system's operational status and environmental factors, thereby improving task adaptability and prediction accuracy. Utilizing this multi-source data, the system can accurately capture the characteristics and changing patterns of the power system in different scenarios, providing strong data support for tasks such as power load forecasting and equipment fault detection. Simultaneously, this data integration method effectively enhances the model's generalization ability, ensuring the system's flexibility and efficiency across different power scenarios.

[0086] In this embodiment, the memory network structure parameters include memory cell capacity, read / write gating mechanism, attention mechanism coefficient, and learning rate.

[0087] This invention significantly improves the learning efficiency and accuracy of models in power task scenarios by optimizing memory network structure parameters, including memory unit capacity, read / write gating mechanism, attention mechanism coefficients, and learning rate. Adjusting the memory unit capacity enables the model to effectively store and utilize long-term information, while the read / write gating mechanism ensures reasonable control of information flow, improving the model's ability to process time-series data. The introduction of attention mechanism coefficients further enhances the model's focus on key data, enabling it to concentrate on important features and thus improve task accuracy. Optimizing the learning rate accelerates the model's convergence process and reduces training time. The effective configuration of these parameters allows the model to adapt more quickly and provide high-precision predictions and decisions when facing complex power tasks.

[0088] In this embodiment, based on a graph structure, each power scenario sample is represented as a node in the graph structure, and connection edges are established through the feature similarity between nodes to generate a power artificial intelligence algorithm model for representing the relationship between power scenarios. The specific method includes:

[0089] S21. Transform the samples of each power scenario into nodes in a graph structure. The nodes are represented by feature vectors. The feature vectors are composed of load data, equipment status and meteorological information from the operational data.

[0090] S22. Calculate the similarity between nodes, use cosine similarity to measure the similarity between nodes, establish the relationship between nodes, and establish a connection edge between each pair of nodes;

[0091] S23. Based on the similarity between nodes, generate an artificial intelligence algorithm model for the power industry. Each power scenario sample is represented by nodes in a graph structure, and the relationship between power scenarios is established between nodes through connecting edges.

[0092] S24. Based on the power artificial intelligence algorithm model, calculate the set of neighboring nodes for each node, and calculate the weighted average of the node feature representations corresponding to the neighboring nodes;

[0093] S25. Combine the node feature representation of the current node with the weighted average value of the neighboring nodes to generate an updated node feature representation, thereby optimizing the power artificial intelligence algorithm model.

[0094] This invention utilizes preprocessed operational data from multiple power scenarios, representing samples of each scenario as nodes using a graph structure. A power AI algorithm model is generated by calculating the cosine similarity between nodes, making knowledge transfer between tasks more efficient. By weighted averaging of neighbor node features and fusing features from the current node, the optimized node feature representation improves the model's accuracy and adaptability, ensuring stability and efficiency in multi-scenario power tasks.

[0095] In this embodiment, the specific method for initializing the structural parameters of the pre-built memory network model using the parrot optimization algorithm with an introduced memory feedback mechanism includes:

[0096] S31. The parrot optimization algorithm is used to initialize the structural parameters of the pre-built memory network model. A set of structural parameters is assigned to each parrot individual, and each parrot individual represents a set of structural parameters of the memory network model.

[0097] S32. In the parrot optimization algorithm, a memory feedback mechanism is introduced to optimize the structural parameters of the memory network model. The structural parameters of each parrot are adjusted according to samples corresponding to different power scenarios, as shown below:

[0098] ;in, Let be the structural parameters of parrot individual i after the (t+1)th iteration; Let be the structural parameters of parrot individual i at the t-th iteration; i is the parrot individual index in the parrot optimization algorithm; This is the learning rate adjustment factor; Let i be the set of neighboring individuals of parrot individual i; For parrot individual i; The fitness of parrot individual k; This is the disturbance intensity coefficient; Let be the modulus of the fitness vector of parrot individual i; Let be the modulus of the fitness vector of parrot individual k; Let i be the set of non-neighbor individuals of parrot individual i; Let be the structural parameters of parrot individual j at the t-th iteration; k be the neighboring individuals of parrot individual i; j be the non-neighboring individuals of parrot individual i; t be the number of iterations used to adjust the structural parameters of each parrot individual. The difference in structural parameters between parrot individual i and parrot individual j at the t-th iteration, The task adaptability coefficient. Contribution coefficient;

[0099] S33. Using the adjusted structural parameters of each individual parrot, adaptively adjust the amount of stored information in the memory network model;

[0100] S34. Using the adjusted structural parameters of each individual parrot, dynamically select important historical operation records for storage and retrieval based on changes in equipment status, and update the memory network model.

[0101] This invention optimizes the memory network model using the Parrot Optimization Algorithm. During initialization, a unique parameter set is assigned to each individual parrot, and a memory feedback mechanism is introduced to achieve dynamic adjustment based on power scenario operation data. In each iteration, the method precisely optimizes the memory network structure based on the fitness differences between neighboring and non-neighboring individuals. By adaptively adjusting the amount of information stored in the memory network model, the network can effectively handle dynamically changing power system data and intelligently select key historical data for storage and retrieval based on changes in equipment status, optimizing the storage and update efficiency of the network model. This process improves the model's adaptability and ensures stability and efficiency in complex power tasks.

[0102] In this embodiment, an artificial intelligence algorithm model based on electricity is used to simulate the imitation behavior of individual parrots, and the structural parameters of the memory network model are optimized. The specific method for a low-fitness parrot individual to imitate the structural parameters of a high-fitness parrot individual includes:

[0103] S41. Using an electric power artificial intelligence algorithm model, in each round of optimization, simulate the imitation behavior of individual parrots and transmit information based on the fitness and structural parameters of neighboring individuals.

[0104] S42. Define an imitation factor for the imitation operation. The imitation factor determines the degree to which a low-fit parrot individual imitates a high-fit parrot individual. The fitness of a low-fit parrot individual is between 0 and 0.3, while the fitness of a high-fit parrot individual is between 0.7 and 1. The low-fit parrot individual will imitate the structural parameters based on the fitness differences of its neighbors.

[0105] S43. Low-fitness parrot individuals optimize the structural parameters of the memory network model by continuously imitating high-fitness parrot individuals:

[0106] ;

[0107] in, Individuals with low fitness; l is Index of individual parrots; These are weighting coefficients; For the neighboring individual The fitness of; To Summing the fitness of all neighboring individuals; For the neighboring individual In the The structure parameters at the next iteration; For low-fitness parrot individuals In the The structure parameters at the next iteration;

[0108] S44. When imitating, a parrot with low fitness will choose a neighboring individual with higher fitness as its imitation object.

[0109] S45. After multiple iterations, the structural parameters of low-fit parrot individuals gradually approach those of high-fit parrot individuals.

[0110] This invention simulates the imitative behavior of individual parrots, transmitting information based on the fitness of neighboring individuals and the structural parameters of the memory network in each round of optimization. The introduction of the imitation factor determines the degree to which low-fitness individuals learn from high-fitness individuals, effectively improving the performance of low-fitness individuals. By imitating the network structure of higher-fitness individuals, low-fitness individuals gradually optimize the structural parameters of their memory network model. After multiple iterations, the structure of low-fitness individuals gradually approaches that of high-fitness individuals, significantly improving the model's adaptability and accuracy in multi-scenario tasks. This optimization process ensures that each parrot individual can quickly adjust and optimize its network structure according to task requirements, thereby improving the stability and efficiency of the power artificial intelligence algorithm model.

[0111] In this embodiment, the specific method for calculating the final fitness of each parrot in the source and target domains by measuring the transfer accuracy, task execution stability, and generalization ability of the memory network model under multiple power scenarios includes:

[0112] S51. Calculate the transfer accuracy of the memory network model in the source and target domains. The transfer accuracy is the ratio of the number of correctly executed samples in the target domain task to the total number of samples in the target domain task. The value of the transfer accuracy ranges from 0 to 1, where 1 indicates that all samples are correctly classified.

[0113] S52. Calculate the task execution stability. Task execution stability is the average absolute deviation of the performance fluctuation of the task in multiple executions. Task execution performance is the performance value at each execution. The number of executions is the total number of times the task is executed. The lower the stability value, the smaller the fluctuation of the task execution and the better the stability.

[0114] S53. Calculate the generalization ability of the memory network model. The generalization ability is the difference between the accuracy on the training set and the accuracy on the test set. The generalization ability value ranges from 0 to 1. 0 means that the accuracy on the training set and the test set are exactly equal, and 1 means that the difference between the accuracy on the training set and the test set is the largest.

[0115] S54. Combining transfer accuracy, task execution stability, and generalization ability, calculate the fitness of each parrot individual. Fitness is composed of a weighted sum of transfer accuracy, task execution stability, and generalization ability, with weighting coefficients used to balance the contributions of the three.

[0116] S55. A transfer accuracy of 0.8 or higher indicates a high execution accuracy of the task in the target domain; a task execution stability of less than 0.1 indicates small fluctuations during execution; and a generalization ability of less than 0.05 indicates small differences in performance under different power scenarios.

[0117] This invention ensures the efficient performance of power AI algorithm models across various power task scenarios by precisely calculating transfer accuracy, task execution stability, and generalization ability. Transfer accuracy measures the execution accuracy of the task in the target domain, ensuring task execution precision. Task execution stability assesses the volatility during execution, ensuring the model's stability over long-term operation. Generalization ability measures the performance difference between the training and test sets, reducing the risk of overfitting and enhancing the model's adaptability. Combining these three metrics, this invention calculates the comprehensive fitness value for each individual parrot, optimizing the model's performance in different scenarios, ensuring the execution efficiency and stability of the task in the target domain, while simultaneously improving the model's adaptability in new scenarios.

[0118] In this embodiment, the specific method for selecting the structural parameters of the memory network model with the highest fitness includes:

[0119] S61. Evaluate the final fitness of each parrot individual in the source and target domains. The final fitness of each parrot individual represents its final fitness in the source and target domains.

[0120] ;

[0121] in, For the final fitness of an individual parrot, The total number of individual parrots. , , These are the 1st, 2nd, and 3rd weighting coefficients, respectively. Let be the migration accuracy of the i-th parrot individual. For the task execution stability of the i-th parrot individual, Let represent the generalization ability of the i-th parrot individual. , , The first, second, and third hyperparameters are used to adjust the impact of migration accuracy, task execution stability, and generalization ability on fitness, respectively.

[0122] S62. Sort all parrot individuals according to their final fitness, and select the top five parrot individuals with the highest final fitness as the population memory. The population memory stores the memory network structure parameters corresponding to the five selected parrot individuals. Each parrot individual represents a set of trainable parameters of the memory network model, including the weight matrix, bias term and memory unit gating parameters. By weighted fusion of the memory network structure parameters in the population memory, the weights of the final memory network structure are obtained, which serve as the basis for the optimized memory network model.

[0123] S63. Compare the memory network structures of the top five parrot individuals with the highest fitness with the remaining parrot individuals, select the optimal structure combination, and ensure that the structure performs optimally in the source domain and target domain. The optimal memory network structure represents the weight of the optimal memory network structure.

[0124] Furthermore, performance metrics are calculated on both the source and target domain data. The memory network structures of the top five parrot individuals (elite individuals) are then compared and validated with the structural parameters of other non-selected individuals to confirm whether a structure performs better in a local task. This can be accomplished through weighted cross-validation or recalculating the fitness metrics. The weight represents the contribution of each parrot individual to the final memory network structure during fusion, and can be expressed as the ratio of the individual parrot's fitness to the total fitness of the five parrot individuals.

[0125] S64. Output the optimal memory network structure as the optimized memory network model and output it.

[0126] This invention accurately calculates the fitness of each parrot individual by evaluating their final fitness values ​​in the source and target domains, combined with transfer accuracy, task execution stability, and generalization ability. By ranking the parrot individuals and selecting the top five individuals with the highest fitness as the population memory, the chosen memory network structure is ensured to perform optimally in different scenarios. This process not only optimizes the model structure but also improves the model's execution efficiency in the source and target domains by comparing and selecting the optimal structure combination. Finally, by outputting the optimal memory network structure, efficient task adaptation and knowledge transfer are achieved in multi-power task scenarios, ensuring the model's high accuracy and stability.

[0127] In this embodiment, the specific method for achieving knowledge transfer and task adaptation through transfer learning and meta-learning, applied to the target power scenario, includes:

[0128] S71. Adjust the optimal memory network structure of the optimized memory network model to adapt it to the feature distribution of the target power task scenario;

[0129] Specifically, by using transfer learning techniques, the model parameters are fine-tuned using target power task scenario data to align the feature distributions of the source and target domains.

[0130] S72. Use transfer learning to transfer knowledge from the source domain to the target domain. Utilize the similarity between the source and target domains to transfer knowledge from the source domain to the target domain, optimize task performance in the target domain, and overcome the small sample problem in the power industry through transfer learning.

[0131] S73. Based on the requirements of the target power scenario, fine-tune the optimized memory network model. Adjust the structural parameters of the optimized memory network model by minimizing the loss function of the target domain to make up for the shortcomings in the small sample power scenario.

[0132] S74. Further optimize the fine-tuned memory network model using meta-learning methods to ensure it can adapt to different power task scenarios and share knowledge across multiple power tasks:

[0133] ;

[0134] in, The adjusted fitness; The number of samples for the target task; For the first The true labels of each target task sample; For the first The predicted output for each target task sample; For migration fitness weighting coefficients; To maintain the weighting coefficients of source domain knowledge; This refers to the number of structural parameters of the memory network model; For the first in the target domain Each memory network model structural parameters, For the first in the source domain Each memory network model structural parameters, For each sample in the target power task dataset, This serves as an index for the various structural parameters of the memory network model;

[0135] S75. In the target power scenario, the task is executed through the optimized memory network model, the effect of the task execution is evaluated, and the evaluation results are output.

[0136] Specifically, the effectiveness can be evaluated using evaluation metrics, which include prediction error (such as mean absolute error MAE, root mean square error RMSE), classification accuracy, and F1 score.

[0137] S76. Based on the evaluation results, further optimize the memory network model, adjust the structural parameters of the memory network model, and complete the knowledge transfer and task adaptation for multiple power scenarios.

[0138] Specifically, if the evaluation results do not meet the requirements, the current structural parameters are used as input to readjust and optimize the structural parameters of the memory network model until the convergence condition is met, thus completing the knowledge transfer and task adaptation for multiple power scenarios.

[0139] This invention overcomes the problem of limited sample size in power tasks by adjusting and optimizing the memory network model to adapt to the characteristics of the target power task scenario and transferring knowledge from the source domain to the target domain through transfer learning. Further improvements in task adaptability are achieved by fine-tuning the network model and minimizing the loss function of the target domain. After applying meta-learning methods, the model can share knowledge across multiple power task scenarios, ensuring efficient task execution and knowledge transfer. This method significantly improves the accuracy, stability, and adaptability of the memory network model in power tasks through continuous optimization and adjustment of its parameters, ensuring efficient task execution and result evaluation in complex, multi-scenario environments.

[0140] This invention also provides a multi-scenario transfer learning system for power artificial intelligence algorithm models, including a data acquisition and preprocessing module, a power artificial intelligence algorithm model generation module, an initialization module, an iterative optimization module, and a transfer learning and meta-learning module;

[0141] The data acquisition and preprocessing module is used to collect operational data from multiple power scenarios and preprocess the operational data to form samples corresponding to different power scenarios.

[0142] The power artificial intelligence algorithm model generation module is used to transform samples of each power scenario into nodes in the graph structure based on the graph structure, and establish connection edges through the feature similarity between nodes to generate power artificial intelligence algorithm models that represent the relationship between power scenarios.

[0143] The initialization module is used to initialize the structural parameters of a pre-built memory network model using a parrot optimization algorithm that incorporates a memory feedback mechanism. The memory network model is used to process temporal information and historical data, and each parrot represents a set of structural parameters of the memory network model.

[0144] The iterative optimization module is used to simulate the imitation behavior of an individual parrot using an electric power artificial intelligence algorithm model. The structural parameters of the memory network model are iteratively optimized by using the transfer accuracy, task execution stability and generalization ability of the memory network model in the source and target domains under multiple power scenarios as fitness.

[0145] The transfer learning and meta-learning module is used to select the structural parameters of the memory network model with the highest fitness and apply them to the target power scenario. Knowledge transfer and task adaptation are completed through transfer learning and meta-learning.

[0146] This invention optimizes the adaptability and knowledge transfer capability of power AI algorithm models through multi-module collaborative work. The data acquisition and preprocessing module collects operational data from multiple power application scenarios, providing a rich information source for subsequent model training. The power AI algorithm model generation module constructs a graph model of power scenarios based on the similarity between nodes, thereby improving the connection between power task scenarios and the structured representation of data. The initialization module initializes the memory network structure parameters using the parrot optimization algorithm, ensuring a good performance foundation for the initial network structure. The iterative optimization module further adjusts the structural parameters of individual parrots using power scenario operational data. The optimized network structure better adapts to different task scenarios, and by calculating the fitness values ​​in the source and target domains, it ensures that each parrot maintains good performance in different task scenarios. The transfer learning and meta-learning module applies the optimized memory network model to the target power task scenario, achieving efficient task adaptation and cross-scenario knowledge transfer. Finally, by evaluating the task execution effect, the superiority of the optimized model in the target power task scenario is further verified, ensuring the model's practical application value.

[0147] Example 2:

[0148] To verify the feasibility of this invention in practice, this embodiment applies the multi-scenario transfer learning method and system of this invention to the task of power load forecasting and equipment fault detection in a certain region. The power system in this region has complex power load fluctuation characteristics, and equipment faults are frequently affected by factors such as weather changes and equipment aging. Therefore, the problems of insufficient data and poor model training performance faced by load forecasting models and equipment fault detection models need to be addressed through more efficient technical means. This invention adopts a method combining transfer learning and meta-learning. For different power task scenarios, through knowledge transfer and sharing between models, it improves the system's adaptability and performance in different tasks, solving the deficiency of traditional methods in effectively dealing with insufficient data.

[0149] In practical applications, the region's electricity load data, equipment status data, and meteorological data for the past year were first collected. Historical operation records, including equipment operation history and maintenance records, were used as features for model training and optimization.

[0150] Because power load data and equipment fault data often suffer from small sample sizes, traditional deep learning models cannot achieve good training results on limited data. Therefore, we adopted a multi-scenario transfer learning system. This system constructs a multi-scenario memory network model through preprocessing and feature extraction, and optimizes the network structure using the parrot optimization algorithm to adapt to different power task scenarios. By simulating the imitation behavior of individual parrots, low-fitness individuals imitate the memory network structure of high-fitness individuals, thereby achieving model optimization.

[0151] Table 1 demonstrates the significant improvement in performance achieved by employing a multi-scenario transfer learning system in power load forecasting and equipment fault detection tasks. Taking power load forecasting as an example, in traditional deep learning methods using LSTM networks, the prediction error is 12.6%. However, after applying the multi-scenario transfer learning system of this invention, the prediction error is reduced to 9.2%. Similarly, in equipment fault detection, the detection accuracy of traditional methods is 81.5%, while the accuracy is increased to 88.3% using the system of this invention.

[0152] Table 1

[0153] index Traditional LSTM networks Multi-scenario transfer learning system Power load forecasting error 12.6% 9.2% Equipment fault detection accuracy 81.5% 88.3% Model training time 18 15 Model generalization ability Poor excellent Data demand high Low

[0154] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-scene transfer learning of an electric power artificial intelligence algorithm model, characterized in that, The application relates to a method for constructing a power artificial intelligence algorithm model. The method comprises the following steps: collecting operation data of multiple power scenes and preprocessing the operation data to form samples corresponding to different power scenes; taking a graph structure as a basis, converting the sample of each power scene into a node in the graph structure, establishing a connection edge between nodes through feature similarity, and generating a power artificial intelligence algorithm model for representing the correlation between power scenes; initializing the structure parameters of a pre-constructed memory network model by using a parrot optimization algorithm with a memory feedback mechanism, the memory network model being used for processing time series information and historical data, and each parrot individual representing a group of structure parameters of the memory network model; simulating the imitating behavior of the parrot individual by using the power artificial intelligence algorithm model, taking the migration accuracy, task execution stability and generalization ability of the memory network model in source and target domains under multiple power scenes as fitness, and iteratively optimizing the structure parameters of the memory network model; selecting the structure parameters of the memory network model with the highest fitness, and applying the structure parameters to a target power scene to complete knowledge migration and task adaptation through transfer learning and meta-learning. The operation data comprises load data, equipment state, meteorological data and historical operation records. The specific method for taking the graph structure as a basis, converting the sample of each power scene into a node in the graph structure, establishing a connection edge between nodes through feature similarity, and generating a power artificial intelligence algorithm model for representing the correlation between power scenes comprises the following steps: The sample of each power scene is converted into a node in the graph structure, and the node is represented by a feature vector; the feature vector is composed of load data, equipment state and meteorological information in the operation data; The similarity between nodes is calculated, the cosine similarity is used to measure the similarity between nodes, the relationship between nodes is established, and a connection edge is established between each pair of nodes; 2.The method of claim 1, wherein, Based on the similarity between nodes, a power artificial intelligence algorithm model is generated, the sample of each power scene is represented by a node in the graph structure, and the relationship between power scenes is established through a connection edge between nodes; 3.The method of claim 2, wherein, Based on the power artificial intelligence algorithm model, the neighbor node set of each node is calculated, and the weighted average value of the node feature representation of the neighbor nodes is calculated; The node feature representation of the current node is fused with the weighted average value of the neighbor nodes to generate an updated node feature representation, thereby optimizing the power artificial intelligence algorithm model. The specific method for initializing the structure parameters of the pre-constructed memory network model by using the parrot optimization algorithm with a memory feedback mechanism comprises the following steps: The structure parameters of the pre-constructed memory network model are initialized by using the parrot optimization algorithm, a group of structure parameters is allocated to each parrot individual, and each parrot individual represents a group of structure parameters of the memory network model; In the parrot optimization algorithm, the structure parameters of the memory network model are optimized by introducing a memory feedback mechanism, the structure parameters of each parrot individual are adjusted according to the samples corresponding to different power scenes, and the adjustment is represented as follows: The storage information amount of the memory network model is adaptively adjusted by using the adjusted structure parameters of each parrot individual.

4. The multi-scene transfer learning method of the power artificial intelligence algorithm model according to claim 3, characterized in that, ​ ​ ​ ； wherein, is the structure parameter of parrot individual i after the t+1th iteration; is the structure parameter of parrot individual i at the tth iteration; i is the index of parrot individual in the parrot optimization algorithm; is the learning rate adjustment coefficient; is the set of neighbor individuals of parrot individual i; is the fitness of parrot individual i; is the fitness of parrot individual k; is the perturbation intensity coefficient; is the norm of the fitness vector of parrot individual i; is the norm of the fitness vector of parrot individual k; is the set of non-neighbor individuals of parrot individual i; is the structure parameter of parrot individual j at the tth iteration; k is the neighbor individual of parrot individual i; j is the non-neighbor individual of parrot individual i; t is the iteration number in adjusting the structure parameter of each parrot individual; is the difference between the structure parameters of parrot individual i and parrot individual j at the tth iteration, is the task adaptability coefficient, is the contribution coefficient; ​ The structural parameters of each parrot individual are adjusted, and important historical operation records are dynamically selected for access according to changes in the device state to update the memory network model. 5.The method of claim 4, wherein, The specific method for simulating the mimic behavior of the parrot individual by using the power artificial intelligence algorithm model, optimizing the structural parameters of the memory network model, and the low-adaptability parrot individual imitating the structural parameters of the high-adaptability parrot individual comprises: The specific method for simulating the mimic behavior of the parrot individual by using the power artificial intelligence algorithm model, optimizing the structural parameters of the memory network model, and the low-adaptability parrot individual imitating the structural parameters of the high-adaptability parrot individual comprises: The final adaptability of each parrot individual in the source domain and the target domain is calculated by measuring the migration accuracy, task execution stability, and generalization ability of the memory network model in multiple power scenarios. 6.The method of claim 5, wherein, The specific method for simulating the mimic behavior of the parrot individual by using the power artificial intelligence algorithm model, optimizing the structural parameters of the memory network model, and the low-adaptability parrot individual imitating the structural parameters of the high-adaptability parrot individual comprises: In each round of optimization, the mimic behavior of the parrot individual is simulated by using the power artificial intelligence algorithm model, and information is transmitted based on the adaptability and structural parameters of the neighbor individuals. A mimic factor is defined for the mimic operation, which determines the degree of imitation of the low-adaptability parrot individual by the high-adaptability parrot individual. The adaptability of the low-adaptability parrot individual is between 0 and 0.3, and the adaptability of the high-adaptability parrot individual is between 0.7 and 1. The low-adaptability parrot individual will imitate the structural parameters of the neighbor individuals according to the difference in adaptability. The low-adaptability parrot individual optimizes the structural parameters of the memory network model by continuously imitating the high-adaptability parrot individual: ； wherein is a low fitness parrot individual; l is an index of a parrot individual in the population; is a weighting factor; is the fitness of a neighbor individual is the fitness of a neighbor individual is the fitness of a neighbor individual is the sum of the fitness of all neighbor individuals in the population; is the fitness of a neighbor individual is the structure parameter at iteration n; is the structure parameter at iteration n; is a low fitness parrot individual is the structure parameter at iteration n; is the structure parameter at iteration n; When imitating, the low-adaptability parrot individual selects individuals with higher adaptability from the neighbor individuals as the imitation objects. After multiple iterations, the low-adaptability parrot individual gradually approaches the structural parameters of the high-adaptability parrot individual.

7. The multi-scene transfer learning method of the power artificial intelligence algorithm model according to claim 6, characterized in that, The specific method for calculating the final adaptability of each parrot individual in the source domain and the target domain by measuring the migration accuracy, task execution stability, and generalization ability of the memory network model in multiple power scenarios comprises: The migration accuracy of the memory network model in the source domain and the target domain is calculated. The migration accuracy is the ratio of the number of correctly executed samples in the target domain task to the total number of samples in the target domain task. The value range of the migration accuracy is 0 to 1, and 1 indicates that all samples are correctly classified. The task execution stability is calculated. The task execution stability is the average absolute deviation of the performance fluctuation in multiple executions. The task execution performance is the performance value in each execution, and the execution times are the total number of task executions. The generalization ability of the memory network model is calculated. The generalization ability is the difference between the accuracy on the training set and the accuracy on the test set. The value range of the generalization ability is 0 to 1, 0 indicates that the accuracy on the training set and the test set is completely equal, and 1 indicates that the accuracy difference between the training set and the test set is the largest. The fitness of each parrot individual is calculated in combination of the migration accuracy, the task execution stability and the generalization ability, and the fitness is composed of the weighted combination of the migration accuracy, the task execution stability and the generalization ability, and the weight coefficient is used to balance the contributions of the three; The migration accuracy is above 0.8, indicating that the execution accuracy of the task in the target domain is high; the task execution stability is below 0.1, indicating that the fluctuation in the execution process is small; and the generalization ability is below 0.05, indicating that the performance difference in different power scenarios is small. 8.The method of claim 7, wherein, The specific method for selecting the structural parameters of the memory network model with the highest fitness includes: The final fitness of each parrot individual in the source domain and the target domain is evaluated, and the final fitness of each parrot individual represents the final fitness of the parrot individual in the source domain and the target domain: ； wherein, is the final fitness of the parrot individual, is the total number of parrot individuals, , , are the first, second and third weighting coefficients, respectively, is the migration accuracy of the i-th parrot individual, is the task execution stability of the i-th parrot individual, is the generalization ability of the i-th parrot individual, , , are the first, second and third adjustment hyperparameters of the migration accuracy, the task execution stability and the generalization ability, respectively, influencing the fitness. According to the final fitness, all parrot individuals are sorted, and the top five parrot individuals with the highest final fitness are selected as the population memory, the population memory stores the memory network structure parameters corresponding to the selected five parrot individuals, and the weight of the final memory network structure is obtained; The memory network structure parameters of the top five parrot individuals with the highest final fitness are compared with those of the remaining parrot individuals, and the optimal structure combination is selected to ensure that the structure performs optimally in the source domain and the target domain, and the optimal memory network structure represents the weight of the optimal memory network structure; The optimal memory network structure is output as the optimized memory network model and is output. 9.The method of claim 8, wherein, The specific method for applying to the target power scenario to complete knowledge transfer and task adaptation through transfer learning and meta-learning includes: Adjusting the optimal memory network structure of the optimized memory network model to adapt to the feature distribution of the target power task scenario; Using the transfer learning method to transfer the knowledge in the source domain to the target domain, utilizing the similarity between the source domain and the target domain to transfer the knowledge in the source domain to the target domain, optimizing the task execution effect in the target domain, and overcoming the power small sample problem through transfer learning; According to the requirements of the target power scenario, the optimized memory network model is fine-tuned, the structure parameters of the optimized memory network model are adjusted by minimizing the loss function of the target domain, and the deficiencies in the power small sample scenario are made up; The fine-tuned memory network model is further optimized using the meta-learning method to ensure that it can adapt to different power task scenarios and share knowledge among multiple power tasks: ； wherein, is an adjusted fitness; is a number of target task samples; is a true label of an i-th target task sample; is a true label of an i-th target task sample; is a predicted output of an i-th target task sample; is a predicted output of an i-th target task sample; is a migration fitness weight coefficient; is a weight coefficient for preserving source domain knowledge; is a number of memory network model structure parameters; is an i-th memory network model structure parameter in a target domain, is an i-th memory network model structure parameter in a source domain, is an i-th memory network model structure parameter in a target domain, is an i-th memory network model structure parameter in a source domain, is an index of each sample in a target power task dataset, is an index of each structure parameter of a memory network model; In the target power scenario, the task execution is performed through the optimized memory network model, and the execution effect is evaluated, and the evaluation result is output; According to the evaluation result, the memory network model is further optimized, the structure parameters of the memory network model are adjusted, and the knowledge transfer and task adaptation in multiple power scenarios are completed. 10.A multi-scene transfer learning system of a power artificial intelligence algorithm model, characterized in that, The method includes a data acquisition and preprocessing module, a power artificial intelligence algorithm model generation module, an initialization module, an iterative optimization module, and a transfer learning and meta-learning module; The data acquisition and preprocessing module is used to collect the operation data of multiple power scenarios, and to preprocess the operation data to form samples corresponding to different power scenarios; The power artificial intelligence algorithm model generation module is configured to convert each power scene sample into a node in a graph structure, establish a connection edge through feature similarity between the nodes, and generate a power artificial intelligence algorithm model for representing an association relationship between power scenes, with the graph structure as a basis; The initialization module is configured to initialize structure parameters of a pre-constructed memory network model by using a parrot optimization algorithm with a memory feedback mechanism introduced, the memory network model being configured to process time series information and historical data, and each parrot individual representing a set of structure parameters of the memory network model; The iterative optimization module is configured to simulate an imitation behavior of the parrot individual by using the power artificial intelligence algorithm model, take a migration accuracy, a task execution stability, and a generalization ability of the memory network model in source domains and target domains in multiple power scenes as fitness, and iteratively optimize the structure parameters of the memory network model. The transfer learning and meta-learning module is configured to select structure parameters of the memory network model with the highest fitness, apply the structure parameters to a target power scene, and complete knowledge migration and task adaptation through transfer learning and meta-learning.

Images