A migration method for source domain learning strategies based on topology change scenarios

Abstract

The present invention discloses a migration method of a source domain learning strategy based on a topology change scenario, which relates to the field of artificial intelligence, including: establishing a corresponding relationship between a source topology scenario and a target topology scenario; analyzing and obtaining the target data type used for training a model to be trained, and using the target training data used for training the model to be trained in the power grid target field as the data to be trained; obtaining source training data similar to the data to be trained in the power grid source field as similar training data; obtaining a source training model corresponding to the model to be trained in the power grid source field as a similar training model; obtaining training supplementary data; and using the training supplementary data and the data to be trained to train the model to be trained. By obtaining similar training data and similar training models, the similar training data is migrated and transformed to ensure that the data meets the characteristics required by the training data of the model to be trained, and the normal training of the model in the power grid target field is ensured.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a migration method of a source domain learning strategy based on a topology change scenario. Background Art

[0002] In the field of machine learning and artificial intelligence, there are a large number of models that need to be trained. The power grid source field is a field with a large amount of data available for training. In this field, the data has been labeled and has corresponding labels, and can be directly called for training. The power grid target field is a field where data has not been labeled. Its data has no labels, so it cannot be directly called for training. Because when the data type does not match the model to be trained, the required training requirements cannot be met. However, due to the huge amount of data for model training, re-labeling the data in the power grid target field will consume a lot of manpower, thus requiring a large amount of capital investment. In order to avoid such a situation, data mapping and conversion is performed between the power grid source field and the power grid target field, so that the data in the power grid source field is used to train the model in the power grid target field, which can save costs.

[0003] In the power grid, due to the huge demand for data processing, corresponding models are trained in the fields corresponding to different categories of data in the power grid, and these models are used to automatically judge the data in the power grid, thereby identifying the faults or abnormalities of the power grid. However, in the actual operation of the power grid, due to the emergence of new problems, which cannot be handled by the models trained in the original power grid source field, it is necessary to train new models in the power grid target field where these problems appear to identify and handle the problems. In order to reduce the training cost, when training a new model in the power grid target field, the data in the power grid source field is called to train the model in the power grid target field. However, due to the large amount of data in the power grid source field and the unclear correspondence between the power grid source field and the power grid target field, it is impossible to determine which data in the power grid source field is called when calling data. At the same time, it is also unclear how to better map the data in the power grid source field to the power grid target field to ensure the normal training of the model in the power grid target field. Summary of the invention

[0004] In order to solve the above technical problems, a migration method of source domain learning strategy based on topology change scenario is provided. This technical solution solves the problems raised in the above background technology.

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0006] A method for migrating source domain learning strategies based on topology change scenarios, comprising:

[0007] Acquire a source topology scenario in a power grid source domain, where the source topology scenario is composed of at least one source training data; acquire a target topology scenario in a power grid target domain, where the target topology scenario is composed of at least one target training data;

[0008] Acquire at least one source training model in the power grid source field, and acquire source training data used in training the source training model, where the source training model is a trained model;

[0009] Acquire at least one target training model in the target field of the power grid, and acquire target training data used when training the target training model, where the target training model is a trained model;

[0010] Based on the recognition result, a corresponding relationship is established between the source topology scene and the target topology scene;

[0011] Acquire a model to be trained in a target domain of a power grid, analyze and obtain a target data type used for training the model to be trained, and based on the target data type, use the target training data used for training the model to be trained in the target domain of the power grid as the data to be trained;

[0012] Determine whether the data to be trained is sufficient. If so, use the data to be trained to train the model to be trained. If not, based on the correspondence between the source topology scenario and the target topology scenario, obtain at least one source training data similar to the data to be trained in the power grid source field as similar training data;

[0013] Based on the similar training data, a source training model corresponding to the model to be trained is obtained in the power grid source field as a similar training model;

[0014] Based on the analysis of the training modes of similar training models and the model to be trained, similar training data is transferred and transformed to obtain supplementary training data;

[0015] The model to be trained is trained using the training supplementary data and the data to be trained.

[0016] Preferably, the establishing a corresponding relationship between the source topology scene and the target topology scene based on the recognition result comprises the following steps:

[0017] Obtaining the source data classification standard of the source training model and obtaining the source data training purpose of the source training model;

[0018] Obtain target data classification standards for the target training model, and obtain target data training purposes for the target training model;

[0019] Obtain a first recognition accuracy rate of the target training data of the target training model input into the source training model, and obtain a second recognition accuracy rate of the source training data of the source training model input into the target training model;

[0020] The first recognition accuracy rate is superimposed on the second recognition accuracy rate to obtain a first approximation between the source training model and the target training model;

[0021] Extract features from the source data classification standard and the target data classification standard respectively, and use the overlap ratio of the features of the source data classification standard and the target data classification standard as the standard overlap degree;

[0022] Extract features of the source data training purpose and the target data training purpose respectively, and use the overlap ratio of the features of the source data training purpose and the target data training purpose as the purpose overlap degree;

[0023] Using the analytic hierarchy process, the weights of the first approximation, standard coincidence, and purpose coincidence are obtained;

[0024] Using the approximation formula, calculate the second approximation between the source training model and the target training model;

[0025] Matching the source training model with the second largest similarity to the target training model to the target training model;

[0026] Establish a correspondence between the source training data of the source training model and the target training data of the corresponding target training model, and summarize the correspondence as the correspondence between the source topology scene and the target topology scene;

[0027] The approximation formula is as follows: ;

[0028] Where A is the second similarity between the source training model and the target training model, is the weight of the first approximation, is the weight of the standard coincidence degree, is the weight of the target overlap, a is the first approximation between the source training model and the target training model, b is the standard overlap between the source training model and the target training model, and c is the target overlap between the source training model and the target training model.

[0029] Preferably, the step of acquiring the model to be trained in the target field of the power grid and analyzing to obtain the target data type used for training the model to be trained comprises the following steps:

[0030] Obtain the data classification standards and data training purposes of the model to be trained;

[0031] Feature extraction is performed on the source data classification standard and the data to be classified standard respectively, and the overlap ratio of the features of the source data classification standard and the data to be classified standard is used as the standard secondary overlap degree;

[0032] Extract features from the source data training purpose and the data to be trained purpose respectively, and use the overlap ratio of the features of the source data training purpose and the data to be trained purpose as the secondary overlap degree of the purpose;

[0033] Use the coincidence synthesis formula to calculate the similarity between the model to be trained and the source training model;

[0034] The source training model with the greatest similarity to the model to be trained is used as the feature source training model;

[0035] The data type of the target training data corresponding to the source training data of the feature source training model is used as the target data type;

[0036] The combined formula is as follows: ;

[0037] Among them, B is the similarity between the model to be trained and the source training model, d is the standard secondary overlap, and e is the target secondary overlap.

[0038] Preferably, the step of using the target training data in the target field of the power grid for training the model to be trained as the data to be trained comprises the following steps:

[0039] In the power grid target domain, target training data of the same type as the target data type is used as the data to be trained.

[0040] Preferably, the determining whether the training data is sufficient comprises the following steps:

[0041] Use the data to be trained to train the model to be trained;

[0042] Feature recognition obtains at least one data feature of the data to be trained, and based on the big data, obtains at least one replaceable data, the data feature of the replaceable data being consistent with the data feature of the data to be trained;

[0043] Input the replaceable data into the trained model to be trained, and count the error processing ratio of the model to be trained;

[0044] Based on historical data, obtain the allowable error tolerance of the model to be trained;

[0045] When the error processing ratio is greater than the allowable fault tolerance rate, the training data is insufficient, otherwise the training data is sufficient.

[0046] Preferably, the step of obtaining at least one source training data similar to the data to be trained in the power grid source field as similar training data comprises the following steps:

[0047] Acquire at least one source training data corresponding to the data to be trained in the power grid source domain as initial source training data;

[0048] Feature recognition obtains at least one data feature of the initial source training data, and in the power grid source field, obtains at least one similar training data, the data features of the similar training data are consistent with the data features of the initial source training data.

[0049] Preferably, the step of obtaining a source training model corresponding to the model to be trained in the power grid source field as a similarity training model comprises the following steps:

[0050] The source training data used in training the source training model is used to count the parts that overlap with the similar training data, and use them as the similar data of the source training model;

[0051] The amount of similar data of the source training model is divided by the amount of source training data used when training the source training model to obtain a similarity ratio;

[0052] The source training model with the largest similarity ratio is used as the similarity training model.

[0053] Preferably, the step of performing migration transformation on similar training data to obtain supplementary training data comprises the following steps:

[0054] The overlapping parts of the source training data of the similar training model and the similar training data are used as the migration benchmark data;

[0055] In the target domain of the power grid, target training data corresponding to the migration benchmark data is obtained as the migration conversion benchmark data;

[0056] The part of the source training data of the similar training model except the migration benchmark data is used as the data to be migrated;

[0057] Use the minimum and maximum values ​​of the migration conversion benchmark data as interval endpoints to form a conversion interval;

[0058] Obtain an initial mapping function, satisfying that the minimum and maximum values ​​of the data obtained by substituting the migration benchmark data into the initial mapping function are used as the interval endpoints to form a verification interval covering the conversion interval;

[0059] Uniformly select at least one identification point in the interval (0, 1), multiply the value at the identification point by the initial mapping function, and obtain a secondary mapping function;

[0060] Substituting the migration benchmark data into the secondary mapping function to obtain secondary migration benchmark data, and matching the migration conversion benchmark data corresponding to the migration benchmark data to the secondary migration benchmark data;

[0061] superimposing the overlapped portion of at least one secondary migration reference data obtained using the identification point with the corresponding migration conversion reference data to obtain a determination value of the identification point;

[0062] The secondary mapping function obtained from the identification point with the largest judgment value is used as the migration mapping function;

[0063] Use the migration mapping function to perform migration transformation on the data to be migrated to obtain the migration target data;

[0064] The part of the migration target data and the migration conversion benchmark data other than the data to be trained is used as supplementary training data.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] By establishing a corresponding relationship between the source topology scenario and the target topology scenario, similar training data and similar training models are obtained, and the similar training data is migrated and transformed. Therefore, when a large number of new problems appear in the power grid and the model in the power grid source field cannot meet the requirements, the new model can be trained in the power grid target field where the large number of new problems are located. At the same time, in order to reduce the cost of obtaining data for training, the power grid source field is converted into data that can be used for training the power grid target field model by migration. During the migration, the data that needs to be called in the power grid source field is determined, that is, similar training data. There is a high correlation between the similar training data and the model to be trained. Therefore, the training effect of the model to be trained can be guaranteed by using the data migrated and transformed from the similar training data for training. At the same time, when performing the migration and transformation, it is ensured that the converted data can meet the characteristics required by the training data of the model to be trained, thereby ensuring the normal training of the model in the power grid target field. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic diagram of a flow chart of a migration method of a source domain learning strategy based on a topology change scenario of the present invention;

[0068] Figure 2 A schematic diagram of a process of establishing a corresponding relationship between a source topology scene and a target topology scene based on an identification result of the present invention;

[0069] Figure 3 A schematic diagram of a process of obtaining a model to be trained in a target field of a power grid and analyzing and obtaining a target data type used for training the model to be trained;

[0070] Figure 4 A schematic diagram of a flow chart of determining whether the training data is sufficient in the present invention;

[0071] Figure 5 A schematic diagram of a process of obtaining at least one source training data similar to the data to be trained in the power grid source field as similar training data according to the present invention;

[0072] Figure 6A schematic diagram of a process of obtaining a source training model corresponding to a model to be trained in a power grid source field as a similar training model according to the present invention;

[0073] Figure 7 The present invention is a flow chart of the process of transferring and transforming similar training data to obtain supplementary training data. DETAILED DESCRIPTION

[0074] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art may think of other obvious variations.

[0075] Reference Figure 1 As shown, a migration method of a source domain learning strategy based on a topology change scenario includes:

[0076] Acquire a source topology scenario in a power grid source domain, where the source topology scenario is composed of at least one source training data; acquire a target topology scenario in a power grid target domain, where the target topology scenario is composed of at least one target training data;

[0077] Acquire at least one source training model in the power grid source field, and acquire source training data used in training the source training model, where the source training model is a trained model;

[0078] Acquire at least one target training model in the target field of the power grid, and acquire target training data used when training the target training model, where the target training model is a trained model;

[0079] Based on the recognition result, a corresponding relationship is established between the source topology scene and the target topology scene;

[0080] Acquire a model to be trained in a target domain of a power grid, analyze and obtain a target data type used for training the model to be trained, and based on the target data type, use the target training data used for training the model to be trained in the target domain of the power grid as the data to be trained;

[0081] Determine whether the data to be trained is sufficient. If so, use the data to be trained to train the model to be trained. If not, based on the correspondence between the source topology scenario and the target topology scenario, obtain at least one source training data similar to the data to be trained in the power grid source field as similar training data;

[0082] Based on the similar training data, a source training model corresponding to the model to be trained is obtained in the power grid source field as a similar training model;

[0083] Based on the analysis of the training modes of similar training models and the model to be trained, similar training data is transferred and transformed to obtain supplementary training data;

[0084] The model to be trained is trained using the training supplementary data and the data to be trained.

[0085] In this scheme, by analyzing at least one source training model completed in the power grid source field and at least one target training model completed in the power grid target field, a corresponding relationship is established between the source topology scenario and the target topology scenario, that is, a corresponding relationship is established between the source training data and the target training data. Therefore, when training the model to be trained, a judgment is made in the power grid target field whether the data to be trained is sufficient. When the data is insufficient, the data used for migration and supplement in the power grid source field can be determined according to the corresponding relationship between the source training data and the target training data, and then the data to be trained can be expanded, thereby ensuring the training effect of the model to be trained without the need for additional manual data labeling, and controlling the cost of training.

[0086] Reference Figure 2 As shown, based on the recognition result, establishing a corresponding relationship between the source topology scene and the target topology scene includes the following steps:

[0087] Obtaining the source data classification standard of the source training model and obtaining the source data training purpose of the source training model;

[0088] Obtain target data classification standards for the target training model, and obtain target data training purposes for the target training model;

[0089] Obtain a first recognition accuracy rate of the target training data of the target training model input into the source training model, and obtain a second recognition accuracy rate of the source training data of the source training model input into the target training model;

[0090] The first recognition accuracy rate is superimposed on the second recognition accuracy rate to obtain a first approximation between the source training model and the target training model;

[0091] Extract features from the source data classification standard and the target data classification standard respectively, and use the overlap ratio of the features of the source data classification standard and the target data classification standard as the standard overlap degree;

[0092] Extract features of the source data training purpose and the target data training purpose respectively, and use the overlap ratio of the features of the source data training purpose and the target data training purpose as the purpose overlap degree;

[0093] Using the analytic hierarchy process, the weights of the first approximation, standard coincidence, and purpose coincidence are obtained;

[0094] Using the approximation formula, calculate the second approximation between the source training model and the target training model;

[0095] Matching the source training model with the second largest similarity to the target training model to the target training model;

[0096] Establish a correspondence between the source training data of the source training model and the target training data of the corresponding target training model, and summarize the correspondence as the correspondence between the source topology scene and the target topology scene;

[0097] The approximation formula is as follows: ;

[0098] Where A is the second similarity between the source training model and the target training model, is the weight of the first approximation, is the weight of the standard coincidence degree, is the weight of the target overlap, a is the first approximation between the source training model and the target training model, b is the standard overlap between the source training model and the target training model, and c is the target overlap between the source training model and the target training model.

[0099] When performing data correspondence, it is mainly determined by the correspondence between the source training model and the target training model. Since both the source training model and the target training model are models that have completed training, all types of data can be obtained. Therefore, the approximation between the source training model and the target training model can be comprehensively judged from multiple angles, and then the source training model with the highest matching degree is paired with the target training model. Therefore, the training data used by the paired source training model and the target training model can also be matched, because the models they trained have a high degree of approximation;

[0100] Here, matching the source training model with the largest second similarity to the target training model to the target training model means that if there is a second similarity between the target training model and each source training model, then the target training model must have the highest matching degree with the source training model corresponding to the largest second similarity, and thus matching is performed as above.

[0101] Reference Figure 3 As shown, obtaining a model to be trained in the target field of the power grid and analyzing and obtaining the target data type used for training the model to be trained includes the following steps:

[0102] Obtain the data classification standards and data training purposes of the model to be trained;

[0103] Feature extraction is performed on the source data classification standard and the data to be classified standard respectively, and the overlap ratio of the features of the source data classification standard and the data to be classified standard is used as the standard secondary overlap degree;

[0104] Extract features from the source data training purpose and the data to be trained purpose respectively, and use the overlap ratio of the features of the source data training purpose and the data to be trained purpose as the secondary overlap degree of the purpose;

[0105] Use the coincidence synthesis formula to calculate the similarity between the model to be trained and the source training model;

[0106] The source training model with the greatest similarity to the model to be trained is used as the feature source training model;

[0107] The data type of the target training data corresponding to the source training data of the feature source training model is used as the target data type;

[0108] The combined formula is as follows: ;

[0109] Among them, B is the similarity between the model to be trained and the source training model, d is the standard secondary overlap, and e is the target secondary overlap.

[0110] Since the model to be trained is a new model, the training data used is unknown. Therefore, it is necessary to determine the training data of the model to be trained in the target domain of the power grid. However, the number of models in the target domain of the power grid is not as large as that in the source domain of the power grid. Therefore, when performing similarity identification, the source training model is used to compare with the model to be trained. The matching degree between the obtained feature source training model and the model to be trained will be higher than the matching degree between the target training model and the model to be trained. Then, according to the data type of the target training data corresponding to the source training data used for training the feature source training model, the target data type is obtained. Here, due to the corresponding relationship between the source topology scenario and the target topology scenario, there is a corresponding relationship between the source training data and the target training data.

[0111] Using target training data in the target field of the power grid for training the model to be trained as the data to be trained includes the following steps:

[0112] In the power grid target domain, target training data of the same type as the target data type is used as the data to be trained.

[0113] Reference Figure 4 As shown, judging whether the training data is sufficient includes the following steps:

[0114] Use the data to be trained to train the model to be trained;

[0115] Feature recognition obtains at least one data feature of the data to be trained, and based on the big data, obtains at least one replaceable data, the data feature of the replaceable data being consistent with the data feature of the data to be trained;

[0116] Input the replaceable data into the trained model to be trained, and count the error processing ratio of the model to be trained;

[0117] Based on historical data, obtain the allowable error tolerance of the model to be trained;

[0118] When the error processing ratio is greater than the allowable fault tolerance rate, the training data is insufficient, otherwise the training data is sufficient.

[0119] The basis for judgment must be based on the data processing of the model to be trained after training, but the data used for processing cannot be the data to be trained. Therefore, it is necessary to form similar replaceable data based on the data to be trained. Since its characteristics are consistent with those of the data to be trained and only differ in the data, its input result has reference value.

[0120] Reference Figure 5 As shown, obtaining at least one source training data similar to the data to be trained in the power grid source field as similar training data includes the following steps:

[0121] Acquire at least one source training data corresponding to the data to be trained in the power grid source domain as initial source training data;

[0122] Feature recognition obtains at least one data feature of the initial source training data, and in the power grid source field, obtains at least one similar training data, the data features of the similar training data are consistent with the data features of the initial source training data.

[0123] Here, according to the steps of obtaining similar training data, the source training data of the feature source training model is further expanded.

[0124] Reference Figure 6 As shown, obtaining a source training model corresponding to the model to be trained in the power grid source domain as a similar training model includes the following steps:

[0125] The source training data used in training the source training model is used to count the parts that overlap with the similar training data, and use them as the similar data of the source training model;

[0126] The amount of similar data of the source training model is divided by the amount of source training data used when training the source training model to obtain a similarity ratio;

[0127] The source training model with the largest similarity ratio is used as the similarity training model.

[0128] Since the source training data of the similar training data and the characteristic source training model are different, the characteristic source training model was originally used to correspond to the model to be trained, but its similarity was judged based on the training standards and purposes, and the accuracy was insufficient. However, based on the similar training data, the similar training model obtained may be different from the characteristic source training model, and its similarity to the model to be trained is higher. Therefore, the similar training model is finally used as the model that finally corresponds to the model to be trained in the power grid source field.

[0129] Reference Figure 7 As shown, performing migration transformation on similar training data to obtain supplementary training data includes the following steps:

[0130] The overlapping parts of the source training data of the similar training model and the similar training data are used as the migration benchmark data;

[0131] In the target domain of the power grid, target training data corresponding to the migration benchmark data is obtained as the migration conversion benchmark data;

[0132] The part of the source training data of the similar training model except the migration benchmark data is used as the data to be migrated;

[0133] Use the minimum and maximum values ​​of the migration conversion benchmark data as interval endpoints to form a conversion interval;

[0134] Obtain an initial mapping function, satisfying that the minimum and maximum values ​​of the data obtained by substituting the migration benchmark data into the initial mapping function are used as the interval endpoints to form a verification interval covering the conversion interval;

[0135] Uniformly select at least one identification point in the interval (0, 1), multiply the value at the identification point by the initial mapping function, and obtain a secondary mapping function;

[0136] Substituting the migration benchmark data into the secondary mapping function to obtain secondary migration benchmark data, and matching the migration conversion benchmark data corresponding to the migration benchmark data to the secondary migration benchmark data;

[0137] superimposing the overlapped portion of at least one secondary migration reference data obtained using the identification point with the corresponding migration conversion reference data to obtain a determination value of the identification point;

[0138] The secondary mapping function obtained from the identification point with the largest judgment value is used as the migration mapping function;

[0139] Use the migration mapping function to perform migration transformation on the data to be migrated to obtain the migration target data;

[0140] The part of the migration target data and the migration conversion benchmark data other than the data to be trained is used as supplementary training data.

[0141] The initial mapping function is not a very precise function. As long as it satisfies the minimum and maximum values ​​of the data obtained by substituting the migration benchmark data into the initial mapping function as the interval endpoints, the verification interval formed covers the conversion interval. This is easy to achieve. Subsequently, the secondary mapping function obtained by the initial mapping function is evaluated. The evaluation is judged based on the migration situation of the migration benchmark data. The secondary mapping function with the best migration effect must be the secondary mapping function with the smallest overall gap between the secondary migration benchmark data and the corresponding migration conversion benchmark data. Therefore, this function is used as the migration mapping function. Therefore, the effect of migration mapping using the migration mapping function is better.

[0142] Due to the correspondence between the source topology scenario and the target topology scenario, there is a correspondence between the data of the two, and thus there is a correspondence between the migration benchmark data and the migration conversion benchmark data.

[0143] Furthermore, the present solution also proposes a storage medium on which a computer-readable program is stored. When the computer-readable program is called, the above-mentioned method for migrating the source domain learning strategy based on the topology change scenario is executed.

[0144] It is understandable that the storage medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid state drive (SSD).

[0145] To sum up, the advantages of the present invention are: by establishing a corresponding relationship between the source topology scenario and the target topology scenario, similar training data and similar training models are obtained, and the similar training data are migrated and transformed, thereby determining the data that needs to be called in the power grid source field, that is, similar training data. There is a high correlation between the similar training data and the model to be trained, so the training effect of the model to be trained can be guaranteed by using the data migrated and transformed from the similar training data for training. At the same time, when performing the migration and transformation, it is ensured that the converted data can meet the characteristics required by the training data of the model to be trained, thereby ensuring the normal training of the model in the target field of the power grid.

[0146] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A migration method for source domain learning strategy based on topology change scenario, characterized in that: include: Acquire a source topology scenario in a power grid source domain, where the source topology scenario is composed of at least one source training data; acquire a target topology scenario in a power grid target domain, where the target topology scenario is composed of at least one target training data; Acquire at least one source training model in the power grid source field, and acquire source training data used in training the source training model, where the source training model is a trained model; Acquire at least one target training model in the target field of the power grid, and acquire target training data used when training the target training model, where the target training model is a trained model; Based on the recognition result, a corresponding relationship is established between the source topology scene and the target topology scene; Acquire a model to be trained in a target domain of a power grid, analyze and obtain a target data type used for training the model to be trained, and based on the target data type, use the target training data used for training the model to be trained in the target domain of the power grid as the data to be trained; Determine whether the data to be trained is sufficient. If so, use the data to be trained to train the model to be trained. If not, based on the correspondence between the source topology scenario and the target topology scenario, obtain at least one source training data similar to the data to be trained in the power grid source field as similar training data; Based on the similar training data, a source training model corresponding to the model to be trained is obtained in the power grid source field as a similar training model; Based on the analysis of the training modes of similar training models and the model to be trained, similar training data is transferred and transformed to obtain supplementary training data; The model to be trained is trained using the training supplementary data and the data to be trained.

2. According to claim 1, a method for migrating source domain learning strategies based on topology change scenarios is characterized in that: The step of establishing a corresponding relationship between the source topology scene and the target topology scene based on the recognition result includes the following steps: Obtaining the source data classification standard of the source training model and obtaining the source data training purpose of the source training model; Obtain target data classification standards for the target training model, and obtain target data training purposes for the target training model; Obtain a first recognition accuracy rate of the target training data of the target training model input into the source training model, and obtain a second recognition accuracy rate of the source training data of the source training model input into the target training model; The first recognition accuracy rate is superimposed on the second recognition accuracy rate to obtain a first approximation between the source training model and the target training model; Extract features from the source data classification standard and the target data classification standard respectively, and use the overlap ratio of the features of the source data classification standard and the target data classification standard as the standard overlap degree; Extract features of the source data training purpose and the target data training purpose respectively, and use the overlap ratio of the features of the source data training purpose and the target data training purpose as the purpose overlap degree; Using the analytic hierarchy process, the weights of the first approximation, standard coincidence, and purpose coincidence are obtained; Using the approximation formula, calculate the second approximation between the source training model and the target training model; Matching the source training model with the second largest similarity to the target training model to the target training model; Establish a correspondence between the source training data of the source training model and the target training data of the corresponding target training model, and summarize the correspondence as the correspondence between the source topology scene and the target topology scene; The approximation formula is as follows: ; Where A is the second similarity between the source training model and the target training model, is the weight of the first approximation, is the weight of the standard coincidence degree, is the weight of the target overlap, a is the first approximation between the source training model and the target training model, b is the standard overlap between the source training model and the target training model, and c is the target overlap between the source training model and the target training model.

3. The method for migrating source domain learning strategies based on topology change scenarios according to claim 2 is characterized in that: The step of obtaining a model to be trained in the target field of the power grid and analyzing to obtain the target data type used for training the model to be trained comprises the following steps: Obtain the data classification standards and training purposes of the model to be trained; Extract features from the source data classification standard and the data to be classified standard respectively, and use the overlap ratio of the features of the source data classification standard and the data to be classified standard as the standard secondary overlap degree; Extract features from the source data training purpose and the data to be trained purpose respectively, and use the overlap ratio of the features of the source data training purpose and the data to be trained purpose as the secondary overlap degree of the purpose; Use the coincidence synthesis formula to calculate the similarity between the model to be trained and the source training model; The source training model with the greatest similarity to the model to be trained is used as the feature source training model; The data type of the target training data corresponding to the source training data of the feature source training model is used as the target data type; The combined formula is as follows: ; Among them, B is the similarity between the model to be trained and the source training model, d is the standard secondary overlap, and e is the target secondary overlap.

4. The method for migrating source domain learning strategies based on topology change scenarios according to claim 3 is characterized in that: The method of using the target training data in the target field of the power grid for training the model to be trained as the data to be trained comprises the following steps: In the power grid target domain, target training data of the same type as the target data type is used as the data to be trained.

5. The method for migrating source domain learning strategies based on topology change scenarios according to claim 4 is characterized in that: The step of determining whether the training data is sufficient comprises the following steps: Use the data to be trained to train the model to be trained; Feature recognition obtains at least one data feature of the data to be trained, and based on the big data, obtains at least one replaceable data, the data feature of the replaceable data being consistent with the data feature of the data to be trained; Input the replaceable data into the trained model to be trained, and count the error processing ratio of the model to be trained; Based on historical data, obtain the allowable error tolerance of the model to be trained; When the error processing ratio is greater than the allowable fault tolerance rate, the training data is insufficient, otherwise the training data is sufficient.

6. The method for migrating source domain learning strategies based on topology change scenarios according to claim 5, characterized in that: The step of acquiring at least one source training data similar to the data to be trained in the power grid source field as similar training data comprises the following steps: Acquire at least one source training data corresponding to the data to be trained in the power grid source domain as initial source training data; Feature recognition obtains at least one data feature of the initial source training data, and in the power grid source field, obtains at least one similar training data, the data features of the similar training data are consistent with the data features of the initial source training data.

7. The method for migrating source domain learning strategies based on topology change scenarios according to claim 6, characterized in that: The step of obtaining a source training model corresponding to the model to be trained in the power grid source domain as a similar training model comprises the following steps: The source training data used in training the source training model is used to count the parts that overlap with the similar training data, and use them as the similar data of the source training model; The amount of similar data of the source training model is divided by the amount of source training data used when training the source training model to obtain a similarity ratio; The source training model with the largest similarity ratio is used as the similarity training model.

8. The method for migrating source domain learning strategies based on topology change scenarios according to claim 7, characterized in that: The transfer and conversion of similar training data to obtain supplementary training data comprises the following steps: The overlapped part between the source training data of the similar training model and the similar training data is used as the migration benchmark data; In the target domain of the power grid, target training data corresponding to the migration benchmark data is obtained as the migration conversion benchmark data; The part of the source training data of the similar training model except the migration benchmark data is used as the data to be migrated; Use the minimum and maximum values ​​of the migration conversion benchmark data as interval endpoints to form a conversion interval; Obtain an initial mapping function, satisfying that the minimum and maximum values ​​of the data obtained by substituting the migration benchmark data into the initial mapping function are used as the interval endpoints to form a verification interval covering the conversion interval; Uniformly select at least one identification point in the interval (0, 1), multiply the value at the identification point by the initial mapping function, and obtain a secondary mapping function; Substituting the migration benchmark data into the secondary mapping function to obtain secondary migration benchmark data, and matching the migration conversion benchmark data corresponding to the migration benchmark data to the secondary migration benchmark data; superimposing the overlapped portion of at least one secondary migration reference data obtained using the identification point with the corresponding migration conversion reference data to obtain a determination value of the identification point; The secondary mapping function obtained from the identification point with the largest judgment value is used as the migration mapping function; Use the migration mapping function to perform migration transformation on the data to be migrated to obtain the migration target data; The part of the migration target data and the migration conversion benchmark data other than the data to be trained is used as supplementary training data.

Images