Federal large model construction method and system in field of electric power procurement

Abstract

The invention relates to the technical field of electric power purchase, in particular to a federated large model construction method and system in the field of electric power purchase. Comprises: establishing a purchase database; setting a plurality of data source points of the basic business model based on the purchase database, and setting a primary training strategy according to all the data source points and the basic business model; and obtaining an iterative service model according to the preset feedback time node, and judging whether to output a global service model or not according to the iterative service model. A plurality of specific tasks are generated by performing global analysis on business requirements in the field of power purchase, and a training strategy is dynamically adjusted based on association degrees between each specific task and different data source points, so that the training effect of each data source point on a basic business model is improved, the influence of a data island is reduced, and the training efficiency is improved. Therefore, the global business model can effectively fuse multi-party knowledge.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power procurement, in particular to a federated large model construction method and system in the field of power procurement. BACKGROUND

[0002] In the power market transaction, market subjects such as power generation enterprises and power selling companies make procurement decisions by relying on accurate analysis of data such as electricity prices and loads. At present, the construction of a decision support system using a centralized large model faces serious challenges.

[0003] The first problem is data barriers and privacy security. Power procurement involves power generation, power grids, and users, and the data such as load, offer, and contract held by them are all core business secrets and are strictly protected by privacy regulations. The data of each subject forms an island and cannot be gathered for centralized training, which makes it difficult for the model to have a global perspective and limits its performance. Secondly, the dispersion of professional knowledge leads to poor model practicability. Power procurement is a highly specialized field, and its knowledge (such as offer strategy and risk control) is scattered among various market subjects. General large models lack these deep and distributed professional knowledge, and often have understanding biases and unreliable decision-making problems when dealing with specific tasks such as offer and price, and contract decomposition. SUMMARY

[0004] The purpose of the present application is to solve the above technical problems, and the present application provides a federated large model construction method and system in the field of power procurement, aiming to improve the construction efficiency and accuracy of intelligent models in the field of power procurement.

[0005] In some embodiments of the present application, a plurality of specific tasks are generated by globally analyzing the business requirements in the field of power procurement, and the training strategy is dynamically adjusted based on the correlation between each specific task and different data source points (i.e. participants in the field of power procurement), which improves the training effect of each data source point on the basic business model, reduces the impact of data islands, and enables the global business model to effectively integrate multi-party knowledge.

[0006] In some embodiments of the present application, each basic sub-model is periodically evaluated, and the training parameters are adjusted in a timely manner to avoid deviations in the iteration process, reduce the overall training cost, improve the iteration optimization efficiency of the intelligent model corresponding to each specific task, and ensure the construction efficiency and accuracy of the intelligent model in the field of power procurement.

[0007] In some embodiments of the present application, a federated large model construction method in the field of power procurement is provided, which comprises:

[0008] Establishing a procurement database;

[0009] Based on the procurement database, multiple data source points are set for the basic business model, and a first-level training strategy is set according to all data source points and the basic business model.

[0010] The iterative business model is obtained based on the preset feedback time node, and the global business model is output based on the iterative business model.

[0011] In some embodiments of this application, a basic business model is defined, including:

[0012] Multiple business substructures are generated based on the procurement business database;

[0013] Establish a business substructure sequence A, A=(a1, a2…a…) i …a n ), where a i Let i be the i-th business substructure; n is the number of business substructures;

[0014] Based on the business substructure sequence A, ai is sequentially set as the target substructure;

[0015] The basic data for generating the target substructure is based on the procurement database;

[0016] Generate a basic sub-model of the target substructure based on the basic data;

[0017] Generate the basic sub-models for each business substructure in sequence;

[0018] Establish the basic sub-model sequence W, W=(w1,w2…w i …w n ), where w i This serves as the basic sub-model for the i-th business substructure.

[0019] Generate a basic business model based on all the basic sub-models.

[0020] In some embodiments of this application, a primary training strategy is set, including:

[0021] Based on the business substructure sequence A, ai is sequentially set as the substructure to be associated;

[0022] Create a data source point sequence B, B=(b1, b2, ..., b...). i …b m ), where b i Let m be the i-th data source point; m is the number of data source points.

[0023] Based on the data source point sequence B, set bi as the target data source point in sequence;

[0024] Obtain the data feature package of the target data source;

[0025] Generate association training values ​​c between the substructure to be associated and the target data source point based on the data feature package;

[0026] c=[ η i *j i] ;

[0027] Where θ1 is the number of related indicators; η i Let j be the influence factor of the i-th correlation indicator; i It is the reference value of the i-th associated indicator generated based on the data feature package;

[0028] Generate association training values ​​between the substructures to be associated and each data source point in sequence;

[0029] Set the fitting sub-strategy for the structure to be associated based on all associated training values;

[0030] Generate fitting sub-strategies for each business substructure in sequence;

[0031] A first-level training strategy is generated based on all fitted sub-strategies.

[0032] In some embodiments of this application, the iterative business model is obtained according to a preset feedback time node, including:

[0033] Multiple training cycles are preset, and the end time of each training cycle is set as the feedback time node;

[0034] Based on the business substructure sequence A, ai is sequentially set as the substructure to be trained;

[0035] Define the base sub-model of the substructure to be trained as the target sub-model;

[0036] At the start of the current training cycle, the target sub-model is sent to each data source point for iterative training.

[0037] Obtain the training results of each data source point for the target sub-model at the current feedback time point;

[0038] Based on all training results, generate training sub-models for each target sub-model at each data source point;

[0039] The fitting strategy for the substructure to be trained is set as the first-level fitting strategy;

[0040] Based on the first-level fitting strategy and all trained sub-models, an iterative sub-model of the substructure to be trained at the current feedback time node is generated.

[0041] The iterative sub-models of each business substructure at the current feedback time node are generated sequentially;

[0042] Generate the iterative business model for the current feedback time node based on all iterative sub-modules.

[0043] In some embodiments of this application, determining whether to output a global business model based on an iterative business model includes:

[0044] Based on the business substructure sequence A, ai is sequentially set as the substructure to be evaluated;

[0045] Obtain the iterative sub-model of the substructure to be evaluated at the current feedback time point;

[0046] Send the iterative sub-model to each data source point;

[0047] Obtain test data packets from each data source point;

[0048] Generate the model running values ​​of the substructure to be evaluated at the current feedback time point based on all test data packets;

[0049] Generate the model runtime values ​​for each business substructure sequentially;

[0050] Determine whether to output the global business model based on the running values ​​of all models.

[0051] In some embodiments of this application, generating the model running value of the substructure to be evaluated at the current feedback time node includes:

[0052] Based on the data source point sequence B, set bi as the target data source point in sequence;

[0053] Obtain the test data packet from the model data source point;

[0054] Generate training fit values ​​for the target data source based on the test data package;

[0055] The iterative sub-models of the substructure to be evaluated are sequentially generated, along with the matching evaluation values ​​of each data source point.

[0056] The model running value d is generated based on all the matching evaluation values;

[0057] d=[ μ i *k i ];

[0058] Where m is the number of data source points; μ i Let k be the influence factor of the i-th data source point; i The value represents the fit evaluation between the iterative sub-model of the substructure to be evaluated and the i-th data source point.

[0059] In some embodiments of this application, determining whether to output a global business model based on all model runtime values ​​includes:

[0060] Establish a sequence of model running values ​​D, where D = (d1, d2, ..., dn) at the current feedback time point.i …d n ), where d i is the model running value of the iterative sub-model of the i-th service sub-structure at the current feedback time node; n is the number of service sub-structures;

[0061] The preset model running value threshold D1

[0062] If d i > D1, generate an output instruction for the i-th service sub-structure, and generate a first-level sub-model of the i-th service sub-structure according to the output instruction.

[0063] If d i < D1, generate a first-level iteration instruction for the i-th service sub-structure, and set the i-th service sub-structure as a sub-structure to be judged according to the first-level iteration instruction.

[0064] Obtain the first-level sub-models of each service sub-structure.

[0065] Construct a global service model based on all the first-level sub-models.

[0066] In some embodiments of the present application, the first-level iteration instruction includes:

[0067] Generate a corrected evaluation value f for the current sub-structure to be judged;

[0068] f = g * Y1(i) * (k 1i - k 2i );

[0069] g = U1 * Y2(i) * (k 1i - k')];

[0070] Where g is a correction compensation coefficient; Y1(i) is a selection coefficient. If (k 1i - k 2i ) > 0, Y1(i) = 0; if (k 1i - k 2i ) < 0, Y1(i) = -1; k 1i is the fitting evaluation value of the iterative sub-model of the sub-structure to be judged and the i-th data source point at the current feedback time node; k 2i is the fitting evaluation value of the iterative sub-model of the sub-structure to be judged and the i-th data source point at the previous feedback time node; U1 is a preset first conversion coefficient; Y2(i) is a selection coefficient; if (k 1i - k') > 0, Y2(i) = 0; if (k 1i - k') < 0, Y2(i) = 1 / (k 1i - k'); k' is a preset fitting evaluation value threshold;

[0071] Preset the correction evaluation value threshold F1;

[0072] If f > F1, generate a first-level correction instruction for the currently to-be-judged sub-structure;

[0073] If f < F1, the currently to-be-judged sub-structure executes the fitting sub-strategy corresponding to the current feedback time node.

[0074] In some embodiments of the present application, a federated large model construction system in the field of power procurement is provided, including:

[0075] A central control unit, used to establish a procurement database;

[0076] A training unit, used to set a basic business model and multiple data source points according to the procurement database, and set a first-level training strategy according to all data source points and the basic business model;

[0077] An evaluation unit, used to obtain an iterative business model according to a preset feedback time node, and judge whether to output a global business model according to the iterative business model.

[0078] In some embodiments of the present application, the training unit further includes:

[0079] A first training module, used to generate multiple business sub-structures according to the procurement business library;

[0080] Establish a business sub-structure sequence A, A=(a1, a2…a i …a n ), where a i is the i-th business sub-structure; n is the number of business sub-structures;

[0081] Set ai as the target sub-structure in sequence according to the business sub-structure sequence A;

[0082] Generate basic data of the target sub-structure based on the procurement database;

[0083] Generate a basic sub-model of the target sub-structure according to the basic data;

[0084] Generate basic sub-models of each business sub-structure in sequence;

[0085] Establish a basic sub-model sequence W, W=(w1, w2…w i …w n ), where w i is the basic sub-model of the i-th business sub-structure;

[0086] Generate a basic business model according to all basic sub-models;

[0087] A second training module, used to set ai as the to-be-associated sub-structure in sequence according to the business sub-structure sequence A;

[0088] Create a data source point sequence B, B=(b1, b2, ..., b...). i …b m ), where b i Let m be the i-th data source point; m is the number of data source points.

[0089] Based on the data source point sequence B, set bi as the target data source point in sequence;

[0090] Obtain the data feature package of the target data source;

[0091] Generate association training values ​​c between the substructure to be associated and the target data source point based on the data feature package;

[0092] c=[ η i *j i] ;

[0093] Where θ1 is the number of related indicators; η i Let j be the influence factor of the i-th correlation indicator; i It is the reference value of the i-th associated indicator generated based on the data feature package;

[0094] Generate association training values ​​between the substructures to be associated and each data source point in sequence;

[0095] Set the fitting sub-strategy for the structure to be associated based on all associated training values;

[0096] Generate fitting sub-strategies for each business substructure in sequence;

[0097] A first-level training strategy is generated based on all fitted sub-strategies.

[0098] Compared with existing technologies, the advantages of the federated large-scale model construction method and system in the field of power procurement proposed in this application are as follows:

[0099] By conducting a global analysis of business needs in the power procurement field, various specific tasks are generated. Based on the correlation between each specific task and different data source points (i.e., participants in the power procurement field), the training strategy is dynamically adjusted to improve the training effect of each data source point on the basic business model, reduce the impact of data silos, realize cross-enterprise and cross-regional sharing and localized training of the model, and ensure data security. It also fully leverages the advantages of data resources from all parties, enabling the global business model to effectively integrate knowledge from multiple parties.

[0100] By periodically evaluating each basic sub-model and adjusting training parameters in a timely manner, deviations in the iteration process can be avoided, the overall training cost can be reduced, the iterative optimization efficiency of the intelligent models corresponding to each specific task can be improved, and the efficiency and accuracy of building intelligent models in the power procurement field can be guaranteed. Attached Figure Description

[0101] Figure 1 This is a flowchart illustrating a method for constructing a federalized large model in the field of power procurement, as described in a preferred embodiment of this application. Detailed Implementation

[0102] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.

[0103] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0104] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0105] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0106] like Figure 1 As shown in the preferred embodiment of this application, a method for constructing a federated large model in the field of power procurement includes:

[0107] S101: Establish a procurement database;

[0108] S102: Based on the procurement database, set up multiple data source points for the basic business model, and set up a first-level training strategy according to all data source points and the basic business model.

[0109] S103: Obtain the iterative business model based on the preset feedback time node, and determine whether to output the global business model based on the iterative business model.

[0110] Specifically, a procurement database is constructed by aggregating publicly available data on power procurement from various participants in the power business sector (different power generation companies, grid-related operation and maintenance companies, suppliers, etc.).

[0111] Specifically, multiple data source points are set up based on all participants, where each data source point represents a reference that has private procurement data (i.e., data about procurement operations that is not publicly disclosed).

[0112] Specifically, the basic business model is established, including:

[0113] Multiple business substructures are generated based on the procurement business database;

[0114] Establish a business substructure sequence A, A=(a1, a2…a…) i …a n ), where a i Let i be the i-th business substructure; n is the number of business substructures;

[0115] Based on the business substructure sequence A, ai is sequentially set as the target substructure;

[0116] The basic data for generating the target substructure is based on the procurement database;

[0117] Generate a basic sub-model of the target substructure based on the basic data;

[0118] Generate the basic sub-models for each business substructure in sequence;

[0119] Establish the basic sub-model sequence W, W=(w1,w2…w i …w n ), where w i This serves as the basic sub-model for the i-th business substructure.

[0120] Generate a basic business model based on all the basic sub-models.

[0121] Specifically, by analyzing the procurement business process data in the business procurement business database, various specific tasks are generated. Based on all the specific tasks, a business substructure sequence is generated, where each business substructure represents a specific task.

[0122] Specifically, the tasks include, but are not limited to, tasks related to various stages of the power procurement and bidding process (bidding, tendering, bid opening, bid evaluation and awarding) such as quantity quotation and contract breakdown.

[0123] Specifically, data in the procurement database is filtered to obtain data related to the target substructure, which is then set as the foundational data for the target substructure. By analyzing this foundational data, an initial intelligent model is generated that provides a basic understanding of the technical terminology and fundamental logic corresponding to the target substructure. This model is then set as the foundational sub-model for the target substructure.

[0124] Specifically, by constructing initial intelligent models corresponding to each business substructure, a basic cloud business model is formed. This basic business model is then sent to various data source points, and the private data from each data source point is used to optimize and iterate the progress of the basic business model. The security of the data at each data source point is ensured.

[0125] Specifically, a primary training strategy is defined, including:

[0126] Based on the business substructure sequence A, ai is sequentially set as the substructure to be associated;

[0127] Create a data source point sequence B, B=(b1, b2, ..., b...). i …b m ), where b i Let m be the i-th data source point; m is the number of data source points.

[0128] Based on the data source point sequence B, set bi as the target data source point in sequence;

[0129] Obtain the data feature package of the target data source;

[0130] Generate association training values ​​c between the substructure to be associated and the target data source point based on the data feature package;

[0131] c=[ η i *j i] ;

[0132] Where θ1 is the number of related indicators; η i Let j be the influence factor of the i-th correlation indicator; i It is the reference value of the i-th associated indicator generated based on the data feature package;

[0133] Generate association training values ​​between the substructures to be associated and each data source point in sequence;

[0134] Set the fitting sub-strategy for the structure to be associated based on all associated training values;

[0135] Generate fitting sub-strategies for each business substructure in sequence;

[0136] A first-level training strategy is generated based on all fitted sub-strategies.

[0137] Specifically, the data feature package includes parameters such as the amount of private data of the target data source for each business substructure, business call frequency parameters, and execution effect (i.e., the processing efficiency and processing accuracy of the specific tasks corresponding to each business substructure).

[0138] Specifically, the number of correlation indicators includes, but is not limited to, the amount of private data from the target data source point for the substructure to be correlated, the frequency of business calls, the processing efficiency for the substructure to be correlated, and data feature parameters related to model updates such as processing accuracy. By quantifying each correlation indicator, the reference values ​​of each correlation indicator are made to be within the same range, and the larger the parameter value of each correlation indicator, the better the training effect of the private data from the target data source point for the substructure to be correlated.

[0139] Specifically, corresponding influence factors are set according to the degree of influence of each related indicator on the model update effect. The greater the degree of influence, the larger the reference value of the corresponding influence factor. The mapping relationship between the two can be set according to historical parameters.

[0140] Specifically, a larger correlation training value indicates a better training effect of the private data of the target data source corresponding to the substructure to be correlated. Dynamic contribution weights are set for the iterative sub-models output by each data source based on the correlation training value; the larger the correlation training value, the larger the corresponding dynamic contribution weight. A fitting sub-strategy for the substructure to be correlated is constructed based on all dynamic contribution weights.

[0141] It is understandable that in the above embodiments, by performing a global analysis of the business needs in the power procurement field to generate a variety of specific tasks, and based on the correlation between each specific task and different data source points (i.e., participants in the power procurement field), the training strategy is dynamically adjusted to improve the training effect of each data source point on the basic business model, reduce the impact of data silos, and enable the global business model to effectively integrate knowledge from multiple parties.

[0142] In a preferred embodiment of this application, the iterative business model is obtained according to a preset feedback time node, including:

[0143] Multiple training cycles are preset, and the end time of each training cycle is set as the feedback time node;

[0144] Based on the business substructure sequence A, ai is sequentially set as the substructure to be trained;

[0145] Define the base sub-model of the substructure to be trained as the target sub-model;

[0146] At the start of the current training cycle, the target sub-model is sent to each data source point for iterative training.

[0147] Obtain the training results of each data source point for the target sub-model at the current feedback time point;

[0148] Based on all training results, generate training sub-models for each target sub-model at each data source point;

[0149] The fitting strategy for the substructure to be trained is set as the first-level fitting strategy;

[0150] Based on the first-level fitting strategy and all trained sub-models, an iterative sub-model of the substructure to be trained at the current feedback time node is generated.

[0151] The iterative sub-models of each business substructure at the current feedback time node are generated sequentially;

[0152] Generate the iterative business model for the current feedback time node based on all iterative sub-modules.

[0153] Specifically, the training cycle duration can be set based on historical parameters, and the training duration is the maximum duration for the basic business model to complete a single optimization training at each data source point.

[0154] Specifically, once each data source obtains the basic business model, it uses local private data to optimize and iteratively train each basic sub-model in the basic business model, and outputs the corresponding trained sub-model.

[0155] Specifically, all training sub-models output from each data source point are aggregated according to the corresponding business sub-structure. All training sub-models of a single business sub-structure are fitted according to the dynamic contribution weights of each data source point set in the corresponding fitting sub-strategy, thereby generating the iterative sub-model corresponding to that business sub-structure.

[0156] Specifically, iterative sub-models for each business substructure at the current feedback time point are generated sequentially, corresponding business iterative models are constructed, and these models are resent to each data source point for evaluation and execution. Corresponding test data packages are generated, and based on all test data packages, it is determined whether to continue optimization and iterative training in the next training cycle. This process is repeated until the global business model is output.

[0157] It is understandable that in the above embodiments, by distributing the basic business model to various data source points, the model can be shared and trained locally across enterprises and regions, while ensuring data security and giving full play to the advantages of data resources from all parties, so that the global business model can effectively integrate knowledge from multiple parties.

[0158] In a preferred embodiment of this application, determining whether to output a global business model based on the iterative business model includes:

[0159] Based on the business substructure sequence A, a is set sequentially. i The substructure to be evaluated;

[0160] Obtain the iterative sub-model of the substructure to be evaluated at the current feedback time point;

[0161] Send the iterative sub-model to each data source point;

[0162] Obtain test data packets from each data source point;

[0163] Generate the model running values ​​of the substructure to be evaluated at the current feedback time point based on all test data packets;

[0164] Generate the model runtime values ​​for each business substructure sequentially;

[0165] Determine whether to output the global business model based on the running values ​​of all models.

[0166] Specifically, the business iteration model is sent to each data source point and run for testing, thereby generating test data packages for each data source point. Based on all test data packages, the difference between the running output of the iterative sub-model at each data source point and the actual needs of that data source point is generated. The smaller the difference, the greater the corresponding fit evaluation value. The mapping relationship between the two can be set according to historical parameters.

[0167] Specifically, the larger the training fit value, the better the performance of the iterative sub-model at the corresponding data source point.

[0168] Specifically, generating the model running values ​​of the substructure to be evaluated at the current feedback time point includes:

[0169] Based on the data source point sequence B, set bi as the target data source point in sequence;

[0170] Obtain the test data packet from the model data source point;

[0171] Generate training fit values ​​for the target data source based on the test data package;

[0172] The iterative sub-models of the substructure to be evaluated are sequentially generated, along with the matching evaluation values ​​of each data source point.

[0173] The model running value d is generated based on all the matching evaluation values;

[0174] d=[ μ i *k i ];

[0175] Among them, m is the number of data source points; μ i is the influence factor of the i-th data source point; k i is the fitness evaluation value between the iterative sub-model of the sub-structure to be evaluated and the i-th data source point.

[0176] Specifically, the influence factors of each data source point are set according to the training correlation values between the sub-structure to be evaluated and each data source point. The greater the training correlation value, the smaller the value of the corresponding influence factor.

[0177] Specifically, the larger the model running value, the higher the fitness of the iterative sub-model of the sub-structure to be evaluated to the global at the current feedback time node, and the better the running effect.

[0178] Specifically, it is judged whether to output the global business model according to all model running values, including:

[0179] Establish a sequence D of model running values at the current feedback time node, D = (d1, d2…d i …d n ), where d i is the model running value of the iterative sub-model of the i-th business sub-structure at the current feedback time node; n is the number of business sub-structures;

[0180] Preset a model running value threshold D1

[0181] If d i > D1, generate an output instruction for the i-th business sub-structure, and generate a first-level sub-model of the i-th business sub-structure according to the output instruction;

[0182] If d i < D1, generate a first-level iteration instruction for the i-th business sub-structure, and set the i-th business sub-structure as a sub-structure to be judged according to the first-level iteration instruction;

[0183] Obtain the first-level sub-models of each business sub-structure;

[0184] Construct a global business model according to all first-level sub-models.

[0185] Specifically, the model running value threshold can be set according to historical parameters. If the model running value of the current iterative sub-model is greater than the preset model running value threshold, it means that the iterative sub-model can complete the global adaptation of each participant in the power procurement field. The iterative sub-model can be used as a component (i.e., the first-level sub-model) of the corresponding business sub-structure in the global business model, and the optimization iterative training of the business sub-model is stopped, thereby reducing the overall training cost.

[0186] Specifically, a global business model is constructed by obtaining the first-level sub-models corresponding to all business sub-structures output at different feedback time nodes.

[0187] In the preferred embodiments of the embodiments of the present application, the first-level iteration instruction includes:

[0188] Generate a corrected evaluation value f for the currently to-be-judged substructure;

[0189] f = g * Y1(i) * (k 1i - k 2i );

[0190] g = U1 * Y2(i) * (k 1i - k')];

[0191] Wherein, g is a correction compensation coefficient; Y1(i) is a selection coefficient. If (k 1i - k 2i ) > 0, Y1(i) = 0; if (k 1i - k 2i ) < 0, Y1(i) = -1; k 1i is the fitting evaluation value of the iterative sub-model of the to-be-judged substructure and the i-th data source point at the current feedback time node; k 2i is the fitting evaluation value of the iterative sub-model of the to-be-judged substructure and the i-th data source point at the previous feedback time node; U1 is a preset first conversion coefficient; Y2(i) is a selection coefficient. If (k 1i - k') > 0, Y2(i) = 0; if (k 1i - k') < 0, Y2(i) = 1 / (k 1i - k'); k' is a preset fitting evaluation value threshold;

[0192] Preset a corrected evaluation value threshold F1;

[0193] If f > F1, generate a first-level correction instruction for the currently to-be-judged substructure;

[0194] If f < F1, the currently to-be-judged substructure executes the fitting sub-strategy corresponding to the current feedback time node.

[0195] Specifically, the corrected evaluation value threshold can be set according to historical parameters.

[0196] Specifically, by presetting the first conversion coefficient, the correction compensation coefficient g is within a preset value range, and the value of Y2(i) * (k 1i - k')] is larger, and the corresponding value of the correction compensation coefficient g is larger. The mapping relationship between the two can be set according to historical parameters, and the value of the correction compensation coefficient g is always greater than 1.

[0197] Specifically, the fit evaluation threshold can be set based on historical parameters. If the fit evaluation value of the iterative sub-model at the current data source point is less than the preset fit evaluation value threshold, it means that the iterative sub-model is completely unable to adapt to the private data in the current data source point, and its ability to process procurement tasks is poor.

[0198] Specifically, the larger the correction evaluation value, the more likely there is a deviation in the training strategy of the intelligent model for the substructure to be judged. When the correction evaluation value is greater than the preset correction evaluation value threshold, the fitting sub-strategy of the substructure to be judged needs to be corrected in a timely manner according to the first-level correction instruction, so as to ensure the accuracy of the construction of the intelligent model corresponding to the substructure to be judged.

[0199] It is understood that in the above embodiments, by periodically evaluating each basic sub-model and adjusting the training parameters in a timely manner, deviations in the iteration process can be avoided, the overall training cost can be reduced, the iterative optimization efficiency of the intelligent model corresponding to each specific task can be improved, and the efficiency and accuracy of building intelligent models in the field of power procurement can be guaranteed.

[0200] In another preferred embodiment of the federated large-scale model construction method in the field of power procurement based on any of the above preferred embodiments, this preferred embodiment provides a federated large-scale model construction system in the field of power procurement, including:

[0201] The central control unit is used to establish the procurement database;

[0202] The training unit is used to set up a basic business model and multiple data source points based on the procurement database, and to set up a primary training strategy based on all data source points and the basic business model.

[0203] The evaluation unit is used to obtain the iterative business model based on the preset feedback time node, and to determine whether to output the global business model based on the iterative business model.

[0204] In a preferred embodiment of this application, the training unit further includes:

[0205] The first training module is used to generate multiple business substructures based on the procurement business database;

[0206] Establish a business substructure sequence A, A=(a1, a2…a…) i …a n ), where a i Let i be the i-th business substructure; n is the number of business substructures;

[0207] Based on the business substructure sequence A, ai is sequentially set as the target substructure;

[0208] The basic data for generating the target substructure is based on the procurement database;

[0209] Generate a basic sub-model of the target substructure based on the basic data;

[0210] Generate the basic sub-models for each business substructure in sequence;

[0211] Establish the basic sub-model sequence W, W=(w1,w2…w i …w n ), where w i This serves as the basic sub-model for the i-th business substructure.

[0212] Generate a basic business model based on all basic sub-models;

[0213] The second training module is used to sequentially set ai as the substructure to be associated according to the business substructure sequence A;

[0214] Create a data source point sequence B, B=(b1, b2, ..., b...). i …b m ), where b i Let m be the i-th data source point; m is the number of data source points.

[0215] Based on the data source point sequence B, set bi as the target data source point in sequence;

[0216] Obtain the data feature package of the target data source;

[0217] Generate association training values ​​c between the substructure to be associated and the target data source point based on the data feature package;

[0218] c=[ η i *j i] ;

[0219] Where θ1 is the number of related indicators; η i Let j be the influence factor of the i-th correlation indicator; i It is the reference value of the i-th associated indicator generated based on the data feature package;

[0220] Generate association training values ​​between the substructures to be associated and each data source point in sequence;

[0221] Set the fitting sub-strategy for the structure to be associated based on all associated training values;

[0222] Generate fitting sub-strategies for each business substructure in sequence;

[0223] A first-level training strategy is generated based on all fitted sub-strategies.

[0224] Based on the first concept of this application, a variety of specific tasks are generated by performing a global analysis of business needs in the power procurement field. Based on the correlation between each specific task and different data source points (i.e., participants in the power procurement field), the training strategy is dynamically adjusted to improve the training effect of each data source point on the basic business model, reduce the impact of data silos, and enable the global business model to effectively integrate knowledge from multiple parties.

[0225] According to the second concept of this application, by periodically evaluating each basic sub-model, the training parameters can be adjusted in a timely manner to avoid deviations in the iteration process, reduce the overall training cost, improve the iterative optimization efficiency of the intelligent models corresponding to each specific task, and ensure the efficiency and accuracy of building intelligent models in the field of power procurement.

[0226] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of this application, and these improvements and substitutions should also be considered within the scope of protection of this application.

Claims

1. A method for constructing a federated large-scale model in the field of power procurement, characterized in that, include: Establish a procurement database; Based on the procurement database, multiple data source points are set for the basic business model, and a first-level training strategy is set according to all data source points and the basic business model. The iterative business model is obtained based on the preset feedback time node, and the global business model is output based on the iterative business model.

2. The method for constructing a federated large-scale model in the field of power procurement as described in claim 1, characterized in that, Establish a basic business model, including: Multiple business substructures are generated based on the procurement business database; Establish a business substructure sequence A, A=(a1, a2…a…) i …a n ), where a i Let i be the i-th business substructure; n is the number of business substructures; Based on the business substructure sequence A, ai is sequentially set as the target substructure; The basic data for generating the target substructure is based on the procurement database; Generate a basic sub-model of the target substructure based on the basic data; Generate the basic sub-models for each business substructure in sequence; Establish the basic sub-model sequence W, W=(w1,w2…w i …w n ), where w i This serves as the basic sub-model for the i-th business substructure. Generate a basic business model based on all the basic sub-models.

3. The method for constructing a federated large-scale model in the field of power procurement as described in claim 2, characterized in that, Define the primary training strategy, including: Based on the business substructure sequence A, ai is sequentially set as the substructure to be associated; Create a data source point sequence B, B=(b1, b2, ..., b...). i …b m ), where b i Let m be the i-th data source point; m is the number of data source points. Based on the data source point sequence B, set bi as the target data source point in sequence; Obtain the data feature package of the target data source; Generate association training values ​​c between the substructure to be associated and the target data source point based on the data feature package; c=[ or i *j i] ? Where θ1 is the number of related indicators; η i Let j be the influence factor of the i-th correlation indicator; i It is the reference value of the i-th associated indicator generated based on the data feature package; Generate association training values ​​between the substructures to be associated and each data source point in sequence; Set the fitting sub-strategy for the structure to be associated based on all associated training values; Generate fitting sub-strategies for each business substructure in sequence; A first-level training strategy is generated based on all fitted sub-strategies.

4. The method for constructing a federated large-scale model in the field of power procurement as described in claim 3, characterized in that, The iterative business model is obtained based on the preset feedback time points, including: Multiple training cycles are preset, and the end time of each training cycle is set as the feedback time node; Based on the business substructure sequence A, ai is sequentially set as the substructure to be trained; Define the base sub-model of the substructure to be trained as the target sub-model; At the start of the current training cycle, the target sub-model is sent to each data source point for iterative training. Obtain the training results of each data source point for the target sub-model at the current feedback time point; Based on all training results, generate training sub-models for each target sub-model at each data source point; The fitting strategy for the substructure to be trained is set as the first-level fitting strategy; Based on the first-level fitting strategy and all trained sub-models, an iterative sub-model of the substructure to be trained at the current feedback time node is generated. The iterative sub-models of each business substructure at the current feedback time node are generated sequentially; Generate the iterative business model for the current feedback time node based on all iterative sub-modules.

5. The method for constructing a federated large-scale model in the field of power procurement as described in claim 4, characterized in that, Determine whether to output the global business model based on the iterative business model, including: Based on the business substructure sequence A, ai is sequentially set as the substructure to be evaluated; Obtain the iterative sub-model of the substructure to be evaluated at the current feedback time point; Send the iterative sub-model to each data source point; Obtain test data packets from each data source point; Generate the model running values ​​of the substructure to be evaluated at the current feedback time point based on all test data packets; Generate the model runtime values ​​for each business substructure sequentially; Determine whether to output the global business model based on the running values ​​of all models.

6. The method for constructing a federated large-scale model in the field of power procurement as described in claim 5, characterized in that, Generate the model running values ​​for the substructure to be evaluated at the current feedback time point, including: Based on the data source point sequence B, set bi as the target data source point in sequence; Obtain the test data packet from the model data source point; Generate training fit values ​​for the target data source based on the test data package; Generate the fitness evaluation values of the iterative sub-models of the sub-structures to be evaluated and each data source point in sequence; Generate the model running value d based on all the fitness evaluation values; d=[ m i *k i ]; Where m is the number of data source points; μ i Let k be the influence factor of the i-th data source point; i The value represents the fit evaluation between the iterative sub-model of the substructure to be evaluated and the i-th data source point.

7. The method for constructing a federated large-scale model in the field of power procurement as described in claim 6, characterized in that, Judge whether to output the global business model based on all the model running values, including: Establish a sequence of model running values ​​D, where D = (d1, d2, ..., dn) at the current feedback time point. i …d n ), where d i This represents the model running value of the iterative sub-model of the i-th business substructure at the current feedback time node; n is the number of business substructures. Preset the model running value threshold D1 If d i >D1, generate the output instruction for the i-th business substructure, and generate the first-level sub-model of the i-th business substructure based on the output instruction; If d i <D1, generate a first-level iteration instruction for the i-th service sub-structure, and set the i-th service sub-structure as the sub-structure to be judged according to the first-level iteration instruction; Obtain the first-level sub-models of each business sub-structure; Construct the global business model based on all the first-level sub-models.

8. The method for constructing a federated large-scale model in the field of power procurement as described in claim 6, characterized in that, The first-level iteration instruction includes: Generate the correction evaluation value f of the currently to-be-judged sub-structure; f=g*[ Y1(i)*(k 1i -k 2i )]; g=U1*[ Y2(i)*(k 1i -k')]; Where g is the correction compensation coefficient; Y1(i) is the selection coefficient, if (k 1i -k 2i If (k)>0, Y1(i)=0; if (k)>0, Y1(i)=0; 1i -k 2i If ) < 0, Y1(i) = -1; k 1i k represents the matching evaluation value between the iterative sub-model of the substructure to be judged and the i-th data source point at the current feedback time node. 2i U1 is the matching evaluation value between the iterative sub-model of the substructure to be judged and the i-th data source point at the previous feedback time node; U1 is the preset first conversion coefficient; Y2(i) is the selection coefficient; if (k 1i -k')>0, Y2(i)=0; if (k 1i -k')<0,Y2(i)=1 / (k 1i -k');k' is the preset threshold for the matching evaluation value; Preset the correction evaluation value threshold F1; If f > F1, generate the first-level correction instruction for the currently to-be-judged sub-structure; If f < F1, the currently to-be-judged sub-structure executes the fitting sub-strategy corresponding to the current feedback time node.

9. A federated large-scale model construction system for the field of power procurement, employing the federated large-scale model construction method for the field of power procurement as described in any one of claims 1-8, characterized in that, Including: The central control unit is used to establish the procurement database; The training unit is used to set the basic business model and multiple data source points according to the procurement database, and set the first-level training strategy according to all the data source points and the basic business model; The evaluation unit is used to obtain the iterative business model according to the preset feedback time node, and judge whether to output the global business model according to the iterative business model.

10. The federated large-scale model construction system for the field of power procurement as described in claim 9, characterized in that, The training unit further includes: Establish a business substructure sequence A, A=(a1, a2…a…) i …a n ), where a i Let i be the i-th business substructure; n is the number of business substructures; The first training module is used to generate multiple business sub-structures according to the procurement business library; Set ai as the target sub-structure in sequence according to the business sub-structure sequence A; Generate the basic data of the target sub-structure based on the procurement database; Generate the basic sub-model of the target sub-structure according to the basic data; Establish the basic sub-model sequence W, W=(w1,w2…w i …w n ), where w i This serves as the basic sub-model for the i-th business substructure. Generate the basic sub-models of each business sub-structure in sequence; Generate the basic business model according to all the basic sub-models; Create a data source point sequence B, B=(b1, b2, ..., b...). i …b m ), where b i Let m be the i-th data source point; m is the number of data source points. The second training module is used to set ai as the to-be-associated sub-structure in sequence according to the business sub-structure sequence A; Set bi as the target data source point in sequence according to the data source point sequence B; Obtain the data feature package of the target data source point; c=[ or i *j i] ? Where θ1 is the number of related indicators; η i Let j be the influence factor of the i-th correlation indicator; i It is the reference value of the i-th associated indicator generated based on the data feature package; Generate the association training value c of the to-be-associated sub-structure and the target data source point according to the data feature package; Generate the association training values of the to-be-associated sub-structure and each data source point in sequence; Set the fitting sub-strategy of the to-be-associated structure according to all the association training values; Generate the fitting sub-strategies of each business sub-structure in sequence; Generate the first-level training strategy according to all the fitting sub-strategies.

Images