Machine learning driven complex transformer electromagnetic design optimization solving method and system

Abstract

The present application belongs to the technical field of transformers, and relates to a machine learning driven complex transformer electromagnetic design optimization solving method and system. The present application trains a model based on historical data to learn design rules, predicts initial values of variables for new tasks, and generates a contracted search domain in combination with experience rules to provide a high-quality starting point and precise boundary for subsequent optimization. In the solving stage, an improved particle swarm algorithm is used to initialize a population in the contracted domain with the predicted values as the center, and a constraint feasibility priority rule is introduced to guide the search direction, supplemented by a backtracking mechanism to avoid falling into local optimum, so as to focus the search on high-potential areas. Through iterative electromagnetic calculation and constraint verification, an optimized design scheme that meets all engineering requirements is efficiently output, effectively solving the problems of low solving efficiency and difficulty in obtaining feasible solutions caused by the multiple transformer electromagnetic design variables and complex constraints in the existing transformer electromagnetic design method, significantly improving the design efficiency and scheme feasibility, and reducing the calculation cost.

Description

Technical Field

[0001] This invention belongs to the field of transformer technology, specifically relating to a machine learning-driven method and system for optimizing the electromagnetic design of complex transformers. Background Technology

[0002] As a core component of a power system, the quality of a transformer's electromagnetic design directly determines its operational efficiency, reliability, and long-term safety. The electromagnetic design process involves determining core electromagnetic parameters such as core structure, winding parameters, current density, leakage flux distribution, and electromagnetic losses, while simultaneously meeting multiple stringent engineering constraints, including temperature rise limits, insulation class, and material selection. This leads to complex combinatorial optimization problems in electromagnetic design, characterized by high variable dimensionality, complex constraints, and highly nonlinear, multimodal objective functions (such as cost and losses).

[0003] Existing optimization schemes typically employ metaheuristic intelligent optimization algorithms, such as particle swarm optimization and genetic algorithms. However, directly applying these metaheuristic algorithms to transformer electromagnetic design still has significant shortcomings. Specifically: the algorithms perform global searches without prior information guidance, resulting in an excessively large search space when faced with a large number of continuous and discrete mixed variables, leading to slow and unstable convergence. Furthermore, existing algorithms often use simple penalty functions to handle complex and coupled engineering constraints, making it difficult to stably find feasible solutions that satisfy all constraints within a finite number of iterations, easily generating a large number of invalid solutions or getting trapped in local optima. Moreover, the overall optimization process lacks a targeted guidance mechanism, resulting in high computational costs and failing to meet the urgent demands of modern engineering design for real-time performance and efficiency.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] One objective of this invention is to effectively solve the problems of low solution efficiency and difficulty in obtaining feasible solutions caused by the large number of electromagnetic design variables and complex constraints in existing transformer electromagnetic design methods by introducing machine learning-guided particle swarm optimization algorithm.

[0006] To achieve the above objectives, this invention provides a machine learning-driven optimization solution for the electromagnetic design of complex transformers, comprising: Receive a transformer electromagnetic design calculation request submitted by a user. The calculation request includes a configuration file and a calculation template corresponding to the transformer type. Parse the configuration file and calculation template to extract the task characteristic parameters, electromagnetic design variables to be optimized, engineering constraints and optimization objectives of the current design task; The task feature parameters are input into a pre-trained machine learning model, and the estimated initial values ​​of each electromagnetic design variable are output using implicit engineering experience learned from historical transformer design data. Based on the preset engineering experience rules corresponding to each electromagnetic design variable, the shrinking search domain of each electromagnetic design variable is generated according to the estimated initial value, so as to form the variable boundary constraints of the particle swarm optimization algorithm. Based on the estimated initial values ​​and the search domain, the particle swarm optimization algorithm constructs an initial population and sampling distribution, so that the optimization solution process prioritizes searching in the high feasibility candidate region, and introduces a constraint feasibility priority criterion and a backtracking mechanism on the basis of the particle swarm optimization algorithm for iterative optimization solution; When the preset termination conditions are met, output a transformer electromagnetic design scheme that meets the engineering constraints.

[0007] Furthermore, the step of inputting the task feature parameters into a pre-trained machine learning model and outputting estimated initial values ​​for each electromagnetic design variable using implicit engineering experience learned from historical transformer design data includes: constructing a set of task feature parameters characterizing design requirements from the extracted electrical parameters, operating conditions, structural presets, and performance targets; wherein the set of task feature parameters includes at least: the transformer's rated capacity, voltage level, frequency, cooling method, temperature rise limit, winding type, core material type, insulation class, and performance target values ​​or upper limits for no-load loss and load loss; and calling the offline preprocessing configuration file corresponding to the transformer type to numerically reduce the set of task feature parameters. The system performs classification and encoding to generate a normalized feature vector that meets the input requirements of a pre-trained machine learning model. This normalized feature vector is then input into a machine learning model pre-trained for the transformer type. The model performs inference in a multi-input multi-output (MIMO) manner, outputting continuous numerical predictions of each electromagnetic design variable in parallel. The pre-trained machine learning model is a decision tree ensemble model based on a gradient boosting framework, trained offline using historical data in a multi-input multi-output manner. Post-processing is then applied to the continuous numerical predictions. For the discrete design variables, a discretization mapping is performed based on a predefined set of selectable engineering values ​​to correct the predictions to the closest valid engineering values, resulting in the final estimated initial values.

[0008] Furthermore, before the step of inputting the task feature parameters into the pre-trained machine learning model, the method further includes: matching the task feature parameters of the current design task with design schemes in the historical database; if a successful historical design scheme that is the same as or highly similar to the current task is matched, the prediction process of the machine learning model is skipped, and the variable values ​​of the successful historical scheme are directly output as the estimated initial values; if a failed historical design scheme that is the same as or highly similar to the current task but does not meet the constraints is matched, the prediction process of the machine learning model is skipped, and the variable values ​​of the failed historical scheme are directly output as the estimated initial values; if no historical design scheme that meets the above conditions is matched, the step of inputting the task feature parameters into the pre-trained machine learning model is executed.

[0009] Furthermore, the step of generating a shrinkage search domain for each electromagnetic design variable based on the preset engineering experience rules corresponding to each electromagnetic design variable and according to the estimated initial value includes: determining the source of the estimated initial value; calling a preset shrinkage strategy corresponding to the source from the configuration file according to the determined source; and for each electromagnetic design variable, applying the called shrinkage strategy according to its variable type, and calculating its shrinkage search domain based on the estimated initial value of the variable.

[0010] Further, the step of determining the source of the estimated initial value and calling the preset shrinkage strategy corresponding to the source from the configuration file according to the determined source includes: if the source of the estimated initial value is obtained from historical feasible design schemes matched from the historical database, the resulting shrinkage search domain range is the smallest; if the source of the estimated initial value is generated by the machine learning model based on the current task feature parameters, the resulting shrinkage search domain range is the largest; if the source of the estimated initial value is obtained from historical failed design schemes matched from the historical database, the resulting shrinkage search domain range is between the largest and the smallest.

[0011] Furthermore, the step of applying the called contraction strategy to calculate the contraction search domain for each electromagnetic design variable, based on its variable type and using the estimated initial value of the variable as a reference, includes: for continuous variables, the contraction strategy defines a numerical contraction parameter for it; based on the contraction parameter, a temporary search interval is calculated with the estimated initial value as the center, and the boundary of the interval is restricted to the global engineering allowable range of the variable to form a continuous contraction search domain; for discrete variables, the contraction strategy defines a neighborhood size parameter N for it; using the estimated initial value in the predefined set of engineering selectable values ​​as the center point, N adjacent legal values ​​forward and backward from the center point are selected to jointly constitute a discrete contraction search domain.

[0012] Further, the step of constructing the initial population and sampling distribution based on the estimated initial values ​​and the search domain-guided particle swarm optimization algorithm includes: receiving the estimated initial values ​​of each electromagnetic design variable and its corresponding shrinking search domain; determining the position of at least one reference particle in the initial population based on the estimated initial values; randomly sampling the initial positions of the remaining particles within the variable boundary constraints defined by the shrinking search domain; and forming the initial population together with the reference particle and the randomly sampled particles.

[0013] Furthermore, the step of introducing a constraint feasibility priority criterion and a backtracking mechanism for iterative optimization based on the particle swarm optimization algorithm includes: the initial value is the initial population constructed based on the estimated initial value and the shrinking search domain; the iterative optimization loop executes the following in each iteration: (1) performing performance calculation and constraint verification on the electromagnetic design scheme represented by each particle position in the population; (2) updating the individual historical best position of each particle and the global historical best position of the population according to the constraint feasibility priority criterion; (3) calculating the new position of each particle according to the updated individual historical best position and global historical best position, following the velocity and position update formula of the particle swarm optimization algorithm; and restricting the components of the new position within a preset search domain, wherein the search domain is a shrinking search domain generated based on machine learning prediction and historical experience. (4) Monitor the optimization status and trigger the backtracking mechanism when the preset stagnation condition is met; end the iteration when the preset termination condition is met; wherein the stagnation condition includes: the global historical optimal position of the population is an infeasible solution and its total constraint violation has not decreased in the first preset number of iterations, and / or the global historical optimal position of the population is a feasible solution and its objective function value has not been improved in the second preset number of iterations; the backtracking mechanism includes: in response to triggering the backtracking mechanism, randomly select a portion of particles from the current population according to a preset ratio, and directly set the position of the selected particles as the coordinates of the global historical optimal position of the current population, or, with the global historical optimal position of the current population as the center, randomly generate a new position for each selected particle in a preset perturbation neighborhood.

[0014] Furthermore, the step of performing performance calculations and constraint verification on the electromagnetic design scheme represented by each particle position in the population includes: decoding the particle position vector into a set of specific values ​​for electromagnetic design variables; inputting the specific values ​​into an electromagnetic calculation template corresponding to the transformer type, driving the electromagnetic calculation template to perform calculations to obtain performance indicators; judging the compliance of each performance indicator according to predefined engineering constraints, and calculating the objective function value, total number of constraint violations, and feasibility determination of the scheme; and / or, the step of updating the individual historical best position of each particle and the global historical best position of the population according to the constraint feasibility priority criterion includes: for each particle in the population, comparing its current position with its individual historical best position, and deciding whether to update the individual historical best position of the particle according to the constraint feasibility priority criterion; after completing the update of the individual historical best positions of all particles, comparing the individual historical best positions of all particles, and determining the global historical best position of the population according to the constraint feasibility priority criterion.

[0015] In other embodiments, a machine learning-driven transformer electromagnetic design optimization solution system is provided, capable of executing the optimization method described above. The system includes: a request receiving and parsing module, used to receive transformer electromagnetic design calculation requests submitted by users, and parse the configuration file and calculation template in the request to extract task feature parameters, electromagnetic design variables to be optimized, engineering constraints, and optimization objectives for the current design task; an initial value prediction module, used to receive the task feature parameters and input them into a pre-trained machine learning model to obtain estimated initial values ​​for each electromagnetic design variable output by the model; and a shrinking search domain generation module, used to generate a set of initial values ​​for each electromagnetic design variable based on the estimated initial values ​​and preset engineering experience rules corresponding to each variable. A shrinking search domain, which together constitute the variable boundary constraints of the particle swarm optimization algorithm; an online optimization solution module is used to perform the following operations: using the estimated initial values ​​as a reference, initialize the particle swarm under the boundary constraints of the shrinking search domain; and perform iterative solution using an improved particle swarm optimization algorithm that integrates the constraint feasibility priority criterion and the backtracking mechanism, wherein the constraint feasibility priority criterion is used to guide the updating of the individual historical best position of the particles and the global historical best position of the population in each iteration, and the backtracking mechanism is used to reset some particles when the search stall is detected; a result output module is used to output the set of transformer electromagnetic design schemes that meet all engineering constraints obtained by the online optimization solution module when the preset termination conditions are met.

[0016] Based on the foregoing description, those skilled in the art will understand that this invention trains a model using historical data to learn design patterns, predicts initial values ​​for variables in new tasks, and generates a shrinking search domain using empirical rules, providing a high-quality starting point and accurate boundaries for subsequent optimization. In the solution phase, an improved particle swarm optimization algorithm is employed, initializing the population within the shrinking domain centered on the predicted values. A constraint feasibility priority criterion is introduced to guide the search direction, supplemented by a backtracking mechanism to avoid getting trapped in local optima, thereby focusing the search on high-potential regions. Through iterative electromagnetic calculations and constraint verification, an optimized design scheme that meets all engineering requirements is efficiently output. This solves the problems of low solution efficiency and difficulty in obtaining feasible solutions caused by numerous transformer electromagnetic design variables and complex constraints in existing transformer electromagnetic design methods, significantly improving design efficiency and scheme feasibility while reducing computational costs. Attached Figure Description

[0017] The accompanying drawings, as part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation of the invention. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart of a machine learning-driven optimization solution method for electromagnetic design of complex transformers in some embodiments of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0019] Those skilled in the art should understand that the embodiments described below are merely a part of the embodiments of the present invention, and not all of the embodiments of the present invention. These partial embodiments are intended to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of the present invention.

[0020] The following reference Figure 1 This paper will provide a detailed explanation of the machine learning-driven optimization solution method for electromagnetic design of complex transformers in some embodiments of the present invention. Figure 1 This is a flowchart of a machine learning-driven optimization solution method for electromagnetic design of complex transformers in some embodiments of the present invention.

[0021] In some embodiments of this invention, a machine learning-driven method for optimizing the electromagnetic design of complex transformers is provided. The system receives user-defined design tasks (i.e., configuration files and calculation templates) and parses them into structured data. Then, based on the priority reuse of historical samples, a pre-trained machine learning model is introduced to extract implicit engineering experience from historical transformer design data, and outputs estimated initial values ​​for each electromagnetic design variable when reusable historical samples are lacking. A search domain is then generated by combining preset engineering experience rules, thereby providing initialization guidance for the optimization solution. Finally, a particle swarm optimization algorithm is used to construct an initial population and sampling distribution based on the initial values ​​and the search domain, and a constraint feasibility priority criterion and a backtracking mechanism are introduced for iterative optimization, ultimately outputting a set consisting of multiple feasible design schemes. This method effectively addresses the problems in traditional transformer electromagnetic design, where the high dimensionality of design variables and the complexity and coupling of engineering constraints lead to low efficiency, difficulty in obtaining feasible solutions, and high computational costs, making it difficult to meet real-time engineering requirements. Machine learning-driven optimization methods for complex transformer electromagnetic design suffer from these issues. This method can shorten the time to obtain feasible solutions and even high-quality solutions, while ensuring that the optimization results meet the hard engineering constraints, thus improving the automation and engineering feasibility of complex transformer electromagnetic design. Furthermore, this method deeply integrates human experience, historical data, and algorithms, automating the process from task reception to solution output, reducing reliance on experienced designers.

[0022] like Figure 1 As shown, a machine learning-driven optimization solution method for the electromagnetic design of complex transformers includes: Step S100: Receive a transformer electromagnetic design calculation request submitted by the user. The request includes a configuration file and a calculation template corresponding to the transformer type.

[0023] Specifically, the system receives optimization task requests submitted by clients. These requests include: a user-edited configuration file, a calculation template identifier corresponding to the selected transformer type, and task metadata. The configuration file and the calculation template together describe the transformer electromagnetic design optimization task.

[0024] The configuration file is an editable structured file (written in a structured language such as JSON or YAML) used to define the configurable parts of the task, including: basic parameters, parameters to be retrieved, definitions and initial ranges of electromagnetic design variables, engineering constraints, optimization objectives, and optimization algorithm parameters. Therefore, the configuration file includes at least one of the following definition information: parameter definitions, variable definitions, constraint definitions, and optimization control definitions. Parameter definitions include basic parameters directly given by the user, and parameters to be retrieved from the built-in database of the calculation template. Variable definitions include the definition of the legal value set for discrete variables, and the definition of the value range and precision for continuous variables; constraint definitions describe the engineering performance index limits that the design scheme must meet. Optimization control definitions include the parameter configuration and calculation termination conditions for the particle swarm optimization algorithm.

[0025] Specifically, the basic parameters are designed by users directly and without relying on database queries. These basic parameters include identification and classification parameters, core electrical parameters, structural and material selection parameters, and environmental and special requirements parameters.

[0026] Identification and classification parameters are used for task identification, tracking, and classification. These parameters must include at least the customer code, product line, production code I, and production code II.

[0027] The core electrical parameters define the basic electrical specifications, winding configuration, and voltage regulation method of the transformer. These parameters include at least the capacity, high-voltage voltage, low-voltage voltage, third coil voltage, frequency, high-voltage connection type, low-voltage connection type, number of low-voltage groups, number of high-voltage sections, low-voltage operating mode, whether the third low-voltage section is reversed, high-voltage regulation type, and tap changer position number. The tap changer position numbers include tap 1 (negative), tap 2 (negative), tap 4 (positive), and tap 5 (positive), with tap X indicating the specific tap number. For example, taps 4 and 5 are used to increase the output voltage, while taps 1 and 2 are used to decrease the output voltage.

[0028] The structural and material selection parameters define the transformer's physical structure, cooling scheme, main materials, and insulation process details. These parameters include at least the structural type, cooling method, casing type, clamp type, conductor type, core type, wrapping method, insulation distance wrapping method between the upper / lower ends of the secondary coil, insulation distance wrapping method between the primary / secondary / tertiary / large secondary coil windings and auxiliary windings, high-voltage end distance adjustment 1 (lower left), high-voltage end distance adjustment 2 (upper right), and the air gap between the insulation casing and the core for delta-connected structures.

[0029] The environmental and special requirements parameters define the operating environment conditions and certain material or performance preset values. These parameters include at least altitude, three-proof requirements, thickness without a sample, thickness with a sample, ratio without a sample, magnetic flux density with a sample, and magnetic flux density without a sample.

[0030] The parameters to be retrieved are not directly input by the user, but are automatically queried, calculated, or matched from the engineering database built into the calculation template based on the input basic parameters. The parameters to be retrieved include performance constraint-related parameters, insulation and safety-related parameters, material and process-related parameters, and economic calculation-related parameters.

[0031] Performance constraint parameters include high-voltage temperature rise limits, low-voltage temperature rise limits, minimum efficiency requirements, impedance requirements, lower / upper limits for low-voltage voltage deviation, lower / upper limits for low-voltage phase shift angle deviation, lower / upper limits for third coil voltage deviation, reference temperature, and efficiency algorithm. The limits and algorithms for these performance constraint parameters are typically specified by national standards, customer agreements, or internal company control standards, and are derived from basic parameters such as capacity, voltage level, and cooling method.

[0032] Insulation and safety-related parameters include high-voltage withstand voltage, low-voltage withstand voltage, withstand voltage between low-voltage groups 12 / 23, withstand voltage of low-voltage subgroups, withstand voltage of the third coil, maximum center distance, maximum value between high-voltage coils, minimum value of high-voltage external air duct, minimum insulation distance of pad blocks, and calculated maximum value between high-voltage coils. These parameters are determined by insulation levels, safety regulations, and structural dimensions, and must be determined by consulting standards or calculating using empirical formulas based on voltage levels and insulation schemes.

[0033] Material and process-related parameters include heat dissipation structure, presence or absence of a bottom fan, high-voltage paper wrapping type, low-voltage basic paper wrapping type, and low-voltage phase-shifting paper wrapping type. These parameters are determined by matching them from the process library based on the cooling method, voltage level, and capacity.

[0034] The relevant parameters for economic calculations include the unit price of copper, the unit price of aluminum, the unit price of silicon steel sheets (with and without silicon steel sheets), and the capacity of the third coil. The unit price of materials is determined by market conditions or purchase contracts, while the capacity of the third coil may be determined by design standards or the voltage and load conditions of the third coil; all of these parameters need to be retrieved from the database to calculate the cost.

[0035] The definition and initial range of electromagnetic design variables define all variables to be optimized and set an initial search space for each variable. The definition and initial range of electromagnetic design variables include discrete variables and continuous variables.

[0036] Discrete variables list the set of legal values ​​for each variable. Discrete variables include at least the thickness of the third coil wire, the width of the third coil wire, the number of coil strands in the third coil, the number of low-voltage coils per group, the thickness of the high-voltage wire, the width of the high-voltage wire, the number of high-voltage strands, the thickness of the low-voltage basic wire, the width of the low-voltage basic wire, the number of low-voltage basic strands, the thickness of the low-voltage phase-shifting wire, the number of low-voltage phase-shifting strands, the first group angle, the adjustment high margin low-voltage, and the adjustment window high margin high-voltage.

[0037] Continuous variables define the initial minimum, maximum, and optimization step size (accuracy) for each variable. Continuous variables include at least the core diameter, high-pressure small air passages, number of high-pressure small air passages, inner main air passage adjustment value, outer main air passage adjustment value, number of high-pressure turns per cake, air passages between the first and second major sections, air passages within the first and second major sections, air passages between the third major sections, and air passages within the third major section.

[0038] Engineering constraints define all conditions that the design scheme must meet. These constraints link the limits in the pulled parameters with the optimization variables through calculation formulas in the calculation template. Engineering constraints include performance constraints, insulation and structural constraints, and process and physical feasibility constraints.

[0039] Performance constraints directly affect the transformer's operating performance and energy efficiency, and must meet national, industry standards, or specific customer requirements. Performance constraints include at least: temperature rise constraints, such as the calculated temperature rise of the high-voltage and low-voltage windings under rated load must be less than or equal to their corresponding allowable temperature rise limits; loss constraints, such as the total losses of the transformer (including no-load and load losses) must not exceed the maximum loss limit specified for the corresponding energy efficiency class; impedance constraints, such as the calculated short-circuit impedance value must be controlled within the reasonable tolerance range required by the customer or standard, typically allowing a deviation of ±10%; and voltage regulation constraints, such as the output voltage deviation under specified load conditions must be between the allowable lower and upper limits to ensure power supply quality.

[0040] Insulation and structural constraints are used to ensure the insulation reliability and mechanical structure rationality of transformers, which are the foundation for the long-term safe and stable operation of the equipment. Insulation and structural constraints include at least: internal insulation distance constraints, such as: the insulation distance between high-voltage winding discs must not exceed the maximum allowable value, and / or the distance from the outside of the high-voltage winding to the tank or other components must not be less than the minimum required value, and / or the distance of the main insulation parts must be greater than or equal to the minimum safe distance calculated according to the insulation withstand voltage level; structural dimension constraints, such as the total height of all windings after assembly must be less than the height of the transformer core window and allowances for necessary process adjustments.

[0041] Process and physical feasibility constraints anchor the design scheme within realistic manufacturability and physical laws. These constraints include at least: material specification standardization constraints, where the core diameter selected in the design must be one of the standard series values ​​conforming to production processes and material specifications, and cannot be arbitrarily chosen; and conductor current-carrying capacity constraints, such as the required conductor cross-sectional area calculated based on winding current, allowable current density, temperature rise, etc., which must be less than or equal to the actual selected conductor cross-sectional area to ensure the conductor does not overheat.

[0042] Optimization objectives are used to define one or more objective functions that need to be minimized (or maximized), primarily cost minimization. Optimization objectives include primary and secondary objectives. Primary objectives include at least total manufacturing cost, for example: Total manufacturing cost = silicon steel sheet cost + wire cost + insulation material cost + ...; silicon steel sheet cost = core weight * (price per silicon steel sheet with selection * proportion with selection + price per silicon steel sheet without selection * (1 - proportion with selection)); wire cost = (weight of high-voltage wire + weight of low-voltage wire + weight of third coil wire) * price per unit of copper (assuming copper wire). Secondary objectives include at least total loss or efficiency.

[0043] The optimization algorithm parameter configuration specifies the control parameters and stopping conditions used to solve the improved particle swarm optimization algorithm. The optimization algorithm parameters include at least population parameters, learning parameters, stopping conditions, and constraint handling parameters. Population parameters include, for example, population size and maximum number of iterations. Learning parameters include at least inertia weights, individual learning factors, and swarm learning factors. Stopping conditions include at least maximum computation time, maximum number of iterations without improvement, and the expected number of feasible solutions. Constraint handling parameters define the penalty weights for different constraints when the total number of constraints is violated.

[0044] The calculation sheet template is a spreadsheet file corresponding to the transformer type, containing predefined engineering calculation logic and verification formulas. It serves as the fixed calculation kernel and variable mapping relationship for the task. All information defined in the configuration file has a corresponding mapping location in the calculation sheet template. Specifically, the calculation sheet template is a spreadsheet file containing macro functions, which includes: a complete chain of electromagnetic, thermal, and mechanical property calculation formulas; a built-in database or data query link for storing or retrieving material properties and standard specification data; and cell mapping relationships corresponding to items defined in the configuration file.

[0045] The calculation template identifier is used to instruct the system to load the corresponding engineering calculation template file. The calculation template is a spreadsheet file stored on the server, containing predefined engineering calculation formulas and verification logic. The calculation template identifier corresponding to the transformer type is automatically generated by the system based on the transformer type selected by the user. Users can select or edit the corresponding configuration file and calculation template according to the transformer type and upload them, thereby achieving flexible configuration of electromagnetic design calculation tasks. Although the configuration files and calculation templates corresponding to different transformer types differ in parameter fields, design variable sets, and engineering constraint forms, they all follow a unified data parsing and optimization solution process.

[0046] After receiving the task, the system first verifies the fields and cell addresses in the configuration file, and then constructs a mapping relationship between the task parameters and the calculation template based on the parsing results. Subsequently, it loads the calculation template and triggers macro initialization and recalculation. The system writes the task parameters into the template according to the mapping relationship, retrieves the corresponding parameters from the template's built-in database, and performs structured parsing and standardized transformation on various constraints, calculation formulas, and discrete variable selections within the template itself. This results in an executable set of parameters, variables, constraints, and objectives, thus automatically instantiating the editable configuration + macro calculation template into an electromagnetic design optimization task that can be called by the optimization solver.

[0047] Specifically, the method includes the step of user editing and generating a configuration file, including: in response to the user selecting a transformer type on the client interface, the system provides a configuration file template corresponding to that type; receiving parameters input by the user on the client interface, the parameters including at least: basic electrical parameters, operating conditions, performance constraints and optimization control parameters; and filling or updating the configuration file template with the parameters input by the user to generate a structured configuration file.

[0048] Understandably, after receiving the user's request for transformer electromagnetic design calculations, the process also includes: performing security verification and parameter pre-check on the received request; after the verification passes, assigning a unique task identifier to the request and asynchronously placing the request into the task queue to await processing; and returning the task identifier to the client so that the user can query the task status and results.

[0049] Step S200 involves parsing the configuration file and calculation template to extract the task characteristic parameters, electromagnetic design variables to be optimized, engineering constraints, and optimization objectives for the current design task. Specifically, step S200 includes: Step S211: Parse the configuration file to obtain the basic electrical parameters of the transformer, define electromagnetic design variables, engineering constraints, optimization algorithm parameters and variable range information.

[0050] Step S212: Establish the mapping relationship between the parameters in the configuration file and the corresponding cells in the calculation template.

[0051] Step S213: Based on the mapping relationship, the parsed parameters are written into the calculation template, and the template is driven to pull the associated parameters from its built-in database.

[0052] In some specific embodiments, the step of writing basic parameters into the calculation template according to the mapping relationship includes: for each basic parameter defined in the configuration file, writing its value into the cell according to its mapping relationship with the cell address in the calculation template; after writing is completed, triggering the recalculation function of the calculation template so that all associated formulas based on the new parameter value are updated.

[0053] In some specific embodiments, the step of driving the calculation template to perform calculations or queries to obtain pulled parameters includes: the calculation template automatically retrieves and calculates the corresponding pulled parameter values ​​based on the written basic parameters through its built-in database query function or macro logic; the system reads these pulled parameter values ​​from the specified output cell of the calculation template to supplement all the input parameters required for the complete optimization task.

[0054] Step S214: Extract and structure the engineering calculation logic, variable relationships and constraint judgment conditions defined in the calculation template to form a set of parameters, variables, constraints and targets that can be called by the optimization algorithm, i.e., the optimization target.

[0055] The parameter set contains input parameters that remain unchanged during the optimization process, parsed from the basic electrical parameters in the configuration file; the variable set defines all electromagnetic design variables to be optimized and their basic properties, parsed from the electromagnetic design variables defined in the configuration file; the constraint set defines all engineering constraints in a structured manner, parsed from the engineering constraints in the configuration file; and the objective set defines the optimization objective function, parsed from the optimization objective, optimization algorithm parameters, and variable range information in the configuration file.

[0056] Step S300 involves inputting the task feature parameters into a pre-trained machine learning model and outputting estimated initial values ​​for each electromagnetic design variable using implicit engineering experience learned from historical transformer design data. Specifically, step S300 includes: Step S310: Construct a set of task feature parameters that characterize the design requirements from the extracted task feature parameters, including electrical parameters, operating conditions, structural presets, and performance targets.

[0057] The set of task characteristic parameters includes at least the following: the transformer's rated capacity, voltage level, frequency, cooling method, temperature rise limit, winding type, core material type, insulation class, and performance target or upper limit values ​​for no-load loss and load loss.

[0058] Step S320: Call the offline preprocessing configuration file corresponding to the transformer type, perform numerical scaling and classification encoding on the task feature parameter set, and generate a normalized feature vector that meets the input requirements of the pre-trained machine learning model. Step S330: The normalized feature vector is input into a machine learning model pre-trained for the transformer type. The model performs inference in a multi-input multi-output manner and outputs continuous numerical predictions of each electromagnetic design variable in parallel. The pre-trained machine learning model is a decision tree ensemble model based on a gradient boosting framework, specifically an XGBoost model. The model is trained offline using historical data in a multi-input multi-output manner and is able to learn and characterize the collaborative change relationship between different electromagnetic design variables.

[0059] Step S340: Post-process the continuous numerical prediction values. For the discrete design variables, perform discretization mapping according to the predefined set of selectable engineering values, and correct the prediction values ​​to the closest legal engineering values ​​to obtain the final estimated initial values.

[0060] Specifically, for each discrete electromagnetic design variable, the difference between the continuous predicted value output by the model and each legal value in the set of selectable values ​​for the project is calculated, and the legal value with the smallest difference is selected as the estimated initial value of the variable.

[0061] In some embodiments, before the step of inputting task feature parameters into a pre-trained machine learning model, the method further includes: matching the task feature parameters of the current design task with design schemes in a historical database; if a historical successful design scheme that is the same as or highly similar to the current task is matched, the prediction process of the machine learning model is skipped, and the variable values ​​of the historical successful scheme are directly output as the initial estimated values; if a historical failed design scheme that is the same as or highly similar to the current task but does not meet the constraints is matched, the prediction process of the machine learning model is skipped, and the variable values ​​of the historical failed scheme are directly output as the initial estimated values; if no historical design scheme that meets the above conditions is matched, the step of inputting task feature parameters into a pre-trained machine learning model is executed.

[0062] Step S400 involves generating a contracted search domain for each electromagnetic design variable based on preset engineering experience rules corresponding to each electromagnetic design variable, according to the estimated initial values, to form the variable boundary constraints for the particle swarm optimization algorithm. Specifically, step S400 includes: Step S410: Determine the source of the estimated initial value, and based on the determined source, call the preset shrinkage strategy corresponding to the source from the configuration file. Specifically, step S410 includes: Step S411: If the estimated initial values ​​are obtained from historical feasible design schemes matched in the historical database, the resulting shrinking search domain will be the smallest. Using the parameter set as a feature, a similarity search is performed in the historical database to locate referable historical design samples.

[0063] It is understood that the method of the present invention supports adaptive expansion for multiple types of transformers. By pre-setting independent configuration files, calculation templates, historical databases and pre-trained models for different types of transformers, and automatically loading matching resources according to the task type during subsequent optimization and solution, it can adapt to different transformer electromagnetic design tasks under the same computing framework.

[0064] Step S412: If the estimated initial value is generated by the machine learning model based on the current task feature parameters, then the resulting shrinkage search domain range is the largest.

[0065] Step S413: If the estimated initial value is obtained from historical failed design schemes matched from the historical database, the resulting shrinking search domain range is between the maximum and minimum. The located historical design sample can be determined as a successful or failed sample based on the constraint set.

[0066] Step S420: For each electromagnetic design variable, based on its variable type, apply the invoked shrinkage strategy and calculate its shrinkage search domain using the estimated initial value of the variable as a reference.

[0067] Specifically, for continuous variables, the contraction strategy defines a numerical contraction parameter, calculates a temporary search interval centered on the estimated initial value based on the contraction parameter, and restricts the boundary of the interval to the allowable range of the global engineering of the variable, forming a continuous contraction search domain.

[0068] The numerical shrinkage parameter can be set as a shrinkage ratio coefficient α or an absolute offset δ, and the corresponding temporary search interval is calculated as [initial value × (1-α), initial value × (1+α)] or [initial value - δ, initial value + δ], respectively. The shrinkage ratio coefficient α and the absolute offset δ can be set according to the operator's experience, and are not specifically limited here.

[0069] For discrete variables, the shrinkage strategy defines a neighborhood size parameter N; the estimated initial value is taken as the center point in the predefined set of selectable values ​​for the project, and N adjacent legal values ​​are selected from the center point forward and backward to form the discrete shrinkage search domain.

[0070] Step S500: Based on the estimated initial values ​​and search domain, the particle swarm optimization algorithm is used to construct the initial population and sampling distribution, so that the optimization solution process prioritizes the search in the high feasibility candidate region. On the basis of the particle swarm optimization algorithm, a constraint feasibility priority criterion and a backtracking mechanism are introduced to perform iterative optimization solution.

[0071] The steps for constructing the initial population and sampling distribution based on estimated initial values ​​and the search domain using the particle swarm optimization algorithm include: receiving estimated initial values ​​of each electromagnetic design variable and its corresponding shrinking search domain; determining the position of at least one reference particle in the initial population based on the estimated initial values; wherein, the step of determining the position of at least one reference particle based on the estimated initial values ​​involves: directly setting the estimated initial values ​​as the initial positions of one or more particles; randomly sampling the initial positions of the remaining particles within the variable boundary constraints defined by the shrinking search domain; specifically, for continuous electromagnetic design variables, random sampling is performed within the continuous interval [lower_bound, upper_bound] defined by the shrinking search domain, according to a uniform or normal distribution; for discrete electromagnetic design variables, random selection is performed from the set of discrete legal values ​​defined by the shrinking search domain. Finally, the reference particles and the randomly sampled particles together constitute the initial population.

[0072] The steps involved in iterative optimization based on the particle swarm optimization algorithm, which incorporate a constraint feasibility priority criterion and a backtracking mechanism, include: The initial value is an initial population constructed based on estimated initial values ​​and a contracted search domain; Iterative optimization loop, executing in each iteration: (1) For each particle position in the population, the performance of the electromagnetic design scheme is calculated and the constraints are checked. Specifically, the particle position vector is decoded into a set of specific values ​​of electromagnetic design variables; the specific values ​​are input into the electromagnetic calculation template corresponding to the transformer type, and the electromagnetic calculation template is driven to perform calculations to obtain performance indicators; according to the predefined engineering constraints, the compliance of each performance indicator is judged, and the objective function value, total number of constraints violated and feasibility of the scheme are calculated.

[0073] The performance indicators include at least one or more of the following: electromagnetic performance indicators, thermal performance indicators, and economic indicators; the electromagnetic performance indicators include no-load loss, load loss, and short-circuit impedance; the thermal performance indicators include the temperature rise of the winding hot spot and the temperature rise of the top oil; and the economic indicators include the estimated manufacturing cost calculated based on the material usage and unit price.

[0074] In some specific examples, the step of inputting specific values ​​into the electromagnetic calculation template corresponding to the transformer type includes: writing each variable value of the particle into the corresponding specified cell of the calculation template according to the initial value of the variable and the mapping relationship established in the search domain estimation module; after writing, driving the calculation template to perform a complete recalculation through the application programming interface (API) or script to trigger its built-in engineering calculation formulas, macros and database queries.

[0075] This step also includes quantifying and recording the results. Specifically, the steps of quantifying and recording the results include: calculating a scalar value representing the overall degree of violation of engineering constraints by the proposed solution; recording the objective function value corresponding to the proposed solution; and labeling the proposed solution as either a "feasible solution" or an "infeasible solution" based on whether the total constraint violation is 0. The total constraint violation, objective function value, and feasible solution label obtained in the evaluation step are used to update the individual historical optimal position and global optimal position of the guiding particle search in subsequent iterations according to the constraint feasibility priority criterion.

[0076] (2) Update the individual historical best position of each particle and the global historical best position of the population according to the constraint feasibility priority criterion. For each particle in the population, compare its current position with its individual historical best position, and decide whether to update the individual historical best position of the particle according to the constraint feasibility priority criterion; after completing the update of the individual historical best positions of all particles, compare the individual historical best positions of all particles, and determine the global historical best position of the population according to the constraint feasibility priority criterion.

[0077] (3) Based on the updated individual historical best position and global historical best position, calculate the new position of each particle according to the velocity and position update formula of the particle swarm optimization algorithm; and restrict the components of the new position to a preset search domain, which is a shrinking search domain generated based on machine learning prediction and historical experience.

[0078] Specifically, based on the dynamic equations of the standard particle swarm optimization algorithm, the new velocity of a particle is calculated according to its current position, velocity, its individual historical best position, and the global best position of the population. The new velocity is then added to the particle's current position to obtain its new position. The velocity update step can be calculated using the following formula. Speed ​​in the (k+1)th iteration: ; in, and They represent the first The particle in the first The position vector and velocity vector at the next iteration Represents particles The updated individual historical best position. This represents the updated global optimal position of the population. Inertial weights are used to balance global and local search capabilities. and As a learning factor, and To take the value of The random number is generated. The position update step calculates the particle's position according to the following formula. Position in the (k+1)th iteration: After calculating the new position of the particle, it is determined whether the new position exceeds the defined search domain set for each electromagnetic design variable. If any component of the new position exceeds its corresponding search domain, boundary processing is performed on that component to ensure that all variable values ​​are within their legal search domains. The boundary processing step can be set as absorption processing or random reset processing. Absorption processing can involve directly setting the value of the component exceeding the search domain to the value of the corresponding boundary. Random reset processing can involve randomly generating a legal value for the out-of-bounds component within the allowed search domain of the variable.

[0079] (4) Monitor the optimization status and trigger the backtracking mechanism when the preset stagnation conditions are met. The stagnation conditions include: the global historical best position of the population is an infeasible solution and the total number of constraint violations has not decreased in the first preset number of consecutive iterations, and / or the global historical best position of the population is a feasible solution and the objective function value has not improved in the second preset number of consecutive iterations. The backtracking mechanism includes: in response to triggering the backtracking mechanism, randomly selecting a portion of particles from the current population according to a preset ratio, and directly setting the position of the selected particles as the coordinates of the global historical best position of the current population, or randomly generating a new position for each selected particle within a preset perturbation neighborhood centered on the global historical best position of the current population.

[0080] In some specific embodiments, the step of randomly selecting particles according to a preset ratio includes: setting a backtracking ratio parameter ρ, where 0 < ρ ≤ 1; calculating the number of particles N to be backtracked, where N is equal to the current population size multiplied by the backtracking ratio parameter ρ and then rounded up or down; and randomly selecting N particles from the current entire population.

[0081] In some specific examples, for continuous design variables, the preset disturbance neighborhood is defined by a disturbance amplitude vector δ, and the new location component is calculated by the formula new value = global historical best value + δ * random number, where the random number is uniformly distributed in the interval [-1, 1]. For discrete design variables, the preset disturbance neighborhood is composed of a set defined by the M adjacent values ​​before and after the global historical best value in the set of selectable values ​​in the project, and the new location component is randomly selected from this set.

[0082] The iteration ends when the preset termination condition is met.

[0083] The preset termination conditions include at least one of the following: reaching the preset maximum number of iterations, reaching the user-defined computation time limit, or reaching a preset threshold for the cumulative number of feasible design solutions satisfying all engineering constraints. Specifically, when the cumulative number of feasible design solutions satisfying all engineering constraints reaches the preset threshold, the optimization process automatically terminates upon reaching any preset threshold, and all currently found feasible solutions or the optimal solution are output. Furthermore, the maximum number of iterations, the set computation time, and the preset threshold can be set by the operator based on actual needs or experience; no specific limitations are imposed here.

[0084] Step S600: When the preset termination condition is met, output the transformer electromagnetic design scheme that meets the engineering constraints.

[0085] The method of this invention supports adaptive expansion for multiple types of transformers, therefore the method further includes: step S700, an offline pre-training model step. Step S700 includes: Step S710 involves standardizing the historical transformer electromagnetic design data, including data cleaning, outlier handling, feature scaling and encoding, and stratified sampling of the training and test sets according to key transformer parameters, outputting a standardized feature set, label set, and preprocessing configuration file.

[0086] The data processing steps include: handling missing values, deduplicating duplicate samples, and unifying units and dimensions in the original historical data; identifying and processing numerical records with abnormal formats and those exceeding the reasonable range of the project; standardizing numerical fields and establishing a mapping table by uniformly encoding non-numerical fields; and dividing the preprocessed samples into training and test sets by using stratified sampling based on capacity or voltage levels.

[0087] Step S720: Using the normalized data, multiple machine learning models are trained in a multi-input multi-output (MIMO) format to learn the mapping relationship between design requirement parameters and electromagnetic design variables, which are represented by implicit engineering experience in historical transformer design data. In the model training step, the multiple machine learning models include at least a decision tree ensemble model (XGBoost, or Extreme Gradient Boosting) based on the gradient boosting framework (LightGBM) and a support vector regression model. The multi-input multi-output format means that multiple electromagnetic design variables are modeled as a joint estimation objective.

[0088] Step S730: Use the test set and error index to evaluate the various trained models, and select the model with the best evaluation result as the final pre-trained estimation model.

[0089] In the model evaluation and selection step, the error metrics used for evaluation include at least one of the following: mean absolute error, mean squared error, root mean squared error, mean absolute percentage error, coefficient of determination, and explained variance score. The final preferred model is a decision tree ensemble model.

[0090] In some specific embodiments, the method further includes step S800, a model update mechanism. Specifically, step S800 includes: continuously accumulating new engineering calculation results and corresponding actual design variable values ​​during the optimization solution process; retraining the pre-trained estimation model based on the newly added engineering data; evaluating and comparing the new training model with the original model using a coverage index, whereby the coverage index measures the ability of the model to estimate the initial values ​​and its corresponding search domain to cover the actual engineering result values; and replacing the original model with the new model when the coverage index of the new model is higher than that of the original model.

[0091] In other embodiments of the present invention, a machine learning-driven transformer electromagnetic design optimization solution system is provided, capable of executing the optimization method described in any of the above descriptions. The system includes a request receiving and parsing module, an initial value prediction module and a shrinking search domain generation module, an online optimization solution module (i.e., POS, short for Pre-trained Online Solver), and a result output module.

[0092] The request receiving and parsing module receives user-submitted transformer electromagnetic design calculation requests and parses the configuration file and calculation template in the request to extract the task feature parameters, electromagnetic design variables to be optimized, engineering constraints, and optimization objectives for the current design task. The initial value prediction module receives the task feature parameters and inputs them into a pre-trained machine learning model to obtain estimated initial values ​​for each electromagnetic design variable output by the model. The shrinking search domain generation module generates a shrinking search domain for each electromagnetic design variable based on the estimated initial values ​​and preset engineering experience rules corresponding to each variable. These shrinking search domains collectively constitute the variable boundary constraints of the particle swarm optimization algorithm. The online optimization solution module performs the following operations: using the estimated initial values ​​as a reference, and under the boundary constraints of the shrinking search domain, initializes the particle swarm and iteratively solves the problem using an improved particle swarm optimization algorithm that integrates constraint feasibility priority criteria and a backtracking mechanism. The constraint feasibility priority criteria guide the updating of the individual historical best position of the particles and the global historical best position of the population in each iteration, while the backtracking mechanism resets some particles when search stagnation is detected. The result output module is used to output a collection of transformer electromagnetic design schemes that meet all engineering constraints obtained by the online optimization solution module when the preset termination conditions are met.

[0093] In some specific embodiments, the shrinking search domain generation module is configured to: execute the sub-step of inputting the task feature parameters into a pre-trained machine learning model and outputting estimated initial values ​​of each electromagnetic design variable using implicit engineering experience learned from historical transformer design data. Specifically, this includes: constructing a set of task feature parameters characterizing design requirements from the extracted electrical parameters, operating conditions, structural presets, and performance targets of the task feature parameters; wherein the set of task feature parameters includes at least: the transformer's rated capacity, voltage level, frequency, cooling method, temperature rise limit, winding type, core material type, insulation class, and performance target values ​​or upper limits for no-load loss and load loss; calling the offline preprocessing configuration file corresponding to the transformer type to perform numerical scaling and classification encoding on the set of task feature parameters, generating a normalized feature vector that meets the input requirements of the pre-trained machine learning model; The normalized feature vector is input into a machine learning model pre-trained for the transformer type. The model performs inference in a multi-input multi-output (MIMO) manner, and outputs continuous numerical predictions of each electromagnetic design variable in parallel. The pre-trained machine learning model is a decision tree ensemble model based on a gradient boosting framework. The model is trained offline using historical data in a multi-input multi-output manner. The continuous numerical predictions are then post-processed. For the discrete design variables, a discretization mapping is performed based on a predefined set of selectable engineering values ​​to correct the predictions to the closest valid engineering values, thus obtaining the final estimated initial values.

[0094] In some specific embodiments, the shrinking search domain generation module is further configured to perform a step prior to the step of inputting task feature parameters into a pre-trained machine learning model. Specifically, this includes: matching the task feature parameters of the current design task with design schemes in a historical database; if a historical successful design scheme that is the same as or highly similar to the current task is matched, the prediction process of the machine learning model is skipped, and the variable values ​​of the historical successful scheme are directly output as the estimated initial values; if a historical failed design scheme that is the same as or highly similar to the current task but does not meet the constraints is matched, the prediction process of the machine learning model is skipped, and the variable values ​​of the historical failed scheme are directly output as the estimated initial values; if no historical design scheme that meets the above conditions is matched, the step of inputting task feature parameters into a pre-trained machine learning model is performed.

[0095] In some specific embodiments, the online optimization solution module is further configured to: execute the sub-step of generating a contracted search domain for each electromagnetic design variable based on the preset engineering experience rules corresponding to each electromagnetic design variable and according to the estimated initial values. Specifically, this includes: determining the source of the estimated initial values; calling a preset contraction strategy corresponding to the source from the configuration file according to the determined source; and for each electromagnetic design variable, applying the called contraction strategy according to its variable type, and calculating its contracted search domain based on the estimated initial values ​​of the variable.

[0096] In some specific embodiments, the online optimization solution module is further configured to: perform the sub-step of determining the source of the estimated initial value, and calling a preset shrinkage strategy corresponding to the source from the configuration file based on the determined source. Specifically, this includes: if the source of the estimated initial value is obtained from historical feasible design schemes matched from a historical database, the resulting shrinkage search domain range is the smallest; if the source of the estimated initial value is obtained from historical failed design schemes matched from a historical database, the resulting shrinkage search domain range is the largest; if the source of the estimated initial value is generated by the machine learning model based on the current task feature parameters, the resulting shrinkage search domain range is between the largest and the smallest.

[0097] In some specific embodiments, the online optimization solution module is further configured to: execute the sub-step of calculating the shrinkage search domain for each electromagnetic design variable, based on its variable type, by applying the invoked shrinkage strategy and using the estimated initial value of the variable as a reference. Specifically, this includes: for continuous variables, the shrinkage strategy defines a numerical shrinkage parameter for it; based on the shrinkage parameter, a temporary search interval is calculated centered on the estimated initial value, and the boundary of the interval is restricted to the global engineering allowable range of the variable to form a continuous shrinkage search domain; for discrete variables, the shrinkage strategy defines a neighborhood size parameter N for it; using the estimated initial value in a predefined set of engineering selectable values ​​as the center point, N adjacent legal values ​​forward and backward from the center point are selected to jointly constitute a discrete shrinkage search domain.

[0098] In some specific embodiments, the online optimization solution module is further configured to: execute the sub-step of constructing an initial population and sampling distribution based on the estimated initial values ​​and the search domain-guided particle swarm optimization algorithm. Specifically, this includes: receiving the estimated initial values ​​of each electromagnetic design variable and its corresponding shrinking search domain; determining the position of at least one reference particle in the initial population based on the estimated initial values; randomly sampling the initial positions of the remaining particles within the variable boundary constraints defined by the shrinking search domain; and forming the initial population together with the reference particle and the randomly sampled particles.

[0099] In some specific embodiments, the online optimization solution module is further configured to: execute the sub-steps of iterative optimization solution based on the particle swarm optimization algorithm by introducing a constraint feasibility priority criterion and a backtracking mechanism, including: initial value: initial population; iterative optimization loop, in each iteration, performing: (1) performance calculation and constraint verification of the electromagnetic design scheme represented by each particle position in the population; (2) updating the individual historical best position of each particle and the global historical best position of the population according to the constraint feasibility priority criterion; (3) calculating the new position of each particle according to the updated individual historical best position and global historical best position, according to the speed and position update formula of the particle swarm optimization algorithm; and restricting the components of the new position within a preset search domain, wherein the search domain is a shrinking search domain generated based on machine learning prediction and historical experience; (4) monitoring the optimization status, and triggering the backtracking mechanism when the preset stagnation condition is met. The iteration ends when the preset termination condition is met. The stagnation conditions include: the global historical optimal position of the population is an infeasible solution, and the total number of constraint violations has not decreased in a first preset number of iterations, and / or the global historical optimal position of the population is a feasible solution, and the objective function value has not improved in a second preset number of iterations; the backtracking mechanism includes: in response to triggering the backtracking mechanism, randomly selecting a portion of particles from the current population according to a preset ratio, and directly setting the position of the selected particles as the coordinates of the global historical optimal position of the current population, or randomly generating a new position for each selected particle within a preset perturbation neighborhood centered on the global historical optimal position of the current population.

[0100] In some specific embodiments, the online optimization solution module is further configured to: execute the sub-step of performing performance calculations and constraint verification for the electromagnetic design scheme represented by each particle position in the population. Specifically, this includes: decoding the particle position vector into specific values ​​of a set of electromagnetic design variables; inputting the specific values ​​into an electromagnetic calculation template corresponding to the transformer type, driving the electromagnetic calculation template to perform calculations to obtain performance indicators; judging the compliance of each performance indicator according to predefined engineering constraints, and calculating the objective function value, total number of constraint violations, and feasibility determination of the scheme; and / or, the step of updating the individual historical best position of each particle and the global historical best position of the population according to the constraint feasibility priority criterion includes: for each particle in the population, comparing its current position with its individual historical best position, and deciding whether to update the individual historical best position of the particle according to the constraint feasibility priority criterion; after completing the update of the individual historical best positions of all particles, comparing the individual historical best positions of all particles, and determining the global historical best position of the population according to the constraint feasibility priority criterion.

[0101] In other embodiments of the present invention, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the relevant steps of the machine learning-driven complex transformer electromagnetic design optimization solution method as described above and applies them to the machine learning-driven transformer electromagnetic design optimization solution system described above.

[0102] In other embodiments of the present invention, a computer-readable storage medium is also provided, which stores a computer program, characterized in that, when the computer program is executed by a processor, it implements the machine learning-driven optimization solution method for complex transformer electromagnetic design as described above and applies it to the machine learning-driven optimization solution system for transformer electromagnetic design described above.

[0103] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, estimation models, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0104] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0105] Those skilled in the art will understand that this invention receives and parses structured design tasks through a system. Based on the priority reuse of historical samples, it introduces a pre-trained machine learning model to extract implicit engineering experience from historical transformer design data. When reusable historical samples are lacking, it outputs estimated initial values ​​for each electromagnetic design variable, and then combines these with engineering experience rules to generate a corresponding search domain, thus providing initialization guidance for optimization. Then, an improved particle swarm optimization algorithm is employed to construct an initial population and sampling distribution based on the estimated initial values ​​and search domain, prioritizing the search within regions more likely to produce feasible solutions. It then combines constraint feasibility priority criteria and a backtracking mechanism for iterative optimization. This method effectively solves the problems of low solution efficiency, difficulty in obtaining feasible solutions, and insufficient automation in traditional optimization methods when dealing with high-dimensional, multi-constraint transformer designs. It improves design efficiency, engineering reliability, and automation levels, and possesses good versatility and scalability.

[0106] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0107] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the claims, any of the claimed embodiments can be used in any combination.

[0108] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. The implementation schemes in the above embodiments can be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A machine learning-driven method for optimizing the electromagnetic design of complex transformers, characterized in that, include: Receive a transformer electromagnetic design calculation request submitted by a user. The calculation request includes a configuration file and a calculation template corresponding to the transformer type. Parse the configuration file and calculation template to extract the task characteristic parameters, electromagnetic design variables to be optimized, engineering constraints and optimization objectives of the current design task; The task feature parameters are input into a pre-trained machine learning model, and the implicit engineering experience learned from historical transformer design data is used to output the estimated initial values ​​of each electromagnetic design variable. Based on the preset engineering experience rules corresponding to each electromagnetic design variable, the shrinking search domain of each electromagnetic design variable is generated according to the estimated initial value, so as to form the variable boundary constraints of the particle swarm optimization algorithm. Based on the estimated initial values ​​and the search domain, the particle swarm optimization algorithm constructs an initial population and sampling distribution, so that the optimization solution process prioritizes searching in the high feasibility candidate region, and introduces a constraint feasibility priority criterion and a backtracking mechanism on the basis of the particle swarm optimization algorithm for iterative optimization solution; When the preset termination conditions are met, output a transformer electromagnetic design scheme that meets the engineering constraints. Before the step of inputting the task feature parameters into the pre-trained machine learning model, the method further includes: matching the task feature parameters of the current design task with design schemes in the historical database; if a successful historical design scheme that is the same as or highly similar to the current task is matched, the prediction process of the machine learning model is skipped, and the variable values ​​of the successful historical scheme are directly output as the initial estimated values; if a failed historical design scheme that is the same as or highly similar to the current task but does not meet the constraints is matched, the prediction process of the machine learning model is skipped, and the variable values ​​of the failed historical scheme are directly output as the initial estimated values; if no historical design scheme that meets the above conditions is matched, the step of inputting the task feature parameters into the pre-trained machine learning model is executed. The step of generating a shrinkage search domain for each electromagnetic design variable based on the preset engineering experience rules corresponding to each electromagnetic design variable and according to the estimated initial value includes: determining the source of the estimated initial value; calling a preset shrinkage strategy corresponding to the source from the configuration file according to the determined source; and for each electromagnetic design variable, applying the called shrinkage strategy according to its variable type, and calculating its shrinkage search domain based on the estimated initial value of the variable. The step of determining the source of the estimated initial value and calling the preset shrinkage strategy corresponding to the source from the configuration file based on the determined source includes: if the source of the estimated initial value is obtained from historical feasible design schemes matched from a historical database, the resulting shrinkage search domain range is the smallest; if the source of the estimated initial value is generated by the machine learning model based on the current task feature parameters, the resulting shrinkage search domain range is the largest; if the source of the estimated initial value is obtained from historical failed design schemes matched from a historical database, the resulting shrinkage search domain range is between the largest and the smallest.

2. The method according to claim 1, characterized in that, The step of inputting the task feature parameters into a pre-trained machine learning model and using implicit engineering experience learned from historical transformer design data to output estimated initial values ​​for each electromagnetic design variable includes: The extracted task feature parameters, including electrical parameters, operating conditions, structural presets, and performance targets, are used to construct a set of task feature parameters that characterize the design requirements. The set of task feature parameters includes at least the following: the transformer's rated capacity, voltage level, frequency, cooling method, temperature rise limit, winding type, core material type, insulation class, and performance target values ​​or upper limits for no-load loss and load loss. The offline preprocessing configuration file corresponding to the transformer type is invoked to perform numerical scaling and classification encoding on the task feature parameter set, generating a normalized feature vector that meets the input requirements of the pre-trained machine learning model. The normalized feature vector is input into a machine learning model pre-trained for the transformer type. The model performs inference in a multi-input multi-output manner and outputs continuous numerical predictions of each electromagnetic design variable in parallel. The pre-trained machine learning model is a decision tree ensemble model based on a gradient boosting framework. The model is trained offline using historical data in a multi-input multi-output manner. The continuous numerical prediction values ​​are post-processed. For the discrete design variables, discretization mapping is performed according to a predefined set of selectable engineering values ​​to correct the prediction values ​​to the closest legal engineering values, thus obtaining the final estimated initial values.

3. The method according to claim 1, characterized in that, The step of calculating the shrinkage search domain for each electromagnetic design variable, based on its variable type and the invoked shrinkage strategy, using the estimated initial value of the variable as a reference, includes: For continuous variables, the contraction strategy defines a numerical contraction parameter for it. Based on the contraction parameter, a temporary search interval is calculated with the estimated initial value as the center, and the boundary of the interval is restricted to the global engineering allowable range of the variable to form a continuous contraction search domain. For discrete variables, the shrinkage strategy defines a neighborhood size parameter N; taking the estimated initial value as the center point in the predefined set of selectable values ​​for the project, the strategy selects N adjacent valid values ​​forward and backward from the center point to form the discrete shrinkage search domain.

4. The method according to claim 3, characterized in that, The steps of constructing the initial population and sampling distribution based on the estimated initial values ​​and the search domain-guided particle swarm optimization algorithm include: Receive the estimated initial values ​​of each electromagnetic design variable and its corresponding shrinkage search domain; Based on the estimated initial values, determine the position of at least one reference particle in the initial population; Within the variable boundary constraints defined by the contracted search domain, the initial positions of the remaining particles are randomly sampled; The initial population is composed of the reference particle and the randomly sampled particles.

5. The method according to claim 4, characterized in that, The steps of introducing a constraint feasibility priority criterion and a backtracking mechanism for iterative optimization based on the particle swarm optimization algorithm include: The initial value is the initial population constructed based on the estimated initial value and the contracted search domain; Iterative optimization loop, executing in each iteration: (1) Perform performance calculations and constraint checks on the electromagnetic design scheme represented by each particle position in the population; (2) Based on the constraint feasibility priority criterion, update the individual historical best position of each particle and the global historical best position of the population; (3) Based on the updated individual historical best position and global historical best position, calculate the new position of each particle according to the velocity and position update formula of the particle swarm optimization algorithm; and restrict the components of the new position within a preset search domain, which is a shrinking search domain generated based on machine learning prediction and historical experience; (4) Monitor the optimization status and trigger the backtracking mechanism when the preset stagnation conditions are met; The iteration ends when the preset termination condition is met.

6. The method according to claim 5, characterized in that... The stagnation conditions include: The global historical optimal position of the population is an infeasible solution, and the total number of constraint violations has not decreased in the first preset number of consecutive iterations, and / or the global historical optimal position of the population is a feasible solution, and the objective function value has not improved in the second preset number of consecutive iterations.

7. The method according to claim 5, characterized in that, The backtracking mechanism includes: In response to triggering the backtracking mechanism, a subset of particles are randomly selected from the current population according to a preset ratio. The position of the selected particles is directly set as the coordinates of the global historical best position of the current population, or a new position is randomly generated for each selected particle within a preset perturbation neighborhood centered on the global historical best position of the current population.

8. The method according to claim 5, characterized in that, The steps of performing performance calculations and constraint checks on the electromagnetic design scheme represented by each particle position in the population include: Decode the particle position vector into a set of specific values ​​for electromagnetic design variables; The specific values ​​are input into the electromagnetic calculation template corresponding to the transformer type, and the electromagnetic calculation template is driven to perform calculations to obtain performance indicators. Based on the predefined engineering constraints, the compliance of each performance index is determined, and the objective function value, total number of constraint violations, and feasibility determination of the proposed scheme are calculated.

9. The method according to claim 5, characterized in that, The step of updating the individual historical best position of each particle and the global historical best position of the population based on the constraint feasibility priority criterion includes: For each particle in the population, compare its current position with its individual historical best position, and decide whether to update the individual historical best position of the particle based on the constraint feasibility priority criterion. After updating the individual historical best positions of all particles, compare the individual historical best positions of all particles, and determine the global historical best position of the population based on the constraint feasibility priority criterion.

10. A machine learning-driven transformer electromagnetic design optimization solution system, characterized in that, The system is capable of performing the optimization method according to any one of claims 1 to 9, the system comprising: The request receiving and parsing module is used to receive the transformer electromagnetic design calculation request submitted by the user, and parse the configuration file and calculation sheet template in the request to extract the task characteristic parameters, electromagnetic design variables to be optimized, engineering constraints and optimization objectives of the current design task. The initial value prediction module is used to receive the task feature parameters and input them into a pre-trained machine learning model to obtain the estimated initial values ​​of each electromagnetic design variable output by the model. The shrinking search domain generation module is used to generate a shrinking search domain for each electromagnetic design variable based on the estimated initial values ​​and the preset engineering experience rules corresponding to each variable. The shrinking search domains together constitute the variable boundary constraints of the particle swarm optimization algorithm. The online optimization and solution module is used to perform the following operations: using the estimated initial values ​​as a reference, initialize the particle swarm under the boundary constraints of the shrinking search domain; and perform iterative solution using an improved particle swarm optimization algorithm that integrates the constraint feasibility priority criterion and the backtracking mechanism. The constraint feasibility priority criterion is used to guide the updating of the individual historical best position of the particles and the global historical best position of the population in each iteration, and the backtracking mechanism is used to reset some particles when the search stalls are detected. The result output module is used to output a collection of transformer electromagnetic design schemes that meet all engineering constraints obtained by the online optimization solution module when the preset termination conditions are met.

Images