Solution method for machine learning models of mixed integer linear programming of industrial optimization

Abstract

This invention discloses a method for solving a machine learning model of mixed-integer linear programming in industrial optimization scenarios, belonging to the field of solving mixed-integer linear programming in industrial optimization scenarios. The method includes: Step 1, collecting and preprocessing multi-source data to generate a training set; Constraint learning model training: dynamically selecting the best-matching machine learning model according to the needs of the industrial scenario; Step 3, Automatic model transformation processing: automatically identifying and parsing the structure of the best-matching machine learning model, obtaining linearized constraints based on SOS1 constraints, and outputting the mixed-integer linear programming constraint set, objective function, and set of binary auxiliary variables; Step 4, Optimization solution: solving using an open-source solver to obtain the corresponding optimal decision variable values; Step 5, Solution result output: parsing the decision variable values ​​obtained in Step 4 into directly executable industrial optimization solutions and outputting them. This invention can achieve general automated modeling for mixed-integer linear programming, ensuring solution efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of mixed-integer linear programming solution technology for industrial optimization scenarios, and more particularly to a machine learning model solution method for mixed-integer linear programming in industrial optimization. Background Technology

[0002] Mixed-Integer Linear Programming (MILP) plays an important role in industrial optimization scenarios (such as production scheduling, energy dispatching, and path planning), but its integration with machine learning models faces significant challenges.

[0003] Current technologies primarily transform machine learning models (such as neural networks and decision trees) into mixed-integer linear programming constraints through manual derivation or limited tools before solving them. However, this approach suffers from several bottlenecks: First, manual modeling requires significant human intervention, leading to inefficiency and the introduction of errors that cause optimization results to deviate from actual requirements. Second, existing automated tools have limited compatibility, supporting only simple models and specific frameworks, making it difficult to adapt to complex structures or cross-platform ecosystems, thus increasing deployment costs. Furthermore, redundant constraints generated by traditional linear modeling methods significantly increase solution complexity, reducing efficiency and stability in industrial scenarios. These issues hinder the large-scale application of data-driven optimization techniques in practical industrial optimization.

[0004] Therefore, the problem to be solved is how to provide a more efficient, robust, and compatible solution method for machine learning models of mixed-integer linear programming for industrial optimization.

[0005] In view of this, the present invention is hereby proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a solution method for machine learning models of mixed-integer linear programming in industrial optimization scenarios. This method can achieve more efficient, robust, and compatible solutions for machine learning models of mixed-integer linear programming in industrial optimization scenarios, improve the solution efficiency and stability of industrial optimization scenarios, and thus solve the aforementioned technical problems existing in the prior art.

[0007] The objective of this invention is achieved through the following technical solution: A method for solving a machine learning model of mixed-integer linear programming in industrial optimization is provided, which is used to solve the machine learning model of mixed-integer linear programming corresponding to the industrial optimization problem, including: Step 1, Industrial Data Acquisition and Preprocessing: We acquire production-related raw industrial data from multiple sources from various industrial application scenarios, and process the acquired raw industrial data to generate a training dataset that can be directly input into a machine learning model. Step 2, Constraint Learning Model Training: The data in the training dataset generated in step 1 is divided into training set, validation set and test set according to a predetermined ratio, so that data from the same production batch or time period are split into the same set. Based on the implicit coupling and complex system relationships that are difficult to analyze in the industrial scenarios corresponding to the industrial problem to be optimized, at least one matching machine learning model is selected from the candidate model pool with multiple machine learning models of different structures. The selected machine learning models are trained and evaluated through training set, validation set and test set. Based on the evaluation results, the machine learning model that best matches the needs of the industrial scenario corresponding to the industrial problem to be optimized is selected. Step 3, Automatic Model Conversion Processing: The machine learning model structure selected in step 2 is automatically identified and parsed as the parsing result, and the parsing result is unified into an intermediate representation. Each primitive operation in the intermediate representation is mapped to a linear form or piecewise linear form that can be accepted by mixed integer linear programming. Based on the SOS1 constraint, the obtained linear form or piecewise linear form is piecewise linearized to obtain linearized constraints. All the obtained linearized constraints are packaged and output as a standard mixed integer linear programming constraint set and objective function, and a set of binary auxiliary variables corresponding to the objective function is generated. Step 4, optimize the solution: The mixed integer programming solver is used to solve the mixed integer linear programming constraint set, objective function and binary auxiliary variable set formed in step 3 to obtain the corresponding optimal decision variable values; Step 5, output the solution results: The optimal decision variable values ​​obtained in step 4 are analyzed into optimization schemes that can be directly executed and output for the industrial problem to be optimized.

[0008] Compared with existing technologies, the solution method for the machine learning model of mixed-integer linear programming for industrial optimization provided by this invention has the following advantages: By employing end-to-end automatic conversion processing applicable to machine learning models across different frameworks, various machine learning models are automatically identified, parsed, and mapped into linear constraints or piecewise linear constraints that can be embedded in mixed-integer linear programming. This significantly simplifies the model conversion process and reduces manual derivation and debugging. Simultaneously, the use of robust modeling techniques based on SOS1 constraints eliminates the dependence on extremely large constants, resulting in better numerical stability and faster convergence speed for the solver during branch and bound processes. This invention can rapidly and automatically embed and solve large-scale models while maintaining high accuracy, fully meeting the dual requirements of efficiency and stability in industrial-grade scenarios. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart illustrating the solution method for a machine learning model of industrial-optimized mixed-integer linear programming provided in this embodiment of the invention.

[0011] Figure 2 A block diagram of the solution system corresponding to the solution method for the machine learning model of industrial optimization mixed integer linear programming provided in the embodiments of the present invention. Detailed Implementation

[0012] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the specific content of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments, which do not constitute a limitation of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0013] First, the following explanations are provided for the terms that may be used in this article: The term "and / or" means that either or both can be achieved simultaneously. For example, X and / or Y means that it includes both "X" or "Y" as well as the three cases of "X and Y".

[0014] The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0015] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.

[0016] Unless otherwise explicitly specified or limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this document according to the specific circumstances.

[0017] When concentration, temperature, pressure, size, or other parameters are expressed as numerical ranges, such ranges should be understood to specifically disclose all ranges formed by any pairing of upper limits, lower limits, or preferred values ​​within that range, regardless of whether the range is explicitly stated; for example, if the numerical range "2 to 8" is stated, then that range should be interpreted to include ranges such as "2 to 7", "2 to 6", "5 to 7", "3 to 4 and 6 to 7", "3 to 5 and 7", "2 and 5 to 7", etc. Unless otherwise stated, the numerical ranges described herein include both their endpoints and all integers and fractions within that range.

[0018] The solution provided by this invention will be described in detail below. Contents not described in detail in the embodiments of this invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, they shall be performed according to conventional conditions in the art or conditions recommended by the manufacturer. Reagents or instruments used in the embodiments of this invention whose manufacturers are not specified are all conventional products that can be purchased commercially.

[0019] like Figure 1 , Figure 2 As shown, this invention provides a method for solving a machine learning model of mixed-integer linear programming for industrial optimization, used to solve a machine learning model of mixed-integer linear programming corresponding to an industrial optimization problem, including: Step 1, Industrial Data Acquisition and Preprocessing: We acquire production-related raw industrial data from multiple sources from various industrial application scenarios, and process the acquired raw industrial data to generate a training dataset that can be directly input into a machine learning model. Step 2, Constraint Learning Model Training: The data in the training dataset generated in step 1 is divided into training set, validation set and test set according to a predetermined ratio, so that data from the same production batch or time period are split into the same set. Based on the implicit coupling and complex system relationships that are difficult to analyze in the industrial scenarios corresponding to the industrial problem to be optimized, at least one matching machine learning model is selected from the candidate model pool with multiple machine learning models of different structures. The selected machine learning models are trained and evaluated through training set, validation set and test set. Based on the evaluation results, the machine learning model that best matches the needs of the industrial scenario corresponding to the industrial problem to be optimized is selected. Step 3, Automatic Model Conversion Processing: The machine learning model structure selected in step 2 is automatically identified and parsed as the parsing result, and the parsing result is unified into an intermediate representation. Each primitive operation in the intermediate representation is mapped to a linear form or piecewise linear form that can be accepted by mixed integer linear programming. Based on the SOS1 constraint, the obtained linear form or piecewise linear form is piecewise linearized to obtain linearized constraints. All the obtained linearized constraints are packaged and output as a standard mixed integer linear programming constraint set and objective function, and a set of binary auxiliary variables corresponding to the objective function is generated. Step 4, optimize the solution: The mixed integer programming solver is used to solve the mixed integer linear programming constraint set, objective function and binary auxiliary variable set formed in step 3 to obtain the corresponding optimal decision variable values; Step 5, output the solution results: The optimal decision variable values ​​obtained in step 4 are analyzed into optimization schemes that can be directly executed and output for the industrial problem to be optimized.

[0020] Preferably, in step 1 of the above method, several industrial application scenarios include: production scheduling industrial application scenario, energy dispatching industrial application scenario, and path planning industrial application scenario. The acquired multi-source raw industrial data types are: work order demand data, quality inspection data, equipment energy consumption data, energy price data, and site topology data.

[0021] Preferably, in step 1 of the above method, multi-source raw industrial data related to production are obtained from several industrial application scenarios in the following manner, and the obtained multi-source raw industrial data are processed to generate a training dataset that can be directly input into a machine learning model, including: By deploying data acquisition units at various locations such as production lines, equipment terminals, quality inspection stations, and energy consumption monitoring points in various industrial application scenarios, multi-source raw industrial data is obtained in real time from various raw industrial data such as PLC signals, SCADA logs, MES work order records, quality inspection results, and environmental sensor readings. The multi-source raw industrial data undergoes preprocessing operations including unified timestamp alignment, resampling interpolation, cleaning of anomalies and missing values, and filtering to remove noise. Fields are then processed based on process knowledge: if a field is a continuous numerical field, it is normalized or standardized; if a field is a discrete categorical field, it is one-hot encoded. This process yields historical sequence data, which is then concatenated into time-series features according to a predetermined window. All obtained time-series features are integrated into a feature matrix and its corresponding label vector, serving as a training set that can be directly input into a machine learning model.

[0022] Preferably, in step 2 of the above method, the machine learning models in the candidate model pool include: Support Vector Machine, Decision Tree, Random Forest, Gradient Boosting Tree, and Multilayer Perceptron; In step 2, the selected machine learning models are trained and evaluated using training sets, validation sets, and test sets in the following manner: Using the training set data, the internal parameters of each selected machine learning model are iteratively adjusted through optimization algorithms to minimize the prediction error as the loss function, enabling the machine learning model to learn complex relationships in the data. During this process, the hyperparameters of the machine learning model are tuned to prevent overfitting in conjunction with the validation set to obtain better fitting results. The optimal parameter configuration is selected according to predefined evaluation metrics to obtain the optimal machine learning model. Finally, the optimal machine learning model after training is evaluated on the test set. In step 2, if only one matching machine learning model is selected from the candidate model pool, then the machine learning model is directly used as the machine learning model that best matches the industrial scenario requirements corresponding to the industrial problem to be optimized, based on the evaluation results. If more than one matching machine learning model is selected from the candidate model pool, the machine learning model that best matches the industrial scenario requirements corresponding to the industrial problem to be optimized is selected based on the evaluation results in the following manner: Choose an appropriate evaluation index based on the nature of the industrial problem to be optimized: if the industrial problem to be optimized is a regression problem (such as energy dispatching problem, production scheduling problem), then use the mean squared error as the evaluation index and select the machine learning model with the smallest mean squared error; If the industrial problem to be optimized is a classification problem (such as fault diagnosis or quality inspection), then classification accuracy and recall are used as evaluation metrics. For industrial scenarios with high accuracy requirements (such as quality control), the machine learning model with the highest accuracy is selected. For industrial scenarios with high recall requirements (such as safety warning), the machine learning model with the highest recall is selected. For industrial scenarios that require comprehensive consideration, the machine learning model with the highest comprehensive index F1-Score is selected. The machine learning models selected above are chosen as the most suitable machine learning models for the industrial scenario requirements corresponding to the industrial problem to be optimized.

[0023] Preferably, in step 3 of the above method, the automatic conversion engine module performs automatic model conversion processing in the following manner, including: The automatic conversion engine module's model parsing layer uses lightweight reflection statements to determine the machine learning library to which the model belongs, automatically identifying and parsing the machine learning model structure selected in step 2 as the parsing result. The machine learning model structure includes: network hierarchy, tree structure, activation function, and splitting condition; the parsing result is unified into an intermediate representation including layer type, weight, threshold, and leaf node rules. The linearization generation layer of the automatic transformation engine module maps each primitive operation in the intermediate representation to a linear or piecewise linear form acceptable to mixed-integer linear programming.

[0024] The above model parsing layer uses lightweight reflection statements to determine the machine learning library to which the model belongs, automatically identifies and parses the machine learning model structure selected in step 2 as the parsing result. The processing flow is as follows: Step A1, Identification based on inheritance relationship: Use a lightweight reflection mechanism to obtain the type inheritance chain of the machine learning model, and make the first judgment by whether the chain contains the iconic base class of a specific machine learning framework; Step A2, Identification based on module path: If the inheritance relationship cannot uniquely determine the source of the framework, the module path of the machine learning model type is obtained through reflection, and a second judgment is made based on the path prefix; Step A3, Route Resolution: Map the framework identifier of the machine learning model identified in the previous steps to a pre-set parser routing table, thereby automatically assigning the machine learning model to the parsing function dedicated to that machine learning model framework.

[0025] Preferably, in the above method, the linearization generation layer maps each primitive operation in the intermediate representation to a linear form or piecewise linear form acceptable to mixed-integer linear programming. If the primitive operation is a ReLU activation function ( If ), then expand as follows: ; in, This represents the intermediate variable before ReLU activation, used as a benchmark for comparison in subsequent activation and linearization constraints; The linear coefficients of the current neuron with respect to each input dimension are related to... Equal-length weight vectors are used to determine the slope direction and magnitude of the affine transformation; This represents the input vector corresponding to the current neuron, used for affine transformation. ; The bias is a scalar used to shift the overall output of the affine transformation. This represents the ReLU output, which is used as the final output of this neuron and passed to subsequent layers or the target function. Representing variables The known lower bound, that is, within the given input range. The minimum achievable value is used to replace the common large constant Big-M during linearization, making the constraints tighter and more stable; Represents a binary indicator variable. This indicates that the neuron is in an activated state. Indicates the off state, used to convert non-linear maximum logic into a mutual exclusion relationship of 0 or 1; Representing variables The known upper bound, that is, within the given input range. The maximum value that can be achieved, used in conjunction with Limited Edition The upper bound when it is 0; If the primitive operation is a branch in a tree model, then let For the input feature vector, For the set of leaf nodes, for each Introducing binary variables For each internal node ,remember For segmentation features, The segmentation threshold is: ;in It is a small positive number.

[0026] The primitive operations in the above method refer to the functional units behind the four core pieces of information in IR (Intermediate Representation)—layer type, weights, thresholds, and leaf node rules: affine layers, activation layers, aggregation layers, and decision splitting. These include: linear / affine primitives (such as linear layers), activation primitives (activation functions such as ReLU and sigmoid), decision tree primitives (decision tree branching, random forests), and other common primitives (softmax, argmax, etc.).

[0027] Preferably, in step 3 of the above method, the obtained linear form or piecewise linear form is linearized based on the SOS1 constraint to obtain linearization constraints in the following manner: Based on SOS1 constraints, the ReLU activation function ( The linearization of ) can be expanded as follows: ; in, This represents the intermediate variable before ReLU activation, which serves as a benchmark for comparison in subsequent activation and linearization constraints. The linear coefficients of the current neuron with respect to each input dimension are related to... Equal-length weight vectors are used to determine the slope direction and magnitude of the affine transformation; This represents the input vector corresponding to the current neuron, which is either the output of the previous layer or the network input, and is used to participate in affine transformation. ; The bias is a scalar used to shift the overall output of the affine transformation. This represents the ReLU output, which is used as the final output of this neuron and passed to subsequent layers or the target function. Represents a nonnegative slack variable, characterizing when The inverse vector that needs to be compensated is used to... The SOS1 constraint is formed to ensure that at most one of the two is positive, thereby realizing ReLU logic; Based on SOS1 constraints, the argmax function ( The linearization of ) can be expanded as follows: ; in, express; Indicates the first One candidate value; This indicates the total number of candidate values; Indicates the first The nth candidate value, i.e., the nth value from the previous layer or the model output vector. Item, used with maximum value Compare; Indicates the first A non-negative slack variable representing the difference between a candidate value and its maximum value; if this candidate is selected as the largest, then... ; This represents the maximum value itself, used to provide a reference benchmark so that the constraint group characterizes the behavior of argmax; This represents a binary indicator variable, where 1 indicates the first... The term is argmax, used in conjunction with... This constitutes an SOS1 constraint; The linear constraints obtained above are packaged and output as a standard mixed-integer linear programming constraint set. The objective function is This generates a set of binary auxiliary variables corresponding to the objective function. .

[0028] Preferably, in step 4 of the above method, the mixed integer programming solver used is the open-source solver SCIP; The open-source solver adopts a hybrid optimization strategy, combining the branch and bound method to solve the mixed integer linear programming constraint set, objective function, and binary auxiliary variable set formed in step 3.

[0029] This approach employs a hybrid optimization strategy in the specific open-source solver, combining branch and bound methods to accelerate convergence while maintaining solution space coverage. Leveraging the native support for SOS1 constraints in the open-source solver, it dynamically generates cutting planes and optimizes branch node selection strategies for large-scale constraint sets embedded in complex models, reducing solution time and memory consumption, thus achieving efficient and stable solutions in industrial-grade scenarios. The solution is then applied to the mixed-integer linear programming constraint set, objective function, and binary auxiliary variable set formed in step 3.

[0030] Preferably, in step 5 of the above method, the optimal decision variable values ​​obtained in step 4 are parsed into an optimization scheme that can be directly executed and output to correspond to the industrial problem to be optimized, in the following manner: If the decision variable value returned by the mixed integer programming solver in step 4 corresponds to the equipment scheduling data, then it is parsed into directly executable equipment scheduling instructions as the corresponding optimization scheme and output. If the decision variable values ​​returned by the mixed integer programming solver in step 4 correspond to energy allocation data, then the energy allocation strategy that can be directly executed is parsed as the corresponding optimization scheme and output. If the decision variable values ​​returned by the mixed integer programming solver in step 4 correspond to logistics path data, then the solution is parsed into a directly executable logistics path sequence as the corresponding optimization scheme and output.

[0031] Preferably, step 5 of the above method further includes: converting the optimization scheme to be output into a structured data file or a visualization chart and then outputting it; and a verification step, which verifies the results of the optimization scheme to be output by using a built-in standardized instance library, and calculates key indicators by comparing the actual optimization results with the expected targets.

[0032] This invention also provides a system for solving a machine learning model of mixed-integer linear programming for industrial optimization that implements the above-described method, comprising: The system includes a multi-source data acquisition module, a constraint learning model training module, an automatic conversion engine module, an optimization solution module, and a result output module; among which, The multi-source data acquisition module can acquire production-related raw industrial data from several industrial application scenarios, and process the acquired raw industrial data to generate a training dataset that can be directly input into a machine learning model. The constrained learning model training module is communicatively connected to the multi-source data acquisition module. It can divide the data in the training dataset generated by the multi-source data acquisition module into training set, validation set and test set according to a predetermined ratio, such as a 6:2:2 ratio, so that data from the same production batch or time period are split into the same set. Based on the implicit coupling and complex system relationships that are difficult to analyze in the industrial scenarios corresponding to the industrial problem to be optimized, the matching machine learning models are selected from the candidate model pool with multiple machine learning models of different structures. The selected machine learning models are trained and evaluated through training sets, validation sets and test sets. Based on the evaluation results, the machine learning model that best matches the needs of the industrial scenario corresponding to the industrial problem to be optimized is selected. The candidate model pool includes machine learning models from different frameworks, such as support vector machines (SVM), decision trees, random forests, gradient boosting trees (GBDT), and neural networks (MLP). The automatic conversion engine module communicates with the constraint learning model training module. It can automatically identify and parse the machine learning model structure selected by the constraint learning model training module as the parsing result, unify the parsing result into an intermediate representation, map each primitive operation in the intermediate representation to a linear form or piecewise linear form acceptable to mixed integer linear programming, perform piecewise linearization on the obtained linear form or piecewise linear form based on SOS1 constraints to obtain linearized constraints, package all obtained linearized constraints, output as a standard mixed integer linear programming constraint set and objective function, and generate a set of binary auxiliary variables corresponding to the objective function. The optimization solution module is connected to the automatic conversion engine module and can solve the mixed integer linear programming constraint set, objective function and binary auxiliary variable set formed in step 3 through the mixed integer programming solver to obtain the corresponding optimal decision variable values. The result output module is communicatively connected to the optimization solution module. It can parse the optimal decision variable values ​​obtained by the optimization solution module into an optimization scheme that can be directly executed and output the solution corresponding to the industrial problem to be optimized.

[0033] Preferably, in the above system, the automatic conversion engine module includes: The model consists of a parsing layer, a linearization generation layer, a linearization constraint processing layer, and an output layer; among which, The model parsing layer can automatically identify and parse the machine learning model structure selected in step 2 as the parsing result by using a lightweight reflection statement to determine the model's affiliation. The model structure includes: network hierarchy, tree structure, activation function and splitting condition; the parsing result is unified into an intermediate representation including layer type, weight, threshold and leaf node rules. The linearization generation layer, which is communicatively connected to the model parsing layer, can map each primitive operation in the intermediate representation output by the model parsing layer to a linear form or piecewise linear form acceptable to mixed-integer linear programming. The constraint encoding output layer is communicatively connected to the linearization generation layer. It can perform piecewise linearization on the linear form or piecewise linear form obtained by the linearization generation layer based on SOS1 constraints to obtain linearized constraints. It also packages the linearized constraints obtained by the linearization constraint processing layer and outputs them as a standard mixed-integer linear programming constraint set and objective function, and generates a set of binary auxiliary variables corresponding to the objective function.

[0034] Preferably, in the above system, the linearization generation layer maps each primitive operation in the intermediate representation to a linear or piecewise linear form acceptable to mixed-integer linear programming. If the primitive operation is a ReLU activation function ( If ), then expand as follows: ; in, This represents the intermediate variable before ReLU activation, used as a benchmark for comparison in subsequent activation and linearization constraints; The linear coefficients of the current neuron with respect to each input dimension are related to... Equal-length weight vectors are used to determine the slope direction and magnitude of the affine transformation; This represents the input vector corresponding to the current neuron, used for affine transformation. ; The bias is a scalar used to shift the overall output of the affine transformation. This represents the ReLU output, which is used as the final output of this neuron and passed to subsequent layers or the target function. Representing variables The known lower bound, that is, within the given input range. The minimum achievable value is used to replace the common large constant Big-M during linearization, making the constraints tighter and more stable; Represents a binary indicator variable. This indicates that the neuron is in an activated state. Indicates the off state, used to convert non-linear maximum logic into a mutual exclusion relationship of 0 or 1; Representing variables The known upper bound, that is, within the given input range. The maximum value that can be achieved, used in conjunction with Limited Edition The upper bound when it is 0; If the primitive operation is a branch in a tree model, then let For the input feature vector, For the set of leaf nodes, for each Introducing binary variables For each internal node ,remember For segmentation features, The segmentation threshold is: ;in It is a small positive number.

[0035] Preferably, in the above system, the linearization constraint processing layer performs piecewise linearization on the obtained linear form or piecewise linear form based on SOS1 constraints to obtain linearization constraints in the following manner: Based on SOS1 constraints, the ReLU activation function ( The linearization of ) can be expanded as follows: ; in, This represents the intermediate variable before ReLU activation, which serves as a benchmark for comparison in subsequent activation and linearization constraints. The linear coefficients of the current neuron with respect to each input dimension are related to... Equal-length weight vectors are used to determine the slope direction and magnitude of the affine transformation; This represents the input vector corresponding to the current neuron, which is either the output of the previous layer or the network input, and is used to participate in affine transformation. ; The bias is a scalar used to shift the overall output of the affine transformation. This represents the ReLU output, which is used as the final output of this neuron and passed to subsequent layers or the target function. Represents a nonnegative slack variable, characterizing when The inverse vector that needs to be compensated is used to... The SOS1 constraint is formed to ensure that at most one of the two is positive, thereby realizing ReLU logic; Based on SOS1 constraints, the argmax function ( The linearization of ) can be expanded as follows: ; in, This indicates the output index of argmax, i.e., the index of the selected candidate value; The index of the candidate value belongs to the set. ; The set representing the candidate value indices, with a total size of ; Indicates the first The nth candidate value, that is, the nth value from the output vector of the previous layer of the neural network or model. Item, used with maximum value Compare; Indicates the first Candidate values ​​and maximum value If the non-negative slack variable between these two values ​​is selected as the maximum value, then... ,otherwise ; This represents the maximum value itself, used to establish a reference baseline and ensure that the constraint group can characterize the behavior of argmax; This represents a binary indicator variable, where 1 indicates the first... The term is the maximum value if it is specified, otherwise it is 0; the SOS1 constraint is used to ensure that when... hour ; The linear constraints obtained above are packaged and output as a standard mixed-integer linear programming constraint set. The objective function is This generates a set of binary auxiliary variables corresponding to the objective function. .

[0036] In summary, this invention utilizes an end-to-end automatic conversion engine applicable to multi-framework training models to parse and map various machine learning models into linear constraints or piecewise linear constraints that can be embedded in mixed-integer linear programming with a single click. This significantly simplifies the model conversion process and reduces manual derivation and debugging. Simultaneously, the adoption of robust modeling techniques based on SOS1 constraints eliminates the dependence on extremely large constants, resulting in better numerical stability and faster convergence speed for the solver during branch and bound processes. Combined with a hybrid global + local optimization strategy and automatic pruning of redundant constraints before solving, both overall solution efficiency and solution quality are significantly improved.

[0037] Furthermore, the method of this invention is compatible with mainstream training frameworks, eliminating the closed-loop dependence on specific commercial ecosystems and significantly reducing the cost and cycle of cross-platform deployment. Application verification in various typical industrial scenarios (such as production scheduling, energy dispatching, and route planning) shows that this invention can not only quickly complete the automatic embedding and solving of large-scale models, but also achieve the integrated implementation of the "prediction-optimization" closed loop while maintaining high accuracy, fully meeting the dual requirements of efficiency and stability in industrial scenarios.

[0038] To more clearly demonstrate the technical solution and its effects provided by the present invention, the following detailed description of the solution provided by the embodiments of the present invention is provided with reference to specific examples.

[0039] Example 1 This embodiment provides a method for solving machine learning models of mixed-integer linear programming (MILP) in industrial optimization. It is an automated modeling method for solving MILP models in industrial optimization. By designing a cross-framework parsing and constraint generation mechanism, it achieves automatic conversion of model structures (including neural networks, decision trees, ensemble learning, etc.) from mainstream machine learning frameworks (such as PyTorch, Scikit-Learn, XGBoost, etc.). This engine can parse the internal logic of the model (such as activation functions, branch conditions, feature weights, etc.) into MILP-compatible linear constraints or piecewise linear constraints, eliminating the need for manual derivation or coding, significantly reducing the complexity and labor costs of multi-model embedding. Compared to the limitations of existing technologies that rely on a single framework or simple model, this method overcomes the cross-platform adaptation bottleneck, supporting "one-click" seamless embedding of complex data-driven models in industrial scenarios, providing flexible and universal optimization modeling capabilities for multiple industries.

[0040] Meanwhile, addressing the numerical instability and surge in redundant variables caused by the setting of external variable boundaries in traditional linearization methods (such as the Big-M method), this invention employs a robust modeling technique based on SOS1 (Special Ordered Set of Type 1) constraints to achieve robust representation of machine learning models. By introducing the synergistic design of mutually exclusive variable sets and piecewise linear constraints, SOS1 constraints can accurately characterize the logical relationships of nonlinear functions (such as ReLU activation functions) without relying on predefined boundaries, effectively avoiding solution space deviation or solution failure caused by boundary errors. Compared to traditional methods, this technique significantly improves the embedding accuracy and solution convergence efficiency of deep models (such as multi-layer neural networks) in MILP, providing a highly stable optimization modeling foundation for complex industrial scenarios.

[0041] The following section will begin with "data-driven constraint learning," explaining the two core innovations of this invention: a "multi-framework compatible automatic model conversion engine" and "robust modeling based on SOS1 constraints." It will also illustrate how these innovations synergistically drive the final solution of mixed-integer linear programming (MILP). Figure 1 As shown, the method includes the following steps: Step 1, Industrial Data Acquisition and Preprocessing: Multi-source raw industrial data is collected from industrial optimization scenarios such as production and sales scheduling, process parameters, energy consumption monitoring, and demand forecasting to construct a feature matrix. and label vectors The raw industrial data is cleaned, normalized, and imputed for missing values ​​to generate a training dataset that can be directly input into machine learning models.

[0042] Step 2, Constraint Learning Model Training:

[0043] The data in the training dataset generated in step 1 is divided into training set, validation set and test set according to a predetermined ratio, so that data from the same production batch or time period are split into the same set. For complex system relationships that are implicit or difficult to resolve, at least one matching machine learning model is selected from the candidate model pool, including Support Vector Machine (SVM), Decision Tree, Random Forest, Gradient Boosting Tree (GBDT), and Neural Network (MLP), and the selection is performed in the candidate model library. The fitting process involves training and evaluating the selected machine learning model using training, validation, and test sets. Evaluation metrics include mean squared error (MSE), classification accuracy, and recall. Based on the evaluation results, the machine learning model that best matches the requirements of the industrial scenario to be optimized is selected.

[0044] Step 3, Automatic Model Conversion Processing: The machine learning model selected in step 2 is parsed by constructing a multi-framework compatible automatic conversion engine module. The core of this multi-framework compatible automatic conversion engine module is that it seamlessly supports mainstream training frameworks such as PyTorch, TensorFlow / Keras, Scikit-Learn, XGBoost, and LightGBM.

[0045] (31) Model analysis processing: The automatic conversion engine module automatically identifies and parses the structure (network hierarchy, tree structure, activation function, splitting condition, etc.) of the machine learning model selected in step 2 (which can be a machine learning model under different frameworks). The parsing results are then unified into an intermediate representation (IR), including layer type, weights, thresholds, leaf node rules, etc.

[0046] At the model parsing layer, the automatic conversion engine directly performs runtime reflection on native objects from frameworks such as PyTorch, TensorFlow / Keras, Scikit-Learn, XGBoost, and LightGBM. It reads the hierarchical connections, weight matrices, and activation functions of neural networks, or the splitting features, thresholds, and leaf values ​​of tree models, and then integrates this heterogeneous information into a unified intermediate representation (IR). The IR records the op type, input dependencies, tensor / scalar shape, and necessary parameters (such as thresholds, weights, and biases) of each node, thus providing a framework-independent but semantically complete "structural blueprint" for subsequent conversions.

[0047] (32) Linearization generation process: The linearization generation layer of the automatic transformation engine module maps each primitive operation in the intermediate representation IR to a linear or piecewise linear form acceptable to MILP. For example:

[0048] (321) ReLU activation function Open as: ;

[0049] (322) Tree model branching: Let For the input feature vector, For the set of leaf nodes, for each Introducing binary variables For each internal node ,remember Segmentation features The segmentation threshold is: ;in It is a small positive number;

[0050] (33) Robust modeling based on SOS1 constraints: To overcome the heavy dependence of the Big-M method on variable boundaries, this embodiment introduces SOS1 constraints for piecewise linearization, improving numerical robustness and solution efficiency: (321) ReLU activation function Expanded as: ; (322) argmax function ( The linearization of ) can be expanded as follows: ; By adopting SOS1 constraints, the dedicated split handler of the mixed integer programming solver (which can be an open-source solver such as SCIP) can directly utilize the mutual exclusion set structure during the branch and bound process, reducing redundant nodes and weak boundaries, and significantly improving branch quality and convergence speed.

[0051] Upon entering the linearization generation layer, the engine traverses the Integer Linear Programming (IR), mapping each primitive operation in real-time to a linear or piecewise linear expression acceptable to Mixed Integer Linear Programming (MILP): continuous activations such as ReLU and Leaky-ReLU are decomposed into Big-M / SOS1 constraints with binary variables; decision tree paths are ensured to have unique leaf nodes by introducing mutually exclusive 0 or 1 variables; discrete operations such as argmax are also translated into mutually exclusive linear systems. To eliminate the dependence of Big-M on upper and lower bounds, the system prioritizes the use of SOS1 templates to obtain more stable numerical performance, and performs interval reasoning on local boundaries when necessary to further reduce the coefficient amplitude. This process, along with the newly added auxiliary variables and local boundaries, is written back to the internal constraint table, preparing for final encoding.

[0052] (34) Constraint encoding output: The above linearization constraints are packaged and output as a standard MILP constraint set. With objective function And at the same time generate a corresponding set of binary auxiliary variables. .

[0053] At the constraint encoding output layer, all processed linear and piecewise linear constraints are packaged at once: the matrix A·x≤b, the integer field specification, the SOS1 mutual exclusion set, and the objective function are output as standard .lp / .mps files, or directly injected into the solver through the mixed integer programming solver interface.

[0054] Step 4, optimize the solution: Efficient optimization is achieved through deep integration of open-source solvers (such as SCIP). Based on automatically generated MILP constraints, redundant constraints are first dynamically identified, eliminating those weakly correlated with or repetitive with the objective function, thus simplifying the problem size. Subsequently, a hybrid optimization strategy is employed, combining the global search capability of branch and bound with local tuning techniques such as neighborhood search and gradient descent to accelerate convergence while ensuring solution space coverage. For large-scale constraint sets embedded in complex models, the solver's native support for SOS1 constraints is utilized to dynamically generate cutting planes and optimize branch node selection strategies, significantly reducing solution time and memory consumption, ensuring efficient and stable solutions in industrial-grade scenarios.

[0055] Step 5, Output the results: The decision variable values ​​returned by the open-source solver in step 4 are parsed into directly executable industrial optimization solutions, such as equipment scheduling instructions, energy allocation strategies, or logistics path sequences, and output as structured data files or visualization charts. Simultaneously, the results are validated using a built-in standardized instance library, comparing the actual optimization effect with the expected target, calculating key indicators (such as cost reduction rate and resource utilization rate). If the deviation exceeds a threshold, a feedback mechanism is triggered to automatically adjust model parameters or optimization strategies, forming a closed-loop process of "modeling-solving-validation-iteration" to ensure the continuous applicability and reliability of the technical solution in dynamic industrial environments.

[0056] Example 2 This embodiment provides a method for solving a machine learning model of mixed-integer linear programming (MILP) for industrial optimization, applicable to chemical production processes. This method requires predicting equipment anomaly risks based on real-time sensor data (such as reactor temperature and pressure) and dynamically optimizing production parameters. First, historical sensor data and anomaly records are collected, and a multilayer perceptron (MLP) model is trained using the PyTorch framework to learn nonlinear fault associations. The method's multi-framework compatible automatic conversion engine module automatically parses the MLP's layer structure and ReLU activation function, employing SOS1 constraints to piecewise linearize the hidden layer outputs, generating MILP-compatible fault probability constraints. Subsequently, a production optimization model is constructed, aiming to maximize output and minimize fault risk. Constraints include material flow rate (continuous variable), valve on / off states (0-1 variables), and anomaly probability limits predicted by the MLP. An open-source solver is invoked to execute a hybrid optimization strategy, combining global branch and bound with local gradient descent to quickly output the optimal parameter combination under safe operating conditions. The results are parsed into real-time control commands. Verification using an example library shows that the obtained optimization scheme significantly reduces fault risk while shortening solution time, without requiring manual intervention in the model conversion process.

[0057] Example 3 This embodiment provides a solution method for a machine learning model of mixed-integer linear programming (MILP) for industrial optimization, applicable to regional energy management. This method requires predicting next day's electricity demand based on weather, historical load, and other data, and optimizing generator scheduling. After integrating multi-source data, a random forest model is trained using Scikit-Learn to capture the nonlinear relationship between electricity demand and meteorological factors. An automatic conversion engine parses the multiple decision tree structure in the forest, assigning binary variables to the leaf nodes of each tree. SOS1 constraints are used to precisely express branching conditions (such as temperature threshold division), outputting a weighted summation form of MILP demand forecast constraints. When constructing the generator scheduling model, the objective is to minimize power generation costs. Constraints cover generator start-up and shutdown (integer variables), output limits (continuous variables), and the electricity demand association rules predicted by the random forest. During the solution process, redundant tree path constraints are dynamically identified, and a global genetic algorithm combined with local neighborhood search is used to generate the most economically optimal power generation plan. The results are verified by comparison with a sample library, showing a significant reduction in demand forecast bias and superior solution stability compared to traditional linearization methods, making it suitable for the complex decision-making needs of multi-energy collaborative scenarios.

[0058] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0059] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for solving a machine learning model of mixed-integer linear programming for industrial optimization, characterized in that, Machine learning models are used to solve mixed-integer linear programming problems corresponding to industrial optimization problems, including: Step 1, Industrial Data Acquisition and Preprocessing: We acquire production-related raw industrial data from multiple sources from various industrial application scenarios, and process the acquired raw industrial data to generate a training dataset that can be directly input into a machine learning model. Step 2, Constraint Learning Model Training: The data in the training dataset generated in step 1 is divided into training set, validation set and test set according to a predetermined ratio, so that data from the same production batch or time period are split into the same set. Based on the implicit coupling and complex system relationships that are difficult to analyze in the industrial scenarios corresponding to the industrial problem to be optimized, at least one matching machine learning model is selected from the candidate model pool with multiple machine learning models of different structures. The selected machine learning models are trained and evaluated through training set, validation set and test set. Based on the evaluation results, the machine learning model that best matches the needs of the industrial scenario corresponding to the industrial problem to be optimized is selected. In step 2, if only one matching machine learning model is selected from the candidate model pool, then the machine learning model is directly used as the machine learning model that best matches the industrial scenario requirements corresponding to the industrial problem to be optimized, based on the evaluation results. If more than one matching machine learning model is selected from the candidate model pool, the machine learning model that best matches the industrial scenario requirements corresponding to the industrial problem to be optimized is selected based on the evaluation results in the following manner: Choose an appropriate evaluation index based on the nature of the industrial problem to be optimized: if the industrial problem to be optimized is a regression problem, then use the mean squared error as the evaluation index and select the machine learning model with the smallest mean squared error. If the industrial problem to be optimized is a classification problem, then classification accuracy and recall are used as evaluation indicators. For industrial scenarios with high accuracy requirements, the machine learning model with the highest accuracy is selected; for industrial scenarios with high recall requirements, the machine learning model with the highest recall is selected; for industrial scenarios requiring comprehensive consideration, the machine learning model with the highest comprehensive indicator F1-Score is selected. The machine learning models selected above are chosen as the most suitable machine learning models for the industrial scenario requirements corresponding to the industrial problem to be optimized. Step 3, Automatic Model Conversion Processing: The machine learning model structure selected in step 2 is automatically identified and parsed as the parsing result, and the parsing result is unified into an intermediate representation. Each primitive operation in the intermediate representation is mapped to a linear form or piecewise linear form that can be accepted by mixed integer linear programming. Based on the SOS1 constraint, the obtained linear form or piecewise linear form is piecewise linearized to obtain linearized constraints. All the obtained linearized constraints are packaged and output as a standard mixed integer linear programming constraint set and objective function, and a set of binary auxiliary variables corresponding to the objective function is generated. In step 3, the automatic conversion engine module performs automatic model conversion processing in the following manner, including: The automatic conversion engine module's model parsing layer uses lightweight reflection statements to determine the machine learning library to which the model belongs, automatically identifying and parsing the machine learning model structure selected in step 2 as the parsing result. The machine learning model structure includes: network hierarchy, tree structure, activation function, and splitting condition; the parsing result is unified into an intermediate representation including layer type, weight, threshold, and leaf node rules. The linearization generation layer of the automatic transformation engine module maps each primitive operation in the intermediate representation to a linear or piecewise linear form acceptable to mixed-integer linear programming. Step 4, optimize the solution: The mixed integer programming solver is used to solve the mixed integer linear programming constraint set, objective function and binary auxiliary variable set formed in step 3 to obtain the corresponding optimal decision variable values; Step 5, output the solution results: The optimal decision variable values ​​obtained in step 4 are analyzed into optimization schemes that can be directly executed and output for the industrial problem to be optimized.

2. The method for solving the machine learning model of mixed-integer linear programming for industrial optimization according to claim 1, characterized in that, In step 1, several industrial application scenarios include: production scheduling industrial application scenario, energy dispatching industrial application scenario, and path planning industrial application scenario. The acquired multi-source raw industrial data types are: work order demand data, quality inspection data, equipment energy consumption data, energy price data, and site topology data.

3. The method for solving the machine learning model of mixed-integer linear programming for industrial optimization according to claim 1 or 2, characterized in that, In step 1, production-related multi-source raw industrial data are acquired from several industrial application scenarios in the following manner, and the acquired multi-source raw industrial data are processed to generate a training dataset that can be directly input into a machine learning model, including: By deploying data acquisition units at various locations such as production lines, equipment terminals, quality inspection stations, and energy consumption monitoring points in various industrial application scenarios, multi-source raw industrial data is obtained in real time from various raw industrial data such as PLC signals, SCADA logs, MES work order records, quality inspection results, and environmental sensor readings. The multi-source raw industrial data undergoes preprocessing operations including unified timestamp alignment, resampling interpolation, cleaning of anomalies and missing values, and filtering to remove noise. Fields are then processed based on process knowledge: if a field is a continuous numerical field, it is normalized or standardized; if a field is a discrete categorical field, it is one-hot encoded. This process yields historical sequence data, which is then concatenated into time-series features according to a predetermined window. All obtained time-series features are integrated into a feature matrix and its corresponding label vector, serving as a training set that can be directly input into a machine learning model.

4. The method for solving the machine learning model of mixed-integer linear programming for industrial optimization according to claim 1 or 2, characterized in that, In step 2, the machine learning models in the candidate model pool include: Support Vector Machine, Decision Tree, Random Forest, Gradient Boosting Tree, and Multilayer Perceptron; In step 2, the selected machine learning models are trained and evaluated using training sets, validation sets, and test sets in the following manner: Using the training set data, the internal parameters of each selected machine learning model are iteratively adjusted through optimization algorithms to minimize the prediction error as the loss function, enabling the machine learning model to learn complex relationships in the data. During this process, the hyperparameters of the machine learning model are tuned to prevent overfitting in conjunction with the validation set, and the optimal parameter configuration is selected according to predefined evaluation metrics to obtain the optimal machine learning model. Finally, the optimal machine learning model after training is evaluated on the test set.

5. The method for solving the machine learning model of mixed-integer linear programming for industrial optimization according to claim 1, characterized in that, The linearization generation layer maps each primitive operation in the intermediate representation to a linear or piecewise linear form acceptable to mixed-integer linear programming, where the primitive operation is a ReLU activation function: Then it expands to: ； in, This represents the intermediate variable before ReLU activation, used as a benchmark for comparison in subsequent activation and linearization constraints; The linear coefficients of the current neuron with respect to each input dimension are related to... Equal-length weight vectors are used to determine the slope direction and magnitude of the affine transformation; This represents the input vector corresponding to the current neuron, used for affine transformation. ; The bias is a scalar used to shift the overall output of the affine transformation. This represents the ReLU output, which is used as the final output of this neuron and passed to subsequent layers or the target function. Representing variables The known lower bound, that is, within the given input range. The minimum achievable value is used to replace the common large constant Big-M during linearization; Represents a binary indicator variable. This indicates that the neuron is in an activated state. Indicates the off state, used to convert non-linear maximum logic into a mutual exclusion relationship of 0 or 1; Representing variables The known upper bound, that is, within the given input range. The maximum value that can be achieved, used in conjunction with Limited Edition The upper bound when it is 0; If the primitive operation is a branch in a tree model, then let For the input feature vector, For the set of leaf nodes, for each Introducing binary variables For each internal node ,remember For segmentation features, The segmentation threshold is: ;in It is a small positive number.

6. The method for solving the machine learning model of mixed-integer linear programming for industrial optimization according to claim 5, characterized in that, In step 3, the obtained linear form or piecewise linear form is linearized based on the SOS1 constraints to obtain linearization constraints in the following manner: ReLU activation function based on SOS1 constraints: The linearization, expanded as follows: ； in, This represents the intermediate variable before ReLU activation, which serves as a benchmark for comparison in subsequent activation and linearization constraints. The linear coefficients of the current neuron with respect to each input dimension are related to... Equal-length weight vectors are used to determine the slope direction and magnitude of the affine transformation; This represents the input vector corresponding to the current neuron, which is either the output of the previous layer or the network input, and is used to participate in affine transformation. ; The bias is a scalar used to shift the overall output of the affine transformation. This represents the ReLU output, which is used as the final output of this neuron and passed to subsequent layers or the target function. Represents a non-negative slack variable, characterizing when The inverse vector that needs to be compensated is used to... The SOS1 constraint is formed to ensure that at most one of the two is positive, thereby realizing ReLU logic; Based on SOS1 constraints, the argmax function is: The linearization, expanded as follows: ； in, This indicates the output index of argmax, i.e., the index of the selected candidate value; The index of the candidate value belongs to the set. ; The set representing the candidate value indices, with a total size of ; Indicates the first The nth candidate value, that is, the nth value from the output vector of the previous layer of the neural network or model. Item, used with maximum value Compare; Indicates the first Candidate values ​​and maximum value If the non-negative slack variable between these two values ​​is selected as the maximum value, then... ,otherwise ; This represents the maximum value itself, used to establish a reference baseline and ensure that the constraint group can characterize the behavior of argmax; This represents a binary indicator variable, where 1 indicates the first... The term is the maximum value if it is specified, otherwise it is 0; the SOS1 constraint is used to ensure that when... hour ; The linear constraints obtained above are packaged and output as a standard mixed-integer linear programming constraint set. The objective function is This generates a set of binary auxiliary variables corresponding to the objective function. .

7. The method for solving the machine learning model of mixed-integer linear programming for industrial optimization according to claim 1 or 2, characterized in that, In step 4, the mixed integer programming solver used is the open-source solver SCIP. The open-source solver adopts a hybrid optimization strategy, combining the branch and bound method to solve the mixed integer linear programming constraint set, objective function, and binary auxiliary variable set formed in step 3.

8. The method for solving the machine learning model of mixed-integer linear programming for industrial optimization according to claim 1 or 2, characterized in that, In step 5, the optimal decision variable values ​​obtained in step 4 are parsed into an optimization scheme that can be directly executed and output, corresponding to the industrial problem to be optimized, in the following manner: If the decision variable value returned by the solver in step 4 corresponds to equipment scheduling data, then it is parsed into directly executable equipment scheduling instructions as the corresponding optimization scheme and output. If the decision variable values ​​returned by the solver in step 4 correspond to energy allocation data, then the energy allocation strategy that can be directly executed is parsed and output as the corresponding optimization scheme. If the decision variable values ​​returned by the solver in step 4 correspond to logistics path data, then the solution is parsed into a directly executable logistics path sequence as the corresponding optimization scheme and output.

9. The method for solving the machine learning model of mixed-integer linear programming for industrial optimization according to claim 8, characterized in that, Step 5 further includes: converting the optimization scheme to be output into a structured data file or a visualization chart and then outputting it; and a verification step, which verifies the results of the optimization scheme to be output by using the built-in standardized instance library, and calculates key indicators by comparing the actual optimization results with the expected target.

Images