Intelligent scrap steel matching method and equipment based on multi-agent collaborative decision-making and medium

Abstract

The invention is suitable for the technical field of iron and steel smelting, and provides an intelligent scrap steel preparation method and device based on multi-agent collaborative decision making and a medium. The method comprises the steps that molten iron information, target steel grade information, scrap steel information, device information, cost information and constraint information are obtained; based on a heat balance intelligent agent, a material balance intelligent agent and a cost control intelligent agent, an intelligent steel scrap distribution model is constructed; the molten iron information, the target steel grade information, the waste steel information, the equipment information, the cost information and the constraint information are input into the intelligent waste steel matching model, heat balance optimization, material balance optimization and component control optimization are conducted through a heat balance intelligent agent, a material balance intelligent agent and a cost control intelligent agent, and a target decision scheme is determined. According to the technical scheme provided by the invention, heat balance, material balance and cost control overall optimization are carried out through a multi-agent collaborative architecture, intelligent scrap steel proportioning with multi-dimensional target collaboration is realized, and the accuracy of scrap steel proportioning is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of iron and steel smelting technology, and particularly relates to an intelligent scrap steel allocation method, equipment and medium based on multi-agent collaborative decision-making. Background Technology

[0002] In the modern steel industry production system, steel smelting is the core link, and its production efficiency, product quality and cost control directly determine the market competitiveness of enterprises. As one of the key raw materials in the steel smelting process, the scientific and rational nature of its proportioning scheme has a decisive impact on the entire smelting process.

[0003] With the rapid development of industrial intelligent technology, the application of technologies such as artificial intelligence, big data analysis, and the Internet of Things in the steel industry is constantly deepening. Utilizing intelligent algorithms (such as optimization algorithms and machine learning algorithms) and automated systems for scrap steel proportioning has become an important development direction for improving the automation level of steel smelting and optimizing production efficiency. Real-time analysis and calculation of scrap steel composition data, smelting process parameters, and production demands through intelligent systems can improve the accuracy and response speed of scrap steel proportioning to a certain extent. However, existing methods often focus only on a single objective (such as minimizing cost or meeting steel quality standards), failing to fully consider the synergistic optimization of multi-dimensional objectives. This results in low accuracy in scrap steel proportioning and makes it difficult to maximize overall production efficiency.

[0004] Therefore, how to achieve intelligent scrap steel ratio optimization with multi-dimensional objectives and improve the accuracy of scrap steel ratio has become an urgent problem to be solved. Summary of the Invention

[0005] To address at least one of the aforementioned technical problems, this invention provides an intelligent scrap steel blending method, equipment, and medium based on multi-agent collaborative decision-making, aiming to achieve intelligent scrap steel blending ratios through multi-dimensional objective collaborative optimization and improve the accuracy of scrap steel blending ratios.

[0006] In a first aspect, the present invention provides an intelligent scrap steel allocation method based on multi-agent collaborative decision-making, the method comprising: The system acquires information on molten iron, target steel grade, scrap steel, equipment, cost, and constraints. The molten iron information includes temperature and composition; the target steel grade information includes type and composition; the scrap steel information includes type, composition, and inventory; the equipment information includes furnace lining condition and the final temperature of the previous furnace; the cost information includes scrap steel price, inventory holding cost, and transportation cost; and the constraint information includes compositional and cost constraints. A smart scrap steel allocation model is constructed based on heat balance intelligent agents, material balance intelligent agents, and cost control intelligent agents. The molten iron information, the target steel grade information, the scrap steel information, the equipment information, the cost information, and the constraint information are input into the intelligent scrap steel matching model. The heat balance intelligent agent, the material balance intelligent agent, and the cost control intelligent agent are used to optimize the heat balance, the material balance, and the composition control, respectively, to determine the target decision scheme, including the batching information, the predicted scrap steel ratio, the predicted blowing temperature, and the predicted total cost.

[0007] In one possible implementation, the step of inputting the molten iron information, the target steel grade information, the scrap steel information, the equipment information, the cost information, and the constraint information into the intelligent scrap steel matching model, and then performing heat balance optimization, material balance optimization, and composition control optimization through the heat balance intelligent agent, the material balance intelligent agent, and the cost control intelligent agent respectively, to determine the target decision scheme, includes: The heat balance intelligent agent optimizes the heat balance based on the molten iron information and the equipment information to determine the upper limit of the scrap steel ratio. The material balance agent optimizes the material balance based on the scrap steel information, the target steel type information, and the composition constraint information, and determines the scrap steel weight information to characterize the weight of each scrap steel type. The cost control agent optimizes cost control based on the cost information and cost constraint information to determine the cost-optimal scrap steel allocation scheme. Based on the upper limit of the scrap steel ratio, the scrap steel weight information, and the cost-optimal scrap steel allocation scheme, the target decision scheme is determined through preset constraints. The preset constraints include any one or more of preset heat constraints, preset composition constraints, preset inventory constraints, and preset cost constraints.

[0008] In one possible implementation, the intelligent scrap steel matching model includes: A multi-source data input module is used to input the molten iron information, the target steel grade information, the scrap steel information, the equipment information, the cost information, and the constraint information; A heat balance intelligent agent, connected to the multi-source data input module, is used to predict the maximum scrap ratio that can be tolerated under the current thermal state based on the molten iron information and the equipment information, and to determine the upper limit of the scrap ratio. The material balance intelligent agent is set up in parallel with the heat balance intelligent agent and connected to the multi-source data input module. It is used to learn the contribution weight distribution of a preset number of different types of scrap steel to the composition of the target steel based on the scrap steel information, the target steel grade information and the composition constraint information, and to determine the scrap steel weight information. The cost control intelligent agent is set up in parallel with the heat balance intelligent agent and the material balance intelligent agent, and is connected to the multi-source data input module. It adopts a deep Q network structure and models the scrap steel ratio optimization as a Markov decision process. It is used to learn the optimal action value function based on the cost information and the cost constraint information, and determine the cost-optimal scrap steel ratio scheme. A multi-agent fusion module, connected to the heat balance agent, the material balance agent, and the cost control agent, is used to obtain a preliminary decision scheme based on the upper limit of scrap steel ratio, the scrap steel weight information, and the cost-optimal scrap steel allocation scheme. A linear programming solution module, connected to the multi-agent fusion module, is used to perform linear programming solutions on the preliminary decision scheme based on preset constraints and preset objectives to obtain candidate decision vectors; The rule post-processing module, connected to the linear programming solution module, is used to modify the candidate decision vector based on a preset modification rule to determine the target decision scheme; The model output module is connected to the rule post-processing module and is used to output the target decision scheme.

[0009] In one possible implementation, the heat balancing agent includes: The first input unit is used to input the molten iron information and the equipment information; A feature splicing unit, connected to the first input unit, is used to splice the molten iron temperature and molten iron composition in the molten iron information to obtain a molten iron feature tensor. The first thermal state feature extraction unit, connected to the feature splicing unit, includes a first convolutional layer and a max pooling layer connected in sequence, used to capture thermal fluctuations based on the molten iron feature tensor and using a first time step size to output the first thermal state feature. The second thermal state feature extraction unit is connected to the first thermal state feature extraction unit and includes a second convolutional layer and a global average pooling layer connected in sequence. It is used to capture the heat accumulation trend based on the first thermal state feature and a second time step size, and output the second thermal state feature. The first time step size is smaller than the second time step size. The heat balance unit, connected to the first input unit and the second thermal state feature extraction unit, is composed of a multilayer sensing mechanism consisting of a three-layer fully connected network. It is used to perform feature fusion and feature mapping on the equipment information and the second thermal state features, and output the upper limit of the scrap steel ratio.

[0010] In one possible implementation, the material balance agent includes a second input unit, a material balance unit, and a feature mapping unit connected in sequence. The second input unit is used to input the scrap steel information, the target steel grade information, and the composition constraint information, and to concatenate the scrap steel inventory, the target steel grade information, and the composition constraint information in the scrap steel information into material static features; The material balance unit adopts a Transformer model with a 4-head self-attention mechanism to encode the scrap steel composition in the scrap steel information together with the static characteristics of the material, and captures the complementary and repulsive relationships of different scrap steel types through a multi-head attention mechanism, and outputs a material balance feature vector. The feature mapping unit is used to map the material balance feature vector to a preset dimension and output the scrap steel weight information through the Softmax function.

[0011] In one possible implementation, the cost control agent includes a third input unit, a value calculation unit, and a strategy optimization unit connected in sequence. The third input unit is used to input the cost information and the cost constraint information, and output the cost state vector after batch normalization processing; The value calculation unit is used to learn the optimal action value function based on the cost state vector and output the expected revenue vector, which is used to represent the expected revenue corresponding to selecting different scrap steel combinations in the current state. The strategy optimization unit is used to output the cost-optimal scrap steel allocation scheme by employing an ε-greedy mechanism and balancing exploration and utilization based on the expected benefit vector.

[0012] In one possible implementation, the multi-agent fusion module includes a feature fusion unit, a weight allocation unit, and a preliminary decision-making unit connected in sequence. The feature fusion unit is used to fuse the upper limit of the scrap steel ratio, the scrap steel weight information and the cost-optimal scrap steel allocation scheme to obtain a fused feature vector. The weight allocation unit is used to dynamically generate weight values ​​through an attention network based on the fused feature vector. The preliminary decision-making unit is used to perform weighted calculations on the fused feature vector based on the weight values ​​to obtain the preliminary decision scheme.

[0013] In one possible implementation, the linear programming solution module includes: A constraint space construction unit is used to construct a constraint space based on each of the preset constraint conditions. The linear programming solution unit, connected to the constraint space construction unit, is used to perform linear programming solutions in the constraint space based on the preset objective using the interior point method to obtain the candidate decision vector.

[0014] In a second aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in the first aspect or any of the implementations thereof.

[0015] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described in the first aspect or any of the implementations thereof.

[0016] Fourthly, the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the steps of the method as described in the first aspect or any of the implementations thereof.

[0017] The advantages of this invention compared to existing technologies are as follows: A multi-dimensional collaborative intelligent scrap steel blending model is constructed based on a heat balance intelligent agent, a material balance intelligent agent, and a cost control intelligent agent. During the model application phase, the acquired information on molten iron, target steel grade, scrap steel, equipment, cost, and constraints is input into the intelligent scrap steel blending model. The heat balance intelligent agent, material balance intelligent agent, and cost control intelligent agent respectively optimize heat balance, material balance, and composition control based on the input information. Through the collaborative computation of these three intelligent agents, the final output includes a target decision scheme containing batching information (scrap steel type and quantity), predicted scrap steel ratio, predicted blowing temperature, and predicted total cost. Compared to traditional scrap steel blending optimization methods that focus only on a single objective, this solution uses a multi-agent collaborative architecture to comprehensively optimize the three core objectives of heat balance, material balance, and cost control. Through multi-dimensional data fusion and intelligent algorithm computation, it achieves intelligent scrap steel blending with multi-dimensional objective collaborative optimization, effectively improving the accuracy of scrap steel blending and thus maximizing overall production efficiency.

[0018] It is understood that the electronic device, computer-readable storage medium, and computer program product provided by the present invention have the same beneficial effects as the above-described intelligent scrap steel allocation method based on multi-agent collaborative decision-making, and will not be repeated here. Attached Figure Description

[0019] To more clearly illustrate the technical solution of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0020] Figure 1This is a flowchart illustrating an intelligent scrap steel allocation method based on multi-agent collaborative decision-making, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an intelligent scrap steel matching model provided in an embodiment of the present invention. Detailed Implementation

[0021] In the following description, specific details such as particular system structures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0022] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0023] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0025] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized. In the modern steel industry production system, steel smelting is a core link, and its production efficiency, product quality, and cost control directly determine the market competitiveness of enterprises. Scrap steel, as one of the key raw materials in the steel smelting process, has a decisive impact on the entire smelting process due to the scientific and rational nature of its proportioning. Specifically, a reasonable scrap steel proportion is not only a core prerequisite for ensuring stable steel composition and meeting mechanical properties to satisfy the diverse needs of downstream industries, but also a key means of achieving heat and material balance during the smelting process. Furthermore, it has significant practical implications for controlling raw material procurement costs and improving resource utilization efficiency. However, in long-term steel production practice, traditional scrap steel blending methods generally rely on the past experience of operators for decision-making. This experience-driven approach has significant limitations: on the one hand, due to the large differences in composition, physical form, and source of different batches of scrap steel, it is difficult for manual calculation of the optimal mixing ratio of various types of scrap steel based solely on experience, which can easily lead to blending deviations, thereby causing fluctuations in molten steel quality and increasing the probability of producing substandard products; on the other hand, the manual blending process requires multiple steps such as data collection, experience judgment, and scheme adjustment, resulting in low overall efficiency. Especially when facing complex and ever-changing production conditions such as frequent changes in smelting orders and unstable raw material supply, it is difficult to respond quickly to production needs, which can easily cause production rhythm disruptions and affect the continuity and stability of the smelting process. With the rapid development of industrial intelligent technology, the application of technologies such as artificial intelligence, big data analysis, and the Internet of Things in the steel industry is constantly deepening. Utilizing intelligent algorithms and automated systems to replace manual scrap steel proportioning has become an important development direction for improving the automation level of steel smelting and optimizing production efficiency. By using intelligent systems to analyze and calculate scrap steel composition data, smelting process parameters, production demands, and other information in real time, the accuracy and response speed of scrap steel proportioning can be improved to a certain extent, alleviating the shortcomings of traditional manual methods. However, existing intelligent scrap steel blending methods still have significant shortcomings. The most important one is that existing methods often focus only on a single objective (such as the lowest cost or meeting steel quality standards), and fail to fully consider the synergistic optimization of multiple dimensions such as scrap steel blending ratio, heat balance, material utilization rate, and environmental protection indicators during the smelting process. This makes it difficult for the blending scheme to maximize overall production efficiency.

[0027] To address the aforementioned technical issues, this invention proposes an intelligent scrap steel allocation method based on multi-agent collaborative decision-making. Through a multi-agent collaborative architecture, it achieves comprehensive optimization of three core objectives: heat balance, material balance, and cost control. By integrating multi-dimensional data and employing intelligent algorithms, it effectively improves the accuracy of scrap steel allocation. For ease of understanding, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Figure 1 This is a flowchart illustrating an intelligent scrap steel allocation method based on multi-agent collaborative decision-making according to an embodiment of the present invention. For ease of explanation, only the parts relevant to this embodiment are shown. The method provided in this embodiment includes the following steps: S1 acquires information on molten iron, target steel grade, scrap steel, equipment, cost, and constraints. Molten iron information includes molten iron temperature and composition; target steel grade information includes the type and composition of the target steel grade; scrap steel information includes scrap steel type, scrap steel composition, and scrap steel inventory; equipment information includes furnace lining condition and the final temperature of the previous furnace; cost information includes scrap steel price, scrap steel inventory holding cost, and transportation cost; and constraint information includes composition constraints and cost constraints.

[0029] Optionally, the molten iron information includes, but is not limited to, molten iron temperature, molten iron weight, molten iron composition, and the content of each component; the main components of molten iron include elements such as silicon (Si), manganese (Mn), phosphorus (P), and sulfur (S). The purpose of the molten iron information is to serve as a calculation basis for the main heat source and the source of metal materials.

[0030] Optionally, the target steel grade is the steel grade planned for output. Target steel grade information includes, but is not limited to, the type, composition, target weight, and target temperature of the target steel grade. This information is used to clarify the final target of the batching. Specifically, the target temperature and the content of elements such as carbon, phosphorus, and sulfur are related to the type of target steel grade. Once the type of target steel grade is determined, information such as composition and target temperature can be identified.

[0031] Optionally, scrap information includes, but is not limited to, scrap type, scrap composition, scrap temperature, and scrap inventory. Scrap types include heavy scrap, medium scrap, light scrap, briquetted scrap, pig iron, iron ore, and sintered ore. Scrap information serves as a decision variable, and each agent in the subsequent intelligent scrap allocation model needs to be optimized under the constraints of this information.

[0032] Optionally, equipment information includes, but is not limited to, furnace lining condition (furnace age), empty furnace time, and the end temperature of the previous furnace (hot state). Equipment information affects the heat balance; for example, a new furnace lining absorbs more heat, while a hot furnace accumulates more heat.

[0033] Optional cost information includes, but is not limited to, scrap steel prices, scrap steel inventory holding costs, and transportation costs.

[0034] Optionally, the constraint information mainly includes composition constraint information and cost constraint information. The composition constraint information can be represented as a three-dimensional vector of carbon, phosphorus, and sulfur content, used by the subsequent material balance agent to perform material balance optimization based on this composition constraint information. The cost constraint information can be the ideal budget value or the maximum budget value, used by the subsequent cost control agent to perform cost control optimization based on this cost constraint information.

[0035] In addition, coolant information, including but not limited to coolant type, composition, inventory and cost, can be obtained as constraint information for subsequent optimization.

[0036] S2, based on heat balance intelligent agent, material balance intelligent agent and cost control intelligent agent, constructs intelligent scrap steel allocation model.

[0037] In one possible implementation, such as Figure 2 As shown, the intelligent scrap steel matching model includes: The multi-source data input module is used to input information such as molten iron, target steel grade, scrap steel, equipment, cost, and constraints. The heat balance intelligent agent, connected to the multi-source data input module, is used to predict the maximum scrap ratio that can be tolerated under the current thermal state based on molten iron information and equipment information, and to determine the upper limit of the scrap ratio; The material balance intelligent agent is set up in parallel with the heat balance intelligent agent and connected to the multi-source data input module. It is used to learn the contribution weight distribution of a preset number of different types of scrap steel to the composition of the target steel based on scrap steel information, target steel grade information and composition constraint information, and determine the scrap steel weight information. The cost control intelligent agent is set up in parallel with the heat balance intelligent agent and the material balance intelligent agent, and is connected to the multi-source data input module. It adopts a deep Q network structure and models the scrap steel ratio optimization as a Markov decision process. It is used to learn the optimal action value function based on cost information and cost constraint information, and determine the cost-optimal scrap steel ratio scheme. The multi-agent fusion module connects with the heat balance agent, the material balance agent, and the cost control agent to obtain a preliminary decision scheme based on the upper limit of scrap steel ratio, scrap steel weight information, and the cost-optimal scrap steel allocation scheme. The rule post-processing module, connected to the linear programming solution module, is used to modify the candidate decision vectors based on preset correction rules to determine the target decision scheme; The model output module, connected to the rule post-processing module, is used to output the target decision scheme.

[0038] Preferably, the multi-source data input module includes a heat balance input unit, a material balance input unit, and a cost control input unit arranged in parallel; the heat balance input unit is used to input molten iron information and equipment information, the material balance input unit is used to input scrap steel information, target steel grade information, and composition constraint information, and the cost control input unit is used to input cost information and cost constraint information.

[0039] As an example, the heat balance input unit is used to input two static scalars: a 5×20-dimensional time series matrix consisting of molten iron temperature and composition, furnace lining status (furnace age percentage), and the end temperature of the previous furnace. The material balance input unit is used to input the scrap steel composition matrix (8 types of scrap steel × 5 main elements), the 20-dimensional One-Hot code of the target steel grade, the available inventory (8-dimensional vector) corresponding to each of the 8 types of scrap steel, and composition constraint information. The composition constraint information is preferably a composition constraint vector of carbon, phosphorus, and sulfur [C,P,S]. The cost control input unit is used to input the 8-dimensional scrap steel price, 8-dimensional scrap steel inventory holding cost, 8-dimensional transportation cost, and 1-dimensional cost constraint scalar, corresponding to each of the 8 types of scrap steel.

[0040] Preferably, the heat balance agent includes: The first input unit is used to input molten iron information and equipment information; The feature splicing unit, connected to the first input unit, is used to splice the molten iron temperature and molten iron composition in the molten iron information to obtain the molten iron feature tensor. The first thermal state feature extraction unit, connected to the feature splicing unit, includes a first convolutional layer and a max pooling layer connected in sequence. It is used to capture thermal fluctuations based on the molten iron feature tensor and the first time step size, and output the first thermal state features. The second thermal state feature extraction unit is connected to the first thermal state feature extraction unit. It includes a second convolutional layer and a global average pooling layer connected in sequence. It is used to capture the thermal accumulation trend based on the first thermal state feature and a second time step size, and output the second thermal state feature. The first time step size is smaller than the second time step size. The heat balance unit, connected to the first input unit and the second thermal state feature extraction unit, is composed of a multi-layer sensing mechanism consisting of a three-layer fully connected network. It is used to perform feature fusion and feature mapping on equipment information and second thermal state features, and output the upper limit of scrap steel ratio.

[0041] As an example, the heat balance unit includes a first fully connected layer, a first dropout layer, a second fully connected layer, a second dropout layer, a third fully connected layer, and a first output layer connected in sequence.

[0042] For example, the heat balance agent employs a hybrid architecture of a one-dimensional convolutional neural network (1D-CNN) and a multilayer perceptron (MLP) to predict the maximum scrap ratio that can be tolerated under the current thermal state. The first input unit receives four types of data: a 5×20-dimensional time-series matrix composed of molten iron temperature and composition, furnace lining state (furnace age percentage), and two static scalars: the end temperature of the previous furnace. The feature concatenation unit concatenates the 5×20-dimensional time-series matrix composed of molten iron temperature and composition into a 120-dimensional vector and reshapes it into a (12, 10) molten iron feature tensor. Subsequently, feature patterns are extracted through two layers of one-dimensional convolutional networks. The first convolutional layer in the first thermal state feature extraction unit includes 64 filters to capture short-term thermal fluctuations, and the second convolutional layer in the second thermal state feature extraction unit includes 128 filters to capture long-range thermal accumulation trends. After passing through a global average pooling layer, a 128-dimensional thermal state code is obtained, which is the second thermal state feature. The furnace lining condition (percentage of furnace age) and the end temperature of the previous furnace are two static scalars, which are concatenated with the second thermal state feature in the heat balance unit to form a 130-dimensional comprehensive feature. The first, second, and third fully connected layers reduce the dimensionality of the 130-dimensional comprehensive feature to 64, 32, and 16 dimensions, respectively. Finally, the upper limit η of the scrap ratio in the range [0, 0.35] is generated by the Sigmoid function in the first output layer, representing the maximum scrap ratio under the heat constraint. At the same time, the heat balance unit uses L2 regularization and Dropout to prevent overfitting, and the first and second Dropout layers ensure robust modeling of dynamic changes in molten iron temperature.

[0043] Preferably, the material balance intelligent agent includes a second input unit, a material balance unit, and a feature mapping unit connected in sequence: The second input unit is used to input scrap steel information, target steel grade information and composition constraint information, and to concatenate the scrap steel inventory, target steel grade information and composition constraint information in the scrap steel information into material static characteristics; The material balance unit uses a Transformer model with a 4-head self-attention mechanism to encode the scrap steel composition and static material characteristics in the scrap steel information. Then, it captures the complementary and repulsive relationships between different types of scrap steel through a multi-head attention mechanism and outputs a material balance feature vector. The feature mapping unit is used to map the material balance feature vector to a preset dimension and output the scrap weight information through the Softmax function.

[0044] As an example, the second input unit includes an embedding layer and a first splicing layer connected in sequence; the material balance unit includes a multi-head attention layer, a feedforward network layer and a residual connection layer connected in sequence; and the feature mapping unit includes a fourth fully connected layer, a fifth fully connected layer and a second output layer connected in sequence.

[0045] For example, the material balance agent is based on a Transformer encoder architecture to learn the contribution weight distribution of eight preset scrap steel types to the composition requirements of the target steel grade. The second input unit takes into account the scrap steel composition matrix (8 types of scrap steel × 5 elements), the 20-dimensional One-Hot code of the target steel grade, the 8-dimensional vector of available inventory, and the 3-dimensional vector of composition constraint information [C,P,R]. The 20-dimensional One-Hot code of the target steel grade is mapped to a 32-dimensional dense vector through an embedding layer, and then concatenated with the 8-dimensional vector of available inventory and the 3-dimensional vector of composition constraint information [C,P,R] through a first concatenation layer to form a 43-dimensional static material feature. The material balancing unit employs a Transformer model with a 4-head self-attention mechanism, jointly encoding the scrap steel composition matrix and the static features of the materials. A multi-head attention layer captures the complementary and repulsive relationships between different scrap steel types (e.g., high-phosphorus scrap steel needs to be balanced with low-phosphorus scrap steel). The feature vector dimension is first increased from 64 to 128 dimensions via a feedforward network layer, then reduced back to 64 dimensions, and finally output as a 64-dimensional material balancing feature vector via a residual connection layer. In the feature mapping unit, the material balancing feature vector is reduced from 64 dimensions to 32 dimensions via a fourth fully connected layer, and further reduced to 16 dimensions via a fifth fully connected layer. The softmax function in the second output layer then generates scrap steel weight information, i.e., an 8-dimensional probability vector α, representing the weight allocation of 8 preset scrap steel types under compositional constraints.

[0046] Preferably, the cost control agent includes a third input unit, a value calculation unit, and a strategy optimization unit connected in sequence: The third input unit is used to input cost information and cost constraint information, and after batch normalization, outputs a cost state vector. The value calculation unit is used to learn the optimal action value function based on the cost state vector and output the expected revenue vector, which is used to represent the expected revenue corresponding to choosing different scrap steel combinations in the current state. The strategy optimization unit is used to output the cost-optimal scrap steel allocation scheme by balancing exploration and utilization based on the expected benefit vector using the ε-greedy mechanism.

[0047] As an example, the third input unit includes a second splicing layer and a batch normalization layer connected in sequence; the value calculation unit includes a sixth fully connected layer, a third Dropout layer, a seventh fully connected layer, a fourth Dropout layer, an eighth fully connected layer, and a third output layer connected in sequence.

[0048] As an example, the cost control agent employs a Deep Q-Network (DQN) structure to model scrap steel ratio optimization as a Markov decision process, learning the optimal action value function in the cost and inventory state space. The third input unit takes a 33-dimensional state vector as input, containing 8-dimensional scrap steel prices, 8-dimensional inventory holding costs, 8-dimensional transportation costs, and a 1-dimensional cost constraint scalar corresponding to 8 preset types of scrap steel. After state concatenation in the second concatenation layer and normalization in the batch normalization layer, the output cost state vector is sent to the value calculation unit. The value calculation unit consists of a network composed of three fully connected layers. The sixth, seventh, and eighth fully connected layers map the cost state vector from 33 dimensions to 64, 32, and 16 dimensions respectively. A Dropout layer (preferably with a dropout rate of 0.2) is inserted between every two fully connected layers to enhance generalization ability. Finally, an 8-dimensional expected revenue vector is output, representing the expected revenue of selecting each type of scrap steel action in the current state. The strategy optimization unit adopts the ε-greedy mechanism to balance exploration and utilization, and outputs the cost-optimal scrap steel allocation scheme to participate in subsequent multi-objective fusion.

[0049] The cost control agent shown in this example directly optimizes cost targets through reinforcement learning, and can dynamically adapt to fluctuations in scrap steel prices and changes in inventory to achieve the most economically optimal decision-making.

[0050] Preferably, the multi-agent fusion module includes a feature fusion unit, a weight allocation unit, and a preliminary decision-making unit connected in sequence: The feature fusion unit is used to fuse the upper limit of scrap steel ratio, scrap steel weight information and cost-optimal scrap steel allocation scheme to obtain a fused feature vector. The weight allocation unit is used to dynamically generate weight values ​​based on the fused feature vector through an attention network. The preliminary decision-making unit is used to perform weighted calculations on the fused feature vector based on the weight values ​​to obtain a preliminary decision scheme.

[0051] As an example, in the multi-agent fusion module, the feature fusion of the three agents is achieved through a hierarchical and progressive mechanism. First, the upper limit η of the scrap steel ratio output by the heat balance agent, the scrap steel weight information α output by the material balance agent, and the cost-optimal scrap steel allocation scheme Q output by the cost control agent are concatenated into a 17-dimensional vector in the feature fusion unit. The weights β1, β2, and β3 are dynamically generated by the weight allocation unit and then weighted by the preliminary decision unit according to d=β1×η×α+β2×α+β3×softmax(Q) to form a preliminary decision scheme that takes into account heat, composition, and cost.

[0052] Preferably, the linear programming solution module includes: Constraint space construction unit, used to construct constraint space based on various preset constraints; The linear programming solution unit, connected to the constraint space construction unit, is used to perform linear programming solutions in the constraint space based on a preset objective using the interior point method, and to obtain candidate decision vectors.

[0053] As an example, if the preset objective is cost minimization, then by constructing a constraint space, the preset heat constraints, preset composition constraints, preset inventory constraints, and preset cost constraints are transformed into a standard linear programming form, thus constructing the constraint space. In the linear programming solution unit, the linear programming solver uses the interior-point method to solve for the candidate decision vector W∈ℝ that minimizes the cost within the aforementioned constraint space. 8 .

[0054] For example, for each batch of scrap steel i, let W_i represent its added weight, H_i represent the unit heat, C be the carbon content, C_min be the preset minimum carbon content, C_max be the preset maximum carbon content, P be the phosphorus content, D_i and s_i be the lower and upper limits of inventory respectively, P_i be the unit price, B be the total budget, and T be the target temperature of the target steel grade; then, the preset heat constraint is ∑(W_i×H_i)≥T, the preset composition constraint is [C]∈[C_min,C_max] and P≤0.035, the preset inventory constraint is D_i≤W_i≤s_i, and the preset cost constraint is ∑(W_i×P_i)≤B.

[0055] As an example, the preset correction rules used in the rule post-processing module to correct candidate decision vectors include proportional reduction of excess inventory, addition of pig iron blocks when the blowing temperature is insufficient, reduction of scrap steel ratio by 2% for new furnace linings, and calling historical scrap steel allocation schemes under the same working conditions when there are no candidate decision vectors.

[0056] The rule post-processing module transforms mathematical solutions into executable recipes. It skips the correction step only when the model has no feasible solution, and directly calls the historical scrap steel matching scheme under the same working conditions, thus ensuring the reliability of the intelligent scrap steel matching scheme.

[0057] As an example, the model output module is used to output target decision schemes including scrap steel type and weight, predicted scrap steel ratio, predicted blow-up temperature, predicted component hit rate and predicted total cost, etc., realizing end-to-end collaboration from multi-source heterogeneous input to executable optimization decision.

[0058] S3 inputs information on molten iron, target steel grade, scrap steel, equipment, cost, and constraints into the intelligent scrap steel blending model. Through the heat balance agent, material balance agent, and cost control agent, heat balance optimization, material balance optimization, and composition control optimization are performed to determine the target decision scheme, including batching information, predicted scrap steel ratio, predicted blowing temperature, and predicted total cost.

[0059] In one possible implementation, step S3 may include: By using a heat balance intelligent agent, based on molten iron information and equipment information, heat balance optimization is performed to determine the upper limit of scrap steel ratio; The material balance agent optimizes the material balance based on scrap steel information, target steel type information and composition constraint information, and determines the scrap steel weight information to characterize the weight of each type of scrap steel. By using a cost control intelligent agent, cost control optimization is performed based on cost information and cost constraint information to determine the cost-optimal scrap steel allocation scheme; Based on the upper limit of scrap steel ratio, scrap steel weight information, and the cost-optimal scrap steel allocation scheme, the target decision scheme is determined by pre-set constraints. The pre-set constraints include any one or more of the following: pre-set heat constraints, pre-set composition constraints, pre-set inventory constraints, and pre-set cost constraints.

[0060] As an example, the heat balance agent, relying on information about molten iron (temperature, composition, etc.), calculates the amount of scrap steel to be added, ensuring that the tapping temperature matches the temperature of the target steel grade, providing a suitable thermal state for converter smelting, and achieving heat balance optimization.

[0061] As an example, the material balance agent combines the elemental requirements of the target steel grade with the current composition of the molten iron, adjusts the amount of scrap steel added, and achieves material balance optimization by matching the composition requirements of the target steel grade through the "dilution effect" or component supplementation of scrap steel.

[0062] As an example, the cost control agent integrates molten iron information with target steel grade information, optimizes the scrap steel ratio structure, prioritizes the use of lower-cost scrap steel combinations, reduces total production costs, and achieves cost control optimization.

[0063] Optionally, through multi-agent collaborative optimization and further constraints based on preset conditions, batching information is generated, clarifying the specific type and weight of each type of scrap steel, which can directly guide the operation of the overhead crane and silo; composition prediction results are obtained, which may include prediction of total loading, scrap steel ratio, and coolant usage; heat balance prediction results are obtained, which may include prediction of starting temperature and required chemical heat, ensuring that the temperature is within a reasonable range at the start of blowing; the total cost of raw materials for this heat is estimated, providing a basis for economic optimization and obtaining the final target decision scheme.

[0064] The technical solution provided by this invention constructs a multi-dimensional collaborative intelligent scrap steel blending model based on a heat balance intelligent agent, a material balance intelligent agent, and a cost control intelligent agent. In the model application stage, the acquired information on molten iron, target steel grade, scrap steel, equipment, cost, and constraints is input into the intelligent scrap steel blending model. The heat balance intelligent agent, material balance intelligent agent, and cost control intelligent agent respectively optimize heat balance, material balance, and composition control based on the input information. Through the collaborative calculation of the three intelligent agents, the final output is a target decision scheme that includes batching information (scrap steel type and quantity), predicted scrap steel ratio, predicted blowing temperature, and predicted total cost. Compared to traditional scrap steel blending optimization methods that focus only on a single objective, this solution uses a multi-agent collaborative architecture to comprehensively optimize the three core objectives of heat balance, material balance, and cost control. Through multi-dimensional data fusion and intelligent algorithm calculation, it achieves intelligent scrap steel blending with multi-dimensional objective collaborative optimization, effectively improving the accuracy of scrap steel blending and thus maximizing overall production efficiency.

[0065] In summary, the improvements of the technical solution provided by this invention are mainly reflected in the following aspects: 1. Precise proportioning: Through multi-agent collaborative decision-making and multi-dimensional optimization, the precise proportioning of scrap steel is achieved, thereby improving the quality of molten steel and production stability.

[0066] 2. Multi-dimensional optimization: Optimize from three dimensions: heat, materials, and cost, to reduce production costs and improve economic efficiency while ensuring the quality of molten steel.

[0067] 3. Intelligent decision-making: Utilizing intelligent algorithms and systems for decision-making reduces human intervention and improves the scientific nature and accuracy of decisions. On the other hand, the present invention also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above-mentioned intelligent scrap steel allocation methods based on multi-agent collaborative decision-making.

[0068] On the other hand, the present invention also provides an electronic device, including a memory and a processor; the memory stores program code executable by the processor; the program code is used to execute any of the above-mentioned intelligent scrap steel allocation methods based on multi-agent collaborative decision-making.

[0069] For example, the program code can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the program code in the electronic device.

[0070] The electronic device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The electronic device may include, but is not limited to, processors and memory. Those skilled in the art will understand that the electronic device may also include input / output devices, network access devices, buses, etc.

[0071] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0072] The memory can be an internal storage unit of the electronic device, such as a hard drive or RAM. It can also be an external storage device, such as a plug-in hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units. The memory is used to store the program code and other programs and data required by the electronic device. The memory can also be used to temporarily store data that has been output or will be output.

[0073] The computer storage medium and electronic device described above are created based on the above method. Their technical functions and beneficial effects will not be elaborated here. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0074] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for intelligent scrap steel allocation based on multi-agent collaborative decision-making, characterized in that, The method includes: The system acquires information on molten iron, target steel grade, scrap steel, equipment, cost, and constraints. The molten iron information includes temperature and composition; the target steel grade information includes type and composition; the scrap steel information includes type, composition, and inventory; the equipment information includes furnace lining condition and the final temperature of the previous furnace; the cost information includes scrap steel price, inventory holding cost, and transportation cost; and the constraint information includes compositional and cost constraints. A smart scrap steel allocation model is constructed based on heat balance intelligent agents, material balance intelligent agents, and cost control intelligent agents. The molten iron information, the target steel grade information, the scrap steel information, the equipment information, the cost information, and the constraint information are input into the intelligent scrap steel matching model. The heat balance intelligent agent, the material balance intelligent agent, and the cost control intelligent agent are used to optimize the heat balance, the material balance, and the composition control, respectively, to determine the target decision scheme, including the batching information, the predicted scrap steel ratio, the predicted blowing temperature, and the predicted total cost.

2. The method according to claim 1, characterized in that, The process involves inputting the molten iron information, the target steel grade information, the scrap steel information, the equipment information, the cost information, and the constraint information into the intelligent scrap steel matching model. Heat balance optimization, material balance optimization, and composition control optimization are then performed through the heat balance intelligent agent, the material balance intelligent agent, and the cost control intelligent agent, respectively, to determine the target decision scheme, including: The heat balance intelligent agent optimizes the heat balance based on the molten iron information and the equipment information to determine the upper limit of the scrap steel ratio. The material balance agent optimizes the material balance based on the scrap steel information, the target steel type information, and the composition constraint information, and determines the scrap steel weight information to characterize the weight of each scrap steel type. The cost control agent optimizes cost control based on the cost information and cost constraint information to determine the cost-optimal scrap steel allocation scheme. Based on the upper limit of the scrap steel ratio, the scrap steel weight information, and the cost-optimal scrap steel allocation scheme, the target decision scheme is determined through preset constraints. The preset constraints include any one or more of preset heat constraints, preset composition constraints, preset inventory constraints, and preset cost constraints.

3. The method according to claim 1, characterized in that, The intelligent scrap steel matching model includes: A multi-source data input module is used to input the molten iron information, the target steel grade information, the scrap steel information, the equipment information, the cost information, and the constraint information; A heat balance intelligent agent, connected to the multi-source data input module, is used to predict the maximum scrap ratio that can be tolerated under the current thermal state based on the molten iron information and the equipment information, and to determine the upper limit of the scrap ratio. The material balance intelligent agent is set up in parallel with the heat balance intelligent agent and connected to the multi-source data input module. It is used to learn the contribution weight distribution of a preset number of different types of scrap steel to the composition of the target steel based on the scrap steel information, the target steel grade information and the composition constraint information, and to determine the scrap steel weight information. The cost control intelligent agent is set up in parallel with the heat balance intelligent agent and the material balance intelligent agent, and is connected to the multi-source data input module. It adopts a deep Q network structure and models the scrap steel ratio optimization as a Markov decision process. It is used to learn the optimal action value function based on the cost information and the cost constraint information, and determine the cost-optimal scrap steel ratio scheme. A multi-agent fusion module, connected to the heat balance agent, the material balance agent, and the cost control agent, is used to obtain a preliminary decision scheme based on the upper limit of scrap steel ratio, the scrap steel weight information, and the cost-optimal scrap steel allocation scheme. A linear programming solution module, connected to the multi-agent fusion module, is used to perform linear programming solutions on the preliminary decision scheme based on preset constraints and preset objectives to obtain candidate decision vectors; The rule post-processing module, connected to the linear programming solution module, is used to modify the candidate decision vector based on a preset modification rule to determine the target decision scheme; The model output module is connected to the rule post-processing module and is used to output the target decision scheme.

4. The method according to claim 3, characterized in that, The heat balance intelligent agent includes: The first input unit is used to input the molten iron information and the equipment information; A feature splicing unit, connected to the first input unit, is used to splice the molten iron temperature and molten iron composition in the molten iron information to obtain a molten iron feature tensor. The first thermal state feature extraction unit, connected to the feature splicing unit, includes a first convolutional layer and a max pooling layer connected in sequence, used to capture thermal fluctuations based on the molten iron feature tensor and using a first time step size to output the first thermal state feature. The second thermal state feature extraction unit is connected to the first thermal state feature extraction unit and includes a second convolutional layer and a global average pooling layer connected in sequence. It is used to capture the heat accumulation trend based on the first thermal state feature and a second time step size, and output the second thermal state feature. The first time step size is smaller than the second time step size. The heat balance unit, connected to the first input unit and the second thermal state feature extraction unit, is composed of a multilayer sensing mechanism consisting of a three-layer fully connected network. It is used to perform feature fusion and feature mapping on the equipment information and the second thermal state features, and output the upper limit of the scrap steel ratio.

5. The method according to claim 3, characterized in that, The material balance intelligent agent includes a second input unit, a material balance unit, and a feature mapping unit connected in sequence: The second input unit is used to input the scrap steel information, the target steel grade information, and the composition constraint information, and to concatenate the scrap steel inventory, the target steel grade information, and the composition constraint information in the scrap steel information into material static features; The material balance unit adopts a Transformer model with a 4-head self-attention mechanism to encode the scrap steel composition in the scrap steel information together with the static characteristics of the material, and captures the complementary and repulsive relationships of different scrap steel types through a multi-head attention mechanism, and outputs a material balance feature vector. The feature mapping unit is used to map the material balance feature vector to a preset dimension and output the scrap steel weight information through the Softmax function.

6. The method according to claim 3, characterized in that, The cost control intelligent agent includes a third input unit, a value calculation unit, and a strategy optimization unit connected in sequence: The third input unit is used to input the cost information and the cost constraint information, and output the cost state vector after batch normalization processing; The value calculation unit is used to learn the optimal action value function based on the cost state vector and output the expected revenue vector, which is used to represent the expected revenue corresponding to selecting different scrap steel combinations in the current state. The strategy optimization unit is used to output the cost-optimal scrap steel allocation scheme by employing an ε-greedy mechanism and balancing exploration and utilization based on the expected benefit vector.

7. The method according to claim 3, characterized in that, The multi-agent fusion module includes a feature fusion unit, a weight allocation unit, and a preliminary decision-making unit connected in sequence. The feature fusion unit is used to fuse the upper limit of the scrap steel ratio, the scrap steel weight information and the cost-optimal scrap steel allocation scheme to obtain a fused feature vector. The weight allocation unit is used to dynamically generate weight values ​​through an attention network based on the fused feature vector. The preliminary decision-making unit is used to perform weighted calculations on the fused feature vector based on the weight values ​​to obtain the preliminary decision scheme.

8. The method according to claim 3, characterized in that, The linear programming solution module includes: A constraint space construction unit is used to construct a constraint space based on each of the preset constraint conditions. The linear programming solution unit, connected to the constraint space construction unit, is used to perform linear programming solutions in the constraint space based on the preset objective using the interior point method to obtain the candidate decision vector.

9. An electronic device comprising 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 method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images