A hot rolling intelligent scheduling method and system of size model cooperation

Abstract

The present application belongs to the technical field of hot rolling scheduling, and relates to a hot rolling intelligent scheduling method and system based on large and small models, which comprises the following steps: based on a hot rolling scheduling large model, analyzing hot rolling scheduling related requirements input by a user and generating structured scheduling parameters and configuration rules; based on a hot rolling scheduling small model, processing the structured scheduling parameters and configuration rules and generating a slab rolling plan; and based on the hot rolling scheduling large model, comprehensively evaluating the slab rolling plan and generating a comprehensive evaluation result. The present application effectively solves the technical difficulties in the hot rolling scheduling scene, such as difficulty in digital modeling of experience rules, various requirements, high requirement for dynamic response, complex multi-objective optimization, and large solution space, and achieves the scheduling goal of maximizing production efficiency, reducing cost, ensuring quality, and timely delivery.

Description

Technical Field

[0001] This invention relates to the field of hot rolling scheduling technology, and in particular to a method and system for intelligent hot rolling scheduling that combines large and small models. Background Technology

[0002] Hot rolling scheduling is a crucial link in steel production planning. The effectiveness of hot rolling scheduling directly affects core performance indicators of production operations, such as the smoothness of logistics connections between upstream and downstream processes, product quality, rolling line capacity, equipment utilization rate, and equipment lifespan. The quality of scheduling is directly related to the production management efficiency and core competitiveness of steel enterprises. With increasingly diversified, smaller-batch, and customized market orders, the production model is shifting from bulk production to multi-variety, small-batch production, significantly increasing scheduling complexity. At the same time, dynamic disturbances in the production environment occur frequently, such as equipment failures, frequent order changes, abnormal slab quality, and sudden process events, placing extremely high demands on the scheduling system's rapid response and dynamic adjustment capabilities.

[0003] Currently, most steel companies still rely on manual scheduling by planners in the hot rolling production planning stage. Because the planning process requires handling multiple tasks simultaneously, including coordinating diverse production plans and casting schemes, organizing and classifying large quantities of slabs, and meeting complex process constraints and dynamic changes in upstream and downstream capacity, planners face a heavy workload. Furthermore, differences in scheduling experience among planners lead to inconsistent scheduling results, affecting the stability of production indicators and hindering the standardization and refined management of the production process. Although some research has attempted to achieve automatic hot rolling scheduling using heuristic algorithms, these methods often struggle to accurately translate practical experience into mathematical models. Coupled with problems such as excessively large solution spaces and insufficient adaptability to complex constraints and real-time changes, they can easily lead to low resource utilization, increased inventory, higher costs, and even affect the timely delivery of critical orders. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to provide a method and system for intelligent scheduling of hot rolling with collaborative large and small models, in order to solve the problems of existing technologies such as reliance on manual labor, difficulty in digital modeling, diverse requirements, inability to achieve dynamic response, complexity of multi-objective optimization, and large solution space.

[0005] On one hand, embodiments of the present invention provide a hot rolling intelligent scheduling method with coordinated large and small models, including:

[0006] Based on a pre-built large model of hot rolling scheduling, the user-input hot rolling scheduling-related requirements are analyzed, and structured scheduling parameters and configuration rules are generated. Based on a pre-built hot rolling scheduling model, the structured scheduling parameters and configuration rules are processed to generate a slab rolling plan. Based on the aforementioned hot rolling scheduling model, a comprehensive evaluation of the slab rolling plan is performed, and a comprehensive evaluation result is generated.

[0007] Furthermore, based on the pre-built large-scale hot rolling scheduling model, the user-input hot rolling scheduling-related requirements are analyzed, and structured scheduling parameters and configuration rules are generated, including: Based on the aforementioned hot rolling scheduling model, the user-input hot rolling scheduling related requirements are analyzed; Based on the results of user needs analysis, guide users to input or upload the corresponding scheduling data and rule files; Based on the user requirements analysis results, scheduling data, and rule files, structured scheduling parameters and configuration rules are generated.

[0008] Furthermore, the training process of the large-scale hot rolling scheduling model includes: Incremental pre-training of a general-purpose base model was performed using a professional knowledge base in the field of metallurgy to generate a basic large model with metallurgical knowledge. Construct a hybrid dataset and divide the hybrid dataset into a training dataset and a test dataset; the hybrid dataset contains at least parameter generation data, rule generation data and rolling mill evaluation data, and each type of data contains task instructions and their corresponding standard answer results; Based on multiple preset hyperparameters and combined with the training dataset, the basic large model is fine-tuned and trained to obtain multiple candidate scheduling large models. Based on the test dataset, each of the candidate scheduling models was tested, and the candidate scheduling model with the highest average score was selected as the final hot rolling scheduling model, in conjunction with expert scoring.

[0009] Furthermore, the process of processing the structured scheduling parameters and configuration rules based on the pre-built hot rolling scheduling small model to generate a slab rolling plan includes: Based on the slab properties, the scheduling parameters are clustered to obtain multiple candidate slab groups containing parameter features; Based on the configuration rules, determine the rule constraints for the preset comprehensive objective optimization function; The slab rolling plan is generated based on the multiple candidate slab groups, the preset comprehensive objective optimization function, and the rule constraints.

[0010] Further, generating the slab rolling plan based on the multiple candidate slab groups, the preset comprehensive objective optimization function, and rule constraints includes: Based on the breadth-first search algorithm, the current optimal solution B, the current highest score, and the solution queue Q to be expanded are initialized; empty rolling plans and slab groups corresponding to special slab types in the user demand analysis results are used as initial solutions and placed into the solution queue Q; The solution queue Q is searched cyclically. During the cyclic search, the current solution S taken from the solution queue Q is expanded according to the multiple candidate slab groups and rule constraints. The current solution S in the solution queue Q that cannot be further expanded and has no state change is scored according to the preset comprehensive objective optimization function. The scoring result is compared with the current highest score, and the current optimal solution B and the current highest score are updated based on the comparison result to perform the next round of search; until the solution queue Q is empty, the current optimal solution B with the highest score is taken as the slab rolling plan.

[0011] Further, the step of expanding the current solution S taken from the solution queue Q according to the multiple candidate slab groups and rule constraints, and scoring the current solution S in the solution queue Q that cannot be further expanded and has no state change according to the preset comprehensive objective optimization function, includes: Obtain the unused slab groups of the current solution S from the plurality of candidate slab groups, and use them as the available slab group set U of the current solution S; Determine whether the set of available slab groups U is empty, where: When the set of available slab groups U is not empty, according to the rule constraints, it is determined one by one whether the current slab group G in the set of available slab groups U can be included in the current solution S; if it can be included, the current slab group G is included in the current solution S and a new solution S is formed. The new solution S is not included in the solution queue Q. At that time, the new solution S Place it into the solution queue Q, and then check again whether the set of available slab groups U is empty; When the available slab set U is empty, determine whether the state of the current solution S has changed after being taken out of the solution queue Q: if it has changed, put the changed current solution S into the solution queue Q and continue to perform the loop search; otherwise, calculate the score of the current solution S based on the preset comprehensive objective optimization function.

[0012] Furthermore, the preset comprehensive objective optimization function is constructed in the following manner: Set multiple KPI production indicators to be optimized, and configure corresponding weight coefficients for each KPI production indicator; By linearly weighting and summing the objective functions of each KPI production indicator with its corresponding weight coefficients, multiple KPI production indicators are aggregated into a comprehensive objective optimization function for a single objective problem.

[0013] Furthermore, the method also includes: Based on the collaborative execution engine of the large and small models and the results of user requirement analysis, model scheduling instructions are generated to realize the collaborative cooperation and functional scheduling between the small hot rolling scheduling model and the large hot rolling scheduling model.

[0014] On the other hand, embodiments of the present invention provide a hot rolling intelligent scheduling system with coordinated large and small models, including: The conversation analysis module is used to analyze user-input hot rolling scheduling-related requirements based on a pre-built large model of hot rolling scheduling, and generate structured scheduling parameters and configuration rules. The plan generation module is used to process the structured scheduling parameters and configuration rules based on a pre-built hot rolling scheduling mini-model and generate a slab rolling plan. The comprehensive evaluation module is used to comprehensively evaluate the slab rolling plan based on the hot rolling scheduling model and generate comprehensive evaluation results.

[0015] Furthermore, the scheduling system also includes: The large and small model collaborative execution module is used to generate model scheduling instructions based on the large and small model collaborative execution engine and user requirement analysis results, so as to realize the collaborative cooperation and functional scheduling between the hot rolling scheduling small model and the hot rolling scheduling large model. The front-end display and interaction module is used to receive user input regarding hot rolling scheduling requirements and improvement requests, and to visually display KPI production indicators, the improved slab rolling plan, and comprehensive evaluation results.

[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: First, unlike related technologies that rely on manual labor and face difficulties in digital modeling, this invention, based on practical experience in production scheduling and a high-quality metallurgical knowledge base, deeply integrates the capabilities of the large model, metallurgical expertise, and business rules through multi-stage incremental pre-training and targeted reinforcement fine-tuning on the foundation of the large model, thereby creating a large hot rolling scheduling model for metallurgical process optimization.

[0017] Secondly, unlike related technologies that cannot achieve dynamic response and face complex multi-objective optimization challenges, this invention uses a large-scale hot rolling scheduling model as its core and innovatively adopts a collaborative approach between large and small models. Leveraging the large model's natural language understanding, complex reasoning, and parameter generation capabilities, combined with the advantages of the intelligent hot rolling scheduling small model in multi-objective optimization, simulation verification, and high-dimensional problem solving, this invention effectively overcomes five major technical challenges in hot rolling scheduling scenarios: the difficulty of digitally modeling empirical rules, diverse requirements, high dynamic response requirements, complex multi-objective optimization, and a large solution space. Therefore, by scientifically planning and dynamically scheduling the production tasks of the hot rolling production line, the invention achieves the goals of maximizing production efficiency, reducing costs, ensuring quality, and delivering on time.

[0018] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 is a flowchart of the intelligent scheduling method for hot rolling with large and small models in accordance with an embodiment of the present invention; Figure 2 This is a system framework diagram for the collaboration of large and small models according to an embodiment of the present invention; Figure 3 This is a system flowchart for the collaboration of large and small models according to an embodiment of the present invention; Figure 4 This is a flowchart of the algorithm for the hot rolling scheduling small model according to an embodiment of the present invention; Figure 5 This is an architecture diagram of the large-scale model collaborative execution engine according to an embodiment of the present invention; Figure 6 This is an architecture diagram of the front-end display and interaction in an embodiment of the present invention; Figure 7 This is a schematic diagram of the main modules of the hot rolling intelligent scheduling system with large and small model collaboration according to an embodiment of the present invention. Detailed Implementation

[0020] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0021] A specific embodiment of the present invention discloses a smart scheduling method for hot rolling with coordinated large and small models, such as... Figure 1 As shown, it includes the following steps S1 to S3: Step S1: Based on the pre-built hot rolling scheduling model, analyze the user-input hot rolling scheduling related requirements and generate structured scheduling parameters and configuration rules.

[0022] Implementation specifically includes: analyzing user-inputted hot-rolling scheduling-related requirements based on the aforementioned hot-rolling scheduling model to obtain key intentions, constraints, and / or optimization preferences; these hot-rolling scheduling-related requirements are essentially user-inputted natural language conversations related to hot-rolling scheduling; based on the user requirement analysis results, guiding users to input or upload scheduling data and rule files corresponding to the key intentions, constraints, and / or optimization preferences, wherein the scheduling data mainly includes the specific attributes of the slab to be scheduled, production order information, and real-time factory status; the rule files define the process specifications and business strategies that must be followed during production, mainly including process restriction rules, quality control standards, and scheduling preference rules; Combining the user requirements analysis results, scheduling data, and rule files, structured scheduling parameters and configuration rules are generated. Specifically, guided by the user requirements analysis results, the scheduling data provided by the user is optimized and integrated to generate structured scheduling parameters. Simultaneously, the rule files provided by the user are parsed and formalized, transforming them into structured configuration rules. Structured scheduling parameters refer to the formatted data set formed after integrating and optimizing the scheduling data, such as slab rolling sequence, key rolling process parameters, time plans, and efficiency indicators. Structured configuration rules refer to the set of logical conditions generated after parsing and formalizing the rule files, which can be directly recognized by the rule engine, such as "the hardness difference between adjacent slabs must not exceed 3 levels," "the width jump must not exceed 200mm," and "must include a specific steel grade," etc.

[0023] In some implementations, when a user submits a hot rolling scheduling request in natural language, such as "Please generate a rolling plan that balances efficiency and slab consumption for next Wednesday's orders, prioritizing XX steel grade," the hot rolling scheduling big data model will analyze the user's hot rolling scheduling request, identify key intentions, constraints, and optimization preferences, and provide a list of relevant data required to fulfill the request, such as orders, inventory, and rolling specifications.

[0024] Afterwards, the large hot-rolling scheduling model will guide users to input or upload relevant data and rule files. This allows pre-defined field configurations (including field types, value ranges, and other field attributes) to be imported into the large model as format constraints along with the input data. This enables the large model to not only validate data formats (such as CSV column names and data types) but also execute business logic validations. For example, it can check whether the thickness and width are within the mill's capacity range and whether the steel grade code conforms to company standards. This transforms user requirements and validated data into structured input that the small hot-rolling scheduling model can execute, generating the user-specified scheduling plan. For details, please refer to [link / reference]. Figure 2 and Figure 3 As shown.

[0025] Preferably, the training process of the large-scale hot rolling scheduling model includes: Incremental pre-training of a general-purpose foundational model is performed using a professional knowledge base in the metallurgical field to generate a basic large model with metallurgical knowledge. In this embodiment, the general-purpose foundational model refers to a deep learning model with broad knowledge representation and general capabilities formed by pre-training on massive and diverse data. A hybrid dataset is constructed and divided into a training dataset and a test dataset. The hybrid dataset contains at least parameter generation data, rule generation data, and rolling process evaluation data, and each type of data contains task instructions and their corresponding standard answer results. The standard answer result refers to the accurate and expected output content that meets the process requirements and is predefined by domain experts or based on production specifications for each task instruction. For example, in the parameter generation data, a task instruction is "Please generate the rolling temperature of the slab", and its corresponding standard answer result is "1150℃". Based on multiple preset hyperparameters and combined with the training dataset, the basic large model is fine-tuned and trained to obtain multiple candidate scheduling large models; by using the training dataset in the mixed training dataset to perform unified supervised fine-tuning of the basic large model, the basic large model can learn and master the processing capabilities of the multiple task types at the same time. Based on the test dataset, each of the candidate scheduling models is tested, and the candidate scheduling model with the highest average score is selected as the final hot rolling scheduling model, in conjunction with expert scoring. This enables the hot rolling scheduling model to automatically invoke the corresponding task processing capabilities based on the user's input session commands.

[0026] It can be understood that through the above training method, the present invention enables a single hot rolling scheduling large model to simultaneously acquire multiple professional capabilities: First, a basic large model is pre-trained using professional texts in the hot rolling scheduling field to enable it to master domain knowledge; then, a mixed training dataset covering all target tasks is constructed, and each training sample consists of an "instruction-output" pair, where the instruction clearly indicates the task type (such as "analyze requirements", "check data", "generate parameters", "evaluate suggestions", etc.), and the output is the standard answer result for the corresponding task; finally, the training data for these different tasks are mixed and the model is uniformly trained, enabling the model to automatically identify the task type based on the user input instruction and call the corresponding capabilities for response, thereby integrating multiple professional functions such as requirement analysis, data inspection, parameter configuration, solution evaluation, and optimization suggestions in one model.

[0027] In some embodiments, the training process of the hot rolling scheduling large model mainly includes the following processes: The first stage is to learn metallurgical knowledge.

[0028] First, to enable the hot rolling scheduling large model to more effectively understand metallurgical professional data, based on a vast professional metallurgical knowledge base, which covers metallurgical books, patents, papers, etc., and use self-supervised learning algorithms to conduct incremental pre-training on the general base large model, so that the model can master basic metallurgical knowledge and understand professional terms. Specifically, it includes: (1) The model architecture adopts a Transformer architecture that only contains a decoder, which is the same as the base large model. This model is trained to predict the next word based on the previous words.

[0029] (2) The training objective is to maximize the likelihood probability. For example, given a text sequence x1, x2,..., x T , this model is an autoregressive model, and its objective is to model the joint probability distribution P(x1, x2,..., x T ), and decompose this joint probability into a product of a series of conditional probabilities. The specific formula is as follows: ; Among them, the parameter x represents the basic data unit in the sequence, T represents the total length of the sequence, t represents the current prediction time step; x<t represents all the words before position t (x1,..., x t 1).

[0030] (3) The training loss function is the cross-entropy loss. For example, for each position t, the model will output a probability distribution P θ (x t|x < t); The training objective is to make the probability distribution predicted by the model for the next word as close as possible to the true one-hot distribution (i.e., the probability of the true next word is 1 and others are 0), that is, to minimize the difference.

[0031] For the loss of a single sample at position t It is obtained from the following cross-entropy calculation formula: ; For a complete text sequence (or a batch of data containing multiple sequences), its total training loss Is the average of the losses at all valid positions in the sequence, and the specific formula is as follows: ; Thus, by minimizing this total training loss through an optimization algorithm (such as the gradient descent algorithm), the model parameters θ are continuously adjusted, so that the probability distribution P θ (xt|x < t) gets closer and closer to the true data distribution, and finally obtains the ability to accurately predict and generate the structured text sequence of hot rolling scheduling.

[0032] The second stage is to learn professional skills.

[0033] In the fine-tuning stage of the large model, the system will perform multiple rounds of supervised fine-tuning by combining a mixed dataset with multiple types of manual annotations. The data includes manually annotated demand response question-and-answer pairs, rolling process evaluation data, parameter generation data, rule generation data, etc. The fine-tuning can adopt the supervised fine-tuning method based on the LoRA algorithm. For example, given the model parameters θ in the aforementioned pre-training stage and its corresponding SFT dataset , where x i is the input instruction, y i is the expected answer, N represents the number of samples, and i represents the i-th sample. The loss is only calculated for the answer part during the fine-tuning process, and the loss function uses the cross-entropy loss, and the specific form is as follows: ; It can be understood that to solve the problems of high training cost and easy catastrophic forgetting in the traditional supervised fine-tuning method, the LoRA (Low-Rank Adaptation) algorithm is introduced in this stage in the SFT framework.

[0034] Specifically, for the weight matrix W0 ∈ R d×k of any fully connected layer in the pre-trained model, the LoRA algorithm does not directly update it, but constrains its incremental update ΔW to a low-rank decomposition form: ; where, B ∈ R d×r , A ∈ Rr×k All are trainable low-rank matrices; With a fixed rank, the number of ΔW parameters is significantly reduced; For input h, the linear layer forward computation formula of LoRA is used: ; Therefore, the SFT loss function combined with LoRA can be expressed as follows: ; Here, θ0 represents the frozen pre-trained parameters, which are adaptively updated only through the low-rank matrix {A,B}. This training method achieves efficient fine-tuning with minimal additional parameters, significantly reducing computational costs, while effectively mitigating the forgetting of pre-trained knowledge.

[0035] The third stage is the evaluation of large-scale models.

[0036] Specifically, this includes: testing multiple models from the above training results using the training and testing sets, conducting manual scoring and evaluation, and selecting the model with the highest average score as the final model. This model has capabilities such as demand analysis, data list generation, parameter generation, and rolling process evaluation.

[0037] It is evident that by leveraging the metallurgical business understanding, natural language understanding, complex reasoning, and parameter generation capabilities of the hot rolling scheduling big model, user needs can be analyzed and generated, raw data can be acquired and verified, rule generation and parameter optimization can be carried out, the rolling process can be evaluated and optimization suggestions can be put forward, and finally, improvement needs can be obtained from users in an interactive manner. This achieves the successful transformation and efficient deployment of the general big model into a professional hot rolling scheduling decision model.

[0038] It should be noted that the detailed training method for the large hot rolling scheduling model can be implemented by those skilled in the art with reference to existing technologies; the improvement of this invention lies in applying the above training method to the specific industrial field of hot rolling scheduling, and constructing a multi-task collaborative training framework for scheduling parameter generation, rule execution and plan evaluation.

[0039] Step S2: Based on the pre-built hot rolling scheduling small model, the structured scheduling parameters and configuration rules are processed to generate a slab rolling plan.

[0040] In practice, this includes: clustering the scheduling parameters according to the slab properties to obtain multiple candidate slab groups containing parameter features; for example, grouping slabs with similar attribute values ​​into the same candidate slab group. Based on the configuration rules, determine the rule constraints for the preset comprehensive objective optimization function; The slab rolling plan is generated based on the multiple candidate slab groups, the preset comprehensive objective optimization function, and the rule constraints.

[0041] The preset comprehensive objective optimization function is constructed in the following manner: Set multiple KPI production indicators to be optimized, and configure corresponding weight coefficients for each KPI production indicator; By linearly weighting and summing the objective functions of each KPI production indicator with its corresponding weight coefficients, multiple KPI production indicators are aggregated into a comprehensive objective optimization function for a single objective problem.

[0042] Preferably, generating the slab rolling plan includes: Based on the breadth-first search algorithm, the current optimal solution B, the current highest score, and the solution queue Q to be expanded are initialized. All of these parameters are internal variables defined and initialized by the algorithm logic itself. Empty rolling plans and slab groups corresponding to special slab types in the user demand analysis results are used as initial solutions and placed into the solution queue Q. The solution queue Q is searched cyclically. During the cyclic search, the current solution S taken from the solution queue Q is expanded according to the multiple candidate slab groups and rule constraints. The current solution S in the solution queue Q that cannot be further expanded and has no state change is scored according to the preset comprehensive objective optimization function. Here, S can be defined as the slab rolling sequence, and Q is a buffer for performing search operations. When a new slab group is added to S, it means that there is still a possibility of continuing the search, and it needs to be placed in the queue Q. The search will continue until the queue Q is empty, which indicates that there is no more space to continue the search.

[0043] Then, the scoring result is compared with the current highest score, and the current optimal solution B and the current highest score are updated based on the comparison result to perform the next round of search; until the solution queue Q is empty, the current optimal solution B with the highest score is taken as the slab rolling plan.

[0044] Furthermore, the current solution S taken from the solution queue Q is expanded, and the current solutions S in the solution queue Q that cannot be further expanded and have no state change are scored according to the preset comprehensive objective optimization function, including: Obtain the unused slab groups of the current solution S from the plurality of candidate slab groups, and use them as the available slab group set U of the current solution S; Determine whether the set of available slab groups U is empty, where: When the set of available slab groups U is not empty, according to the rule constraints, it is determined one by one whether the current slab group G in the set of available slab groups U can be included in the current solution S; if it can be included, the current slab group G is included in the current solution S and a new solution S is formed. The new solution S is not included in the solution queue Q. At that time, the new solution S Place it into the solution queue Q, and then check again whether the set of available slab groups U is empty; When the available slab set U is empty, determine whether the state of the current solution S has changed after being taken out of the solution queue Q: if it has changed, put the changed current solution S into the solution queue Q and continue to perform the loop search; otherwise, calculate the score of the current solution S based on the preset comprehensive objective optimization function.

[0045] It is understandable that whether S changes can be used to distinguish whether the current solution S still has the potential for expansion or has become a complete rolling process to be scored; if the state of the entire solution changes, it indicates that a new available slab has been found since S was taken out, and it can still be matched again in the set of available slabs U; if the state of S does not change, it proves that all possibilities have been tried and this path has reached its end, at which point its KPI score needs to be calculated.

[0046] In some preferred embodiments, the hot rolling scheduling mini-model is mainly configured with the following sub-functional modules: (1) Data preparation module: When a scheduling process is initiated, the large-scale hot-rolling scheduling model first parses the relevant user-input demand data, converting the scheduling data into an internally usable data structure. This generates structured scheduling parameters, primarily including slab inventory information (e.g., slab number, slab length, width, thickness, steel grade), casting information (e.g., casting number, casting start time, slab length, width, thickness, steel grade), and steel grade information (e.g., steel grade hardness, temperature). Simultaneously, based on the smaller hot-rolling scheduling model, the structured scheduling parameters are clustered according to slab attributes, grouping parameters with similar attribute values ​​into the same slab group.

[0047] (2) Rule parsing and execution module: The large-scale hot-rolling scheduling model can transform its generated structured configuration rules into multi-objective optimization objects and constraints. For example, the structured configuration rule "the hardness difference between two consecutive slabs should not exceed 3" can be transformed into a rule or mathematical expression that the system can understand. Then, based on the input current slab sequence and candidate slab groups, and according to the rules after the transformation, the hot rolling scheduling mini-model outputs the result of whether the candidate slab group can be placed in the current slab sequence.

[0048] (3) Indicator Calculation Module: Based on specific business definitions and preset objective functions, KPI production indicator values ​​are calculated for display or evaluation by a large-scale hot rolling scheduling model. Specific indicator values ​​include at least hot charging rate and rolling length. For example, the objective function for hot charging rate is typically defined as the ratio of the weight (or quantity) of hot-charged slabs to the total weight (or quantity) of slabs in the furnace within a certain statistical period; the objective function for rolling length can generally be quantified as the number of slabs continuously rolled in a single rolling cycle, the total length, or the total weight. The specific form and weight of the objective functions corresponding to the above KPI production indicators can be customized by relevant technical personnel based on actual production needs, process constraints, and business priorities.

[0049] (4) Multi-objective optimization module: Considering the coordination issues among modules, this embodiment adopts a linear weighting method to integrate multiple objectives into a single-objective optimization problem. Specifically, assuming there are k objective functions to be optimized, they can typically be expressed as: F(x) = [f1(x), f2(x), ..., f k [(x)] ; where f1(x), f2(x), ..., f k (x) represents the objective function of different KPI production indicators, and x represents the specific scheduling scheme.

[0050] First, each objective needs to be standardized, compressing its value range to between 0 and 1, which can be represented as: NormF(x) = [Norm(f1(x)), Norm(f2(x)), ..., Norm(f... k (x))]; Secondly, by using the linear weighting method, it is transformed into the following single-objective optimization problem: ; Where w1+w2+...+w k =1; w1, w2…w k These are the weighting coefficients for each KPI production indicator, and their values ​​are determined by the specific applicable scenarios and needs.

[0051] Thus, the complex multi-objective optimization is transformed into a single-objective form that is more computationally efficient and easier to solve, providing clear and quantifiable evaluation criteria for subsequent search algorithms, thereby ensuring the comprehensive optimality of the final rolling plan in multiple dimensions such as efficiency, quality, and cost.

[0052] (5) Rolling stroke generation module: Based on a breadth-first search algorithm, the outputs of the multi-objective optimization module and the rule execution module are integrated to generate a complete rolling process including the slab rolling sequence, i.e., the final hot-rolled slab rolling plan. Combined with... Figure 4 As shown, the specific algorithm is as follows: (a) Initialize the current highest score to 0; initialize the current optimal solution B to be empty; initialize the solution queue Q, and put the empty rolling schedule into Q as the initial solution.

[0053] (b) If the user requirements include a special slab type, then the slab group that meets the type should be included in the initial solution.

[0054] (c) Determine if the solution queue Q is empty: If Q is empty, output the current optimal solution B; otherwise, take a solution from Q as the current solution S.

[0055] (d) Obtain the unused slab groups in the current solution S, and use them as the set of available slab groups U for the current solution S.

[0056] (e) Determine if the set of available slab groups U is empty: If U is empty, continue to determine if the current solution S has changed after being taken from the solution queue Q. If it has changed, return to step (c) and perform a new round of judgment and query; if it has not changed, proceed to step (h). If U is not empty, proceed to step (f).

[0057] (f) Take a current slab group G from the available slab group set U and delete G from U; then input S and G into the rule parsing and execution module, and combine the constraints obtained from the structured configuration rules to determine whether the current slab group G can be put into the current solution S. If it cannot be put into the solution, proceed to step (e); otherwise, proceed to step (g).

[0058] (g) A new solution S can be formed after the current slab group G is placed into the current solution S. If the new solution is not in the solution queue Q, then add the new solution to Q and proceed to step (e); if the new solution S If it is already in the dequeue Q, proceed directly to step (e).

[0059] (h) Calculate the score of the current solution S through the index calculation module and the multi-objective optimization module. If the score of S is greater than the current highest score, set the current highest score as the score of S, and take the current solution S as the current optimal solution B. Then proceed to step (c) to continue the loop search. Otherwise, retain the current optimal solution B and its highest score, and proceed directly to step (c).

[0060] To facilitate understanding of the algorithm flow in (a) to (h) above, a specific embodiment is given below.

[0061] Assume the current production line has three slab groups to be scheduled, with corresponding process parameters as follows: Slab group G1 (steel grade A, width 1250 mm), slab group G2 (steel grade A, width 1280 mm), and slab group G3 (steel grade B, width 1250 mm). The scheduling process must adhere to a process constraint rule: the width difference between adjacent rolled slabs cannot exceed 50 mm to ensure production stability and product quality.

[0062] Specifically, when the algorithm starts running, it first initializes by setting the current highest score to 0, the current optimal solution B to empty, and creating a solution queue Q to store the plans to be processed. Then, an empty rolling plan is placed in queue Q as the starting point for exploration. The core of the algorithm is a cyclical search process. As long as queue Q is not empty, the algorithm will retrieve a plan to be processed from it, called the current solution S. Next, the system will find all slab groups that have not yet been used by this current solution S, forming a new set, denoted as the available slab group set U. At this point, the algorithm determines the next operation based on the situation in set U. Where: If set U is not empty, it means that the current solution S may still have the potential to add new slab groups to extend the rolling process. The algorithm will select a candidate slab group G from U and make a judgment based on the width difference constraint rules mentioned above. For example, for the current solution S=[G1], when trying to add G2, the width difference between the two slabs is calculated to be 30 mm, which satisfies the preset rule constraint. Therefore, a new, longer plan, i.e., a new solution S, can be generated. =[G1, G2]. Next, the system checks if this new plan is already in the unqueue Q. If not, it adds it to the tail of the queue, waiting to be retrieved and expanded later. This process tries each slab group in U one by one until U is cleared.

[0063] When set U is empty, it indicates that the current solution S cannot add any new slab groups. At this time, the algorithm needs to determine whether the state of the current solution S has changed since it was taken out of the solution queue Q. If it has changed, it means that S still has the potential to expand into other possible plans (for example, there are currently 3 groups of slabs, and after one round of queries, a usable 4th group of slabs G4 is found, and S can continue to expand). In this case, the algorithm will put the new plan generated based on the new slab back into queue Q, waiting to be taken out again in future cyclic searches.

[0064] If set U is empty and the current solution S remains unchanged, it means that the current solution S is a complete rolling plan that cannot be further expanded, and all its possibilities have been explored. Taking the complete plan S=[G1, G2, G3] in the above scenario as an example, when it is processed, all slabs have been used, U is empty, and no new solutions have been generated in this round (no change), which indicates that it is a path that has been explored to its limit. At this time, the algorithm will trigger the scoring mechanism, calling the index calculation module and the multi-objective optimization module to evaluate the complete plan. Assuming that the score of S is 85 points, the system will compare this score with the currently recorded highest score (initially 0 points), and update the current highest score to 85 based on the comparison result, and update the current optimal solution B to [G1, G2, G3], until the loop ends.

[0065] In the iterative search process, the solution queue Q can be understood as a dynamic work list, continuously adding and removing newly generated partial plans. Each complete plan is evaluated after exhausting all possible expansions. Finally, when the solution queue Q is empty, it means that all possible plan combinations have been searched and evaluated. At this point, the algorithm outputs the optimal solution B with the highest score recorded throughout the search process, which serves as the final hot rolling scheduling scheme, i.e., the slab rolling plan. This ensures that the rolling plan has optimal overall performance, achieving the optimization goals of maximizing production efficiency, reducing costs, ensuring quality, and delivering on time.

[0066] It is evident that by leveraging the expertise of the hot rolling scheduling small model in multi-objective optimization, simulation verification, and high-dimensional problem solving, and by utilizing the structured configuration rules and scheduling parameters generated by the hot rolling scheduling large model, efficient and reliable hot rolling plan formulation can be achieved.

[0067] Step S3: Based on the hot rolling scheduling model, perform a comprehensive evaluation of the slab rolling plan and generate a comprehensive evaluation result.

[0068] Preferably, the method further includes: generating an improved slab rolling plan and a comprehensive evaluation result based on the comprehensive evaluation result and the user's input natural language improvement request. This invention achieves continuous optimization through human-machine collaboration, enabling users to propose clear improvement requests in the form of natural language conversation based on the rolling plan and comprehensive evaluation output by the two models, and to initiate a new round of scheduling optimization, generating an improved slab rolling plan and its corresponding updated evaluation result.

[0069] During implementation, the large-scale hot rolling scheduling model will receive the rolling plan and various KPI production indicators generated by the small-scale hot rolling scheduling model, and conduct a comprehensive evaluation from an expert perspective. The evaluation approach is consistent with the standards and the training data provided by experts. It typically includes analyzing key indicators of the plan (such as rolling kilometers, number of roll changes, energy consumption estimation, etc.), identifying potential risks (such as whether a certain transition is too abrupt), comparing with historical best practices, and proposing specific and feasible optimization suggestions (such as "suggesting to swap the order of two slabs to reduce hardness jumps and improve surface quality").

[0070] After the slab rolling plan is generated, users can also propose interactive improvement requirements based on the site conditions or optimization suggestions. Then, based on the improvement requirements, steps S1 and S2 are repeated to carry out a new round of demand analysis, plan generation and comprehensive evaluation operations.

[0071] Preferably, the method further includes: generating model scheduling instructions based on the collaborative execution engine of the large and small models and the user demand analysis results, so as to realize the collaborative cooperation and functional scheduling between the small hot rolling scheduling model and the large hot rolling scheduling model.

[0072] In some preferred embodiments, such as Figure 5 As shown, the large and small model collaborative execution engine is mainly configured with the following sub-functional modules: (1) Session Management Module: Using a natural language conversational interaction mode, it maintains independent context information for each user session to ensure the continuity and isolation of multi-turn dialogues.

[0073] (2) Workflow Management Module: The entire hot rolling scheduling task can be organized into a workflow execution diagram. This module is responsible for driving and calling different sub-functions of the large and small models to carry out collaborative work based on different states. This module specifically includes the following states: State 1 Requirements Analysis: Call the hot rolling scheduling model to trigger the requirements analysis function.

[0074] Status 2 Data Verification: Call the hot rolling scheduling model to trigger the verification function.

[0075] State 3 Parameter Generation: Call the large hot rolling scheduling model, trigger the parameter generation function, and transmit the output results to the small hot rolling scheduling model.

[0076] State 4 Rolling Stroke Generation: Call the hot rolling scheduling mini-model and execute the aforementioned rolling stroke generation algorithm.

[0077] State 5 Result Evaluation: Return the slab rolling plan output by the hot rolling scheduling small model to the hot rolling scheduling large model, and trigger the evaluation function of the hot rolling scheduling large model.

[0078] Status 6 User Feedback: Wait for user input and jump to the corresponding status based on user feedback to execute the corresponding function. For example, if the user inputs "recalculate", it will jump to status 4, and if the user inputs "modify requirements", it will jump to status 1.

[0079] (3) Context storage module: responsible for storing and managing the context dialogue input by the user, including historical dialogue records, system execution status and uploaded files, which helps to realize the continuous operation of the user.

[0080] Preferably, users can achieve two-way interaction with the system through the front-end display and interaction module: this module can clearly present the scheduling plan, evaluation results, and key indicators generated by the model to the user; at the same time, it can also receive user-input parameters, instructions, and feedback, and transmit them to other functional modules in real time, thereby forming a closed loop of human-machine collaborative decision-making. Figure 6 As shown, this front-end display and interaction module is mainly configured with the following sub-functional modules: (1) Visual dashboard: Displays a rolling plan diagram (including slab attribute information and rolling sequence), KPI indicator chart, and the achievement status of optimization goals.

[0081] (2) Interactive dialog window: natural language chat window.

[0082] (3) Data upload and template download interface: mainly responsible for the visual configuration and editing of parameters and rules.

[0083] Therefore, it can be seen that the embodiments of the present invention can achieve one of the following beneficial effects: First, based on practical experience in production scheduling and a high-quality metallurgical knowledge base, this invention, on the basis of a large-scale base model, deeply integrates the capabilities of the large model, metallurgical expertise and business rules through multi-stage incremental pre-training and targeted reinforcement fine-tuning, creating a large-scale hot rolling scheduling model for metallurgical process optimization.

[0084] Secondly, this invention takes a large-scale hot rolling scheduling model as its core and innovatively adopts a collaborative approach between large and small models. It leverages the large model's natural language understanding, complex reasoning, and parameter generation capabilities, combined with the advantages of a small-scale intelligent hot rolling scheduling model in multi-objective optimization, simulation verification, and solving high-dimensional problems. This effectively overcomes five major technical challenges in hot rolling scheduling scenarios: the difficulty of digitally modeling empirical rules, diverse requirements, high dynamic response requirements, complex multi-objective optimization, and a large solution space. Therefore, by scientifically planning and dynamically scheduling the production tasks of the hot rolling production line, the invention achieves the goals of maximizing production efficiency, reducing costs, ensuring quality, and delivering on time.

[0085] In another embodiment of the present invention, a hot rolling intelligent scheduling system with coordinated large and small models is proposed, such as... Figure 7 As shown, it specifically includes the following modules: The conversation analysis module is used to analyze user-input hot rolling scheduling-related requirements based on a pre-built large model of hot rolling scheduling, and generate structured scheduling parameters and configuration rules. The plan generation module is used to process the structured scheduling parameters and configuration rules based on a pre-built hot rolling scheduling mini-model and generate a slab rolling plan. The comprehensive evaluation module is used to comprehensively evaluate the slab rolling plan based on the hot rolling scheduling model and generate comprehensive evaluation results.

[0086] Preferably, the scheduling system further includes: The large and small model collaborative execution module is used to generate model scheduling instructions based on the large and small model collaborative execution engine and user requirement analysis results, so as to realize the collaborative cooperation and functional scheduling between the hot rolling scheduling small model and the hot rolling scheduling large model. The front-end display and interaction module is used to receive user input regarding hot rolling scheduling requirements and improvement requests, and to visually display KPI production indicators, the improved slab rolling plan, and comprehensive evaluation results.

[0087] The above-described method and system embodiments are based on the same principles, and their related aspects can be referenced from each other to achieve the same technical effects. For specific implementation processes, please refer to the foregoing embodiments, which will not be repeated here.

[0088] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0089] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes 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.

Claims

1. A smart scheduling method for hot rolling with coordinated large and small models, characterized in that, include: Based on a pre-built large model of hot rolling scheduling, the user-input hot rolling scheduling-related requirements are analyzed to generate structured scheduling parameters and configuration rules. Based on a pre-built hot rolling scheduling model, the structured scheduling parameters and configuration rules are processed to generate a slab rolling plan. Based on the aforementioned hot rolling scheduling model, a comprehensive evaluation of the slab rolling plan is performed, and a comprehensive evaluation result is generated.

2. The scheduling method according to claim 1, characterized in that, The pre-built hot rolling scheduling model analyzes user-input hot rolling scheduling requirements and generates structured scheduling parameters and configuration rules, including: Based on the aforementioned hot rolling scheduling model, the user-input hot rolling scheduling-related requirements are analyzed; based on the user requirement analysis results, the user is guided to input or upload the corresponding scheduling data and rule files; Based on the user requirements analysis results, scheduling data, and rule files, structured scheduling parameters and configuration rules are generated.

3. The scheduling method according to claim 2, characterized in that, The training process of the large-scale hot rolling scheduling model includes: Incremental pre-training of a general-purpose base model was performed using a professional knowledge base in the field of metallurgy to generate a basic large model with metallurgical knowledge. Construct a hybrid dataset and divide the hybrid dataset into a training dataset and a test dataset; the hybrid dataset contains at least parameter generation data, rule generation data and rolling mill evaluation data, and each type of data contains task instructions and their corresponding standard answer results; Based on multiple preset hyperparameters and combined with the training dataset, the basic large model is fine-tuned and trained to obtain multiple candidate scheduling large models. Based on the test dataset, each of the candidate scheduling models was tested, and the candidate scheduling model with the highest average score was selected as the final hot rolling scheduling model, in conjunction with expert scoring.

4. The scheduling method according to claim 1, characterized in that, The process, based on a pre-built hot rolling scheduling model, involves processing the structured scheduling parameters and configuration rules to generate a slab rolling plan, including: Based on the slab properties, the scheduling parameters are clustered to obtain multiple candidate slab groups containing parameter features; Based on the configuration rules, determine the rule constraints for the preset comprehensive objective optimization function; The slab rolling plan is generated based on the multiple candidate slab groups, the preset comprehensive objective optimization function, and the rule constraints.

5. The scheduling method according to claim 4, characterized in that, The step of generating the slab rolling plan based on the multiple candidate slab groups, the preset comprehensive objective optimization function, and rule constraints includes: Based on the breadth-first search algorithm, the current optimal solution B, the current highest score, and the solution queue Q to be expanded are initialized; empty rolling plans and slab groups corresponding to special slab types in the user demand analysis results are used as initial solutions and placed into the solution queue Q; The solution queue Q is searched cyclically. During the cyclic search, the current solution S taken from the solution queue Q is expanded according to the multiple candidate slab groups and rule constraints. The current solution S in the solution queue Q that cannot be further expanded and has no state change is scored according to the preset comprehensive objective optimization function. The scoring result is compared with the current highest score, and the current optimal solution B and the current highest score are updated based on the comparison result to perform the next round of search; until the solution queue Q is empty, the current optimal solution B with the highest score is taken as the slab rolling plan.

6. The scheduling method according to claim 5, characterized in that, The step of expanding the current solution S taken from the solution queue Q based on the multiple candidate slab groups and rule constraints, and scoring the current solution S in the solution queue Q that cannot be further expanded and has no state change according to the preset comprehensive objective optimization function, includes: Obtain the unused slab groups of the current solution S from the plurality of candidate slab groups, and use them as the available slab group set U of the current solution S; Determine whether the set of available slab groups U is empty, where: When the set of available slab groups U is not empty, according to the rule constraints, it is determined one by one whether the current slab group G in the set of available slab groups U can be included in the current solution S; if it can be included, the current slab group G is included in the current solution S and a new solution S is formed. The new solution S is not included in the solution queue Q. At that time, the new solution S Place it into the solution queue Q, and then check again whether the set of available slab groups U is empty; When the available slab set U is empty, determine whether the state of the current solution S has changed after being taken out of the solution queue Q: if it has changed, put the changed current solution S into the solution queue Q and continue to perform the loop search; otherwise, calculate the score of the current solution S based on the preset comprehensive objective optimization function.

7. The scheduling method according to claim 6, characterized in that, The preset comprehensive objective optimization function is constructed in the following manner: Set multiple KPI production indicators to be optimized, and configure corresponding weight coefficients for each KPI production indicator; By linearly weighting and summing the objective functions of each KPI production indicator with its corresponding weight coefficients, multiple KPI production indicators are aggregated into a comprehensive objective optimization function for a single objective problem.

8. The scheduling method according to claim 1, characterized in that, The method further includes: Based on the collaborative execution engine of the large and small models and the results of user requirement analysis, model scheduling instructions are generated to realize the collaborative cooperation and functional scheduling between the small hot rolling scheduling model and the large hot rolling scheduling model.

9. A hot rolling intelligent scheduling system with coordinated large and small models, characterized in that, include: The conversation analysis module is used to analyze user-input hot rolling scheduling-related requirements based on a pre-built large model of hot rolling scheduling, and generate structured scheduling parameters and configuration rules. The plan generation module is used to process the structured scheduling parameters and configuration rules based on a pre-built hot rolling scheduling mini-model and generate a slab rolling plan. The comprehensive evaluation module is used to comprehensively evaluate the slab rolling plan based on the hot rolling scheduling model and generate comprehensive evaluation results.

10. The scheduling system according to claim 9, characterized in that, Also includes: The large and small model collaborative execution module is used to generate model scheduling instructions based on the large and small model collaborative execution engine and user requirement analysis results, so as to realize the collaborative cooperation and functional scheduling between the hot rolling scheduling small model and the hot rolling scheduling large model. The front-end display and interaction module is used to receive user input regarding hot rolling scheduling requirements and improvement requests, and to visually display KPI production indicators, the improved slab rolling plan, and comprehensive evaluation results.

Images