Group subordinate electrolytic aluminum branch plant yield distribution decision-making method based on large model

Abstract

The invention discloses a group subordinate electrolytic aluminum branch yield distribution decision method based on a large model, and the method can comprehensively consider the conditions and historical production conditions of each group subordinate electrolytic aluminum branch, carries out the comprehensive evaluation and decision, and gives a reasonable yield distribution scheme, thereby improving the production efficiency and economic benefits of an aluminum industry group. The method comprises the following steps: collecting and processing historical data of each electrolytic aluminum branch plant; training an electrolytic aluminum production management large model of the group based on the industrial basic large model; collecting and processing production and operation data of the recent year to finely adjust the large model; inputting a production plan and external condition data of the next year, and reasoning a yield distribution scheme of each electrolytic aluminum branch plant through the large model; calculating an expected operation condition and economic data, and issuing a yield distribution decision result after manual auditing; according to the method, various factors of each branch plant can be comprehensively considered, a scientific and accurate yield allocation scheme is provided, and resource allocation is optimized.

Description

Technical Field

[0001] The present invention relates to the technical fields of electrolytic aluminum production operation management and artificial intelligence, and specifically provides a method for allocating production output decisions for electrolytic aluminum branches under a group based on a large model. Background Art

[0002] As a representative of high-energy-consuming industries, the electrolytic aluminum industry has a complex production process and extremely high requirements for refined management. A group usually has multiple electrolytic aluminum branches, and each branch has different situations, including various cell types, different cell ages, efficiencies, and performances of electrolytic cells, as well as differences in raw material components, raw material prices, and sales profits. These factors jointly affect the production capacity, labor efficiency, and economic benefits of each electrolytic aluminum plant; In recent years, with the advancement of industrial informatization and industrial intelligence, electrolytic aluminum plants have successively established data acquisition and information management systems, accumulating a large amount of working state data of electrolytic cells under different working conditions. Many aluminum industry groups have also established data middle platforms to collect and centrally manage and use the data of each electrolytic aluminum branch. These data can implicitly reflect the equipment conditions, management levels, production capacities, and production efficiencies of each electrolytic aluminum branch, but it is difficult for manual experience and managers to fully obtain and master this implicit knowledge; Currently, when an aluminum industry group formulates production plans for its subordinate electrolytic aluminum branches, it lacks reasonable and effective guidance, making it difficult to maximize the use of production conditions and production resources. Ultimately, it is impossible to give the optimal production output allocation plan, let alone maximize the group's interests. However, traditional data mining methods and small models are difficult to fully mine and utilize the rich knowledge and information contained in these valuable data; Therefore, it is necessary to introduce large model technologies with more powerful performance and stronger correlation learning capabilities to deeply analyze the data. Relying on the massive data foundation, it can assist production managers in completing complex production management and decision-making tasks. A method for allocating production output decisions for electrolytic aluminum branches under a group based on a large model proposed in this invention patent can comprehensively consider the own conditions and historical production situations of each electrolytic aluminum branch, conduct comprehensive evaluation and decision-making, give a reasonable production output allocation plan, enable each electrolytic aluminum branch to operate in the most economical way, and thus maximize the group's interests, which has important practical significance and application value. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention provides a method for allocating production output decisions for electrolytic aluminum branches under a group based on a large model to help an aluminum industry group allocate appropriate production output for its subordinate electrolytic aluminum branches, thereby ensuring the maximization of the group's economic benefits.

[0004] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for making production allocation decisions for electrolytic aluminum branch factories under a group based on a large model, comprising the following steps in sequence: Step S1: Collect historical data of each electrolytic aluminum branch factory under the group; Step S2: Process the historical data collected in Step S1 for use in training the large model; Step S3: Select a large model based on the industrial field, and use the data processed in Step S2 to train the large model to form a large model for electrolytic aluminum production management; Step S4: Obtain the production and operation data of electrolytic aluminum from multiple branch factories within a set time period, and process this data using Step S2 to constitute a dataset for fine-tuning the large model; Step S5: Use the well-processed large model fine-tuning dataset from Step S4 to fine-tune the parameters of the large model. The large model uses a neural network model with more than 100 million parameters; Step S6: Input the production plan and external condition data for the next year into the well-trained and fine-tuned large model; Step S7: Infer the production allocation plan for each electrolytic aluminum branch factory through the large model; Step S8: Calculate the expected operation situation and economic data, and issue the production allocation decision result after manual review.

[0005] Preferably, the historical data collected from each electrolytic aluminum branch factory under the group in Step S1 includes: process data, quality data, equipment data, raw material composition data, raw material cost, labor cost, manufacturing cost, and operation data in electrolytic production.

[0006] Preferably, the processing of the historical data collected in Step S1 in Step S2 includes: Step S201: Data cleaning, checking and processing missing values, outliers, and noise data in the data, using methods such as deletion or interpolation.

[0007] Step S202: Normalize and standardize the data, convert the data to a unified dimension, ensure the consistency and comparability of the data, and facilitate model training.

[0008] Step S203: Sample the data to ensure the time consistency and integrity of the data, and unify data with different time intervals to the same time scale.

[0009] Step S204: Convert the data into a format suitable for large model training, and numericalize or labelize the data according to the data type.

[0010] Step S205: Re-verify the processed data to ensure data quality and applicability.

[0011] Preferably, the method for fine-tuning the large model parameters in step S5 includes: Step S501: Fine-tuning method selection. According to specific requirements and data characteristics, select a suitable fine-tuning method, including LoRA (Low-Rank Adaptation), Prompt Tuning, transfer learning, and incremental learning; Step S502: Parameter adjustment. According to the selected fine-tuning method, adjust the parameters of the large model, add a low-rank matrix in LoRA, and add specific prompt parameters in Prompt Tuning; Step S503: Fine-tuning training. Use the fine-tuning dataset constructed in step S4 to perform fine-tuning training on the large model, monitor the loss function and performance metrics during the training process, and observe the fine-tuning effect; Step S504: Performance evaluation. Use the validation set to evaluate the performance of the fine-tuned model, and check the accuracy and generalization ability of the model, including whether the model can output a compliant aluminum electrolysis production allocation plan, and the output result does not deviate from the actual production range of the aluminum electrolysis branch factory; Step S505: Model optimization. Perform several rounds of model optimization according to the evaluation results, adjust the hyperparameters multiple times, and increase the number of training rounds until the training loss loss and the validation loss val_loss are stable, and the difference between the two is less than 0.1.

[0012] Preferably, the production plan and external condition data in step S6 include: the expected total aluminum production of the group in the next year, the number of electrolytic cells in normal production in each aluminum electrolysis branch factory, the expected raw material composition and contract price of each aluminum electrolysis branch factory, the normal production quantity of electrolytic cells in each aluminum electrolysis branch factory, the sales plan, and the production data.

[0013] Beneficial effects The present invention provides a method for making a production allocation decision for an aluminum electrolysis branch factory under a group based on a large model, having the following beneficial effects: 1. Through in-depth mining and analysis of massive production data by the large model, a more reasonable production allocation plan can be formulated, maximizing the utilization of production conditions and resources, thereby improving the overall production efficiency and economic benefits of the group; 2. Utilizing big data and large model technologies to overcome the limitations of traditional manual experience decision-making, providing a more scientific and accurate production allocation plan, and reducing human errors; 3. The present invention can comprehensively consider various factors such as the cell type, cell age, efficiency, raw material composition, and cost of each aluminum electrolysis branch factory, perform refined management and decision-making, and enable each branch factory to operate in the best state (prolonging the life of electrolytic cells) and in the most economical way on the premise of maximizing the interests of the group; 4. The present invention can quickly adapt to new production conditions and external environmental changes through the fine-tuning and incremental learning of large models, and maintain the timeliness and accuracy of decision-making schemes. 5. The present invention can optimize resource allocation and production plans, contribute to energy conservation and emission reduction, and support the sustainable development of the electrolytic aluminum industry. 6. Based on the existing information platform and large model technology, the present invention has strong operability and popularization, and is easy to implement and apply within the aluminum industry group. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flowchart of the decision-making method of the present invention; Figure 2 is a flowchart of the present invention for processing the collected historical data; Figure 3 is a flowchart of the present invention for fine-tuning the large model. DETAILED DESCRIPTION OF THE INVENTION

[0015] The following further elaborates on the specific implementation manners of the present invention in conjunction with the drawings, making the technical content of the present invention easier to understand. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0016] The above-mentioned large model refers to a neural network model with more than 100 million parameters.

[0017] The management mode of the above-mentioned electrolytic aluminum branch factories is unified management by the aluminum industry group or aluminum industry section. The group uniformly assigns production tasks to each electrolytic aluminum branch factory, and the group has a unified information platform or data center, which has the conditions for unified data collection, centralized management, and use.

[0018] As Figure 1 shown, a decision-making method for output allocation of electrolytic aluminum branch factories under a group based on a large model includes a total of 8 steps from S1 to S8.

[0019] Step S1: Collect historical data of each electrolytic aluminum branch factory under the group.

[0020] The types of collected data include but are not limited to: Basic conditions of each electrolytic cell: cell number, cell type, cell age, hearth voltage drop, overhaul situation, failure rate, etc.

[0021] Process data: cell voltage, series current, current efficiency, etc.

[0022] Quality data: molecular ratio, element content, primary aluminum quality.

[0023] Raw material components: alumina components.

[0024] Cost data: raw material cost, manufacturing cost, labor cost, etc.

[0025] Business data: raw material procurement data, market demand, sales data, inventory status, The data collection method is information integration, including the cell control management system, quality system, MES system, ERP system, and financial system of each electrolysis plant. The specific data extraction method uses the form data extraction work of the database and the method of API interface call. For the data not covered by the current information construction, the method of manual regular filling and uploading is adopted.

[0026] Step S2: Process the collected historical data. As shown in the appendix Figure 2 The specific implementation examples of the processing are as follows: Step S201: Data cleaning. Check and process the missing values, outliers, and noise data in the data, using the methods of deletion or interpolation. In this embodiment, the following formula is used for interpolation of the missing values:

[0027] Where, is the missing value at time is the value at the previous time nearest to time and is the value at the next time nearest to time

[0028] Step S202: Data standardization. Normalize and standardize the data, convert the data into a unified dimension, ensure the consistency and comparability of the data, and facilitate model training. This embodiment uses the maximum-minimum normalization:

[0029] Where, is the sum of the dimensions of the indicators in the dataset, is the value after normalization, and are the maximum and minimum values in the dimension where the variable is located.

[0030] ​​Step S203: Data Sampling. The data is sampled to ensure the temporal consistency and integrity of the data, and the data at different time intervals is unified to the same time scale. In this embodiment, all data is unified to the single-day time scale. For data smaller than this scale (process data in Step S1), summation is performed; for data larger than this scale (cost data and operation data in Step S1), it is split into single-day averages.

[0031] Step S204: Data Transformation. The data is transformed into a format suitable for large model training, and the data is numericalized or labeled according to the data type. In this embodiment, all time tags are converted into Unix timestamp values to distinguish the data sources; the basic conditions of each electrolytic cell in Step S1 are labeled. For example, 0 and 1 are used to distinguish whether the electrolytic cell is in normal production status.

[0032] Step S205: Data Verification. The processed data is checked and verified to ensure data quality and applicability.

[0033] Step S3: Train the aluminum electrolysis production management large model of the group based on the industrial basic large model. Select a large model based on the industrial field. You can purchase mature industrial large model products from AI companies, saving the pre-training link of the large model and endowing the large model with the basic ability to analyze industrial data. In this embodiment, the Pangu large model of Huawei is selected, and the large model is trained based on the AI development platform of ModelArts. The historical data processed in Step S2 is used to train the aluminum electrolysis production management large model of the group itself. In this embodiment, functions such as data import, dataset establishment, Pangu basic industrial model, and model custom configuration of the ModelArts platform are used to establish a dataset on this platform and train a large model for the output allocation decision of the electrolysis plants under the group. The trained model is sent to the edge inference server in the group information center through the IEF service.

[0034] Step S4: Collect, integrate, and process the production and operation data of each aluminum electrolysis plant in the recent year. Collect the production and operation data of each aluminum electrolysis plant in the recent year, and use the data processing method in Step S2 to clean, standardize, sample, etc. the data in the recent year. In this embodiment, the processed data is imported into the ModelArts platform to construct a dataset for large model fine-tuning to ensure data quality and applicability.

[0035] Step S5: Fine-tune the large model. Please refer to the specific steps shown in Figure 3 as follows: Step S501: Fine-tuning method selection. According to specific requirements and data characteristics, select an appropriate fine-tuning method, such as LoRA (Low-Rank Adaptation), prompt tuning, transfer learning, or incremental learning. In this embodiment, the method of AI Gallery is selected, which belongs to the fine-tuning technology of transfer learning. The large model is fine-tuned with the dataset of the latest year of the group, so that the knowledge of the large model is closer to the latest situation of the group's production.

[0036] Step S502: Parameter adjustment. Adjust the parameters of the large model according to the selected fine-tuning method. For example, add a low-rank matrix in LoRA, or add specific prompt parameters in prompt tuning. The fine-tuning parameters set by AI Gallery in this embodiment include hyperparameters such as fine-tuning learning rate, batch size, and number of iterations.

[0037] Step S503: Fine-tuning training. Use the fine-tuning dataset constructed in Step S4 to perform fine-tuning training on the large model. Monitor the loss function and performance metrics during the training process to observe the fine-tuning effect.

[0038] Step S504: Performance evaluation. Use the validation set to evaluate the performance of the fine-tuned model, and check the accuracy and generalization ability of the model. In particular, whether the model can output a compliant electrolytic aluminum production allocation plan, and the output result will not deviate from the actual production range of the electrolytic aluminum branch factory.

[0039] Step S505: Model optimization. According to the evaluation results, perform several rounds of model optimization, adjust hyperparameters multiple times, increase the number of training rounds, etc., until satisfactory performance is achieved. The desired result is that both the training loss loss and the validation loss val_loss gradually decrease until they stabilize, and the difference between the two is less than 0.1. When both loss and val_loss are high and the decrease is slow, or val_loss drops to a stable level but the difference from loss is large, it is necessary to reselect the feature data or extend the number of training rounds; when loss decreases slowly and val_loss drops and then stabilizes, it is necessary to adjust the learning rate or add a learning rate scheduler, or an adaptive optimizer can also be used to dynamically adjust the learning rate; when loss continues to decrease, but val_loss drops for a period of time and then stabilizes or even increases, it is necessary to increase regularization or stop training early to prevent overfitting; when both loss and val_loss decrease slowly but do not reach a low level, it is necessary to increase the number of training rounds or use a more efficient optimizer.

[0040] Step S6: Input the production plan for the next year and external condition data into the well-trained and fine-tuned large model. The input data specifically includes: the expected total aluminum production of the group for the next year, the number of normally operating electrolytic cells in each electrolytic aluminum plant, the expected raw material components and contract prices to be purchased by each electrolytic aluminum plant, the normal production quantity of electrolytic cells in each electrolytic aluminum plant, the sales plan, and the production data.

[0041] Step S7: The large model infers the output allocation plan for each electrolytic aluminum plant. The model will, based on the input information: the equipment status, production efficiency, and cost factors of each plant, calculate and infer the output allocation plan for each electrolytic aluminum plant.

[0042] Step S8: Estimate the expected operation situation and economic data, and issue the output allocation decision result after manual review. Use the operation economic model of the factory and the group (cost, revenue, and profit calculation model), and based on the output allocation plan given by the large model, calculate the overall expected operation data of the group for the next year. The specific estimation content includes the expected output, cost, profit, and resource utilization rate of each electrolytic aluminum plant. The content reviewed manually includes the rationality, feasibility, risk control, etc. of the plan. After the review is error-free, the output allocation plan is officially issued to each electrolytic aluminum plant. Each plant confirms the plan and executes the production plan according to the plan.

[0043] The above are only the preferred embodiments of the present invention and are not intended to limit the idea of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for making output allocation decisions for electrolytic aluminum branches under a group based on large models, characterized in that, It includes the following steps in sequence: Step S1: Collect the historical data of each electrolytic aluminum plant under the group; Step S2: Process the historical data collected in Step S1 for use in large model training; Step S3: Select a large model based on the industrial field, and use the data processed in Step S2 to train the large model to form a large model for electrolytic aluminum production management; Step S4: Obtain the production and operation data of electrolytic aluminum from multiple plants within a set time period, and process this data using Step S2 to form a dataset for large model fine-tuning; Step S5: Use the well-processed large model fine-tuning dataset in Step S4 to fine-tune the parameters of the large model. The large model uses a neural network model with more than 100 million parameters; Step S6: Input the production plan and external condition data for the next year into the well-trained and fine-tuned large model; Step S7: Infer the production output allocation plan for each electrolytic aluminum plant through the large model; Step S8: Estimate the expected operation situation and economic data, and issue the production output allocation decision result after manual review.

2. A method for decision-making on the output allocation of electrolytic aluminum branches under a group based on a large model according to claim 1, characterized in that, The historical data collected from each electrolytic aluminum plant under the group in Step S1 includes: process data, quality data, equipment data, raw material composition data, raw material cost, labor cost, manufacturing cost, and operation data in electrolytic production.

3. A method for decision-making on the output allocation of the electrolytic aluminum branch factories under a group based on a large model according to claim 1, characterized in that, The processing of the historical data collected in Step S1 in Step S2 includes: Step S201: Data cleaning, check and process missing values, outliers, and noise data in the data, using methods such as deletion or interpolation; Step S202: Normalize and standardize the data, convert the data to a unified dimension, ensure the consistency and comparability of the data, and facilitate model training; Step S203: Sample the data to ensure the time consistency and integrity of the data, and unify data with different time intervals to the same time scale; Step S204: Convert the data into a format suitable for large model training, and numericalize or label the data according to the data type; Step S205: Re-verify the processed data to ensure data quality and applicability.

4. A decision-making method for output allocation of an electrolytic aluminum branch factory under a group based on a large model according to claim 1, characterized in that, The methods for fine-tuning the parameters of the large model in Step S5 include: Step S501: Fine-tuning method selection, according to specific requirements and data characteristics, select a suitable fine-tuning method, including LoRA (Low-Rank Adaptation), Prompt Tuning, transfer learning, and incremental learning; Step S502: Parameter adjustment, according to the selected fine-tuning method, adjust the parameters of the large model, add a low-rank matrix in LoRA, and add specific prompt parameters in Prompt Tuning; Step S503: Fine-tuning training, use the fine-tuning dataset constructed in Step S4 to perform fine-tuning training on the large model, monitor the loss function and performance metrics during the training process, and observe the fine-tuning effect; Step S504: Performance evaluation, use the validation set to evaluate the performance of the fine-tuned model, check the accuracy and generalization ability of the model, including whether the model can output a compliant electrolytic aluminum production output allocation plan, and the output result will not deviate from the actual production range of the electrolytic aluminum plant. Step S505: Model optimization. Based on the evaluation results, perform several rounds of model optimization, adjust hyperparameters multiple times, and increase the number of training epochs until the training loss (loss) and validation loss (val_loss) are stable and the difference between them is less than 0.

1.

5. A method for decision-making on the output allocation of an electrolytic aluminum branch factory under a group based on a large model according to claim 1, characterized in that, The production plan and external condition data in Step S6 include: the expected total aluminum production of the group in the next year, the number of normal production electrolytic cells in each electrolytic aluminum plant, the expected raw material composition and contract price for procurement in each electrolytic aluminum plant, the normal production quantity of electrolytic cells in each electrolytic aluminum plant, the sales plan, and the production data.

Images