A novel multi-scenario simulation method for dry bulk ports based on a large language model

Abstract

This invention discloses a novel multi-scenario simulation method for a dry bulk port based on a large language model. This method belongs to the field of port production and operation technology. The method utilizes the general knowledge of the large language model to expand training samples. Based on the Adam optimizer, the initial model is trained using a training set and verified using a validation set. The multi-scenario simulation results guided by the model are evaluated based on the test set results. A prompt learning mechanism is used to iteratively expand training with multi-scenario data to obtain a trained large language model. Based on the current port operation scenario, the trained large language model is used to extract key semantics of scenario features. Simulation resource allocation reasoning is performed on the extracted key semantics. AnyLogic simulation is used to construct a new dry bulk port simulation model for the corresponding scenario. The large language model training corpus and training set are updated based on simulation preview results and actual port operation data. This invention can optimize production and operation decisions.

Description

Technical Field

[0001] The present invention belongs to the technical field of port production and operation, and more specifically, relates to a novel multi-scenario simulation method for dry bulk ports based on a large language model. Background Art

[0002] With the rapid growth of global trade, dry bulk ports, as crucial transportation hubs and trading nodes for global food and industrial raw materials, have become increasingly important. Their operational efficiency and management are directly linked to the stability and efficiency of the global supply chain. However, traditional dry bulk port management methods, limited by manual experience, struggle to cope with increasingly complex and changing operating environments and operational demands. Therefore, promoting intelligent port management has become crucial for port development.

[0003] The new type of "bulk-to-container" dry bulk port uses containers to store and transport bulk cargo, and builds horizontal transportation and loading and unloading facilities for bulk containers based on the air-rail collection and distribution system. It can effectively solve the problems of low port circulation and transportation efficiency and heavy port pollution in existing bulk cargo ports, and has good application prospects and market potential.

[0004] As a new technology in natural language processing, the Big Language Model (LLM) leverages deep learning capabilities trained on massive amounts of text data, enabling it to deeply understand textual meaning and handle a variety of complex natural language tasks. Applying LLM to multi-scenario simulations of dry bulk ports effectively addresses operational challenges such as limited historical data and the complex configuration of multi-scenario simulation models, providing new insights for dry bulk port management.

[0005] By collecting data related to port operations, such as equipment performance, operation time, and material attributes, a large language model is trained to reason about resource allocation in multi-scenario port simulations. Based on this data, a virtual port environment is constructed to simulate real-world port operations, enabling multi-scenario simulations. This simulation method provides ports with multi-scenario simulation preview results, enabling intelligent decision-making support for dry bulk ports, such as equipment scheduling and optimized operation plans. This improves port operational efficiency, reduces operating costs, and ensures operational safety.

[0006] In summary, the multi-scenario simulation method for new dry bulk ports based on the large language model can effectively promote the improvement of the intelligent management level and sustainable development of new dry bulk ports with its powerful natural language processing capabilities and intelligent decision-making support.

[0007] Traditional dry bulk ports rely on manual experience for production and operational management, making it difficult to respond promptly to diverse production environments and decision-making requirements. A novel multi-scenario simulation method for dry bulk ports based on a large language model can be applied to new "bulk-to-container" dry bulk ports, promoting the intelligent management level and sustainable development of these ports.

[0008] (1) Data limitations and difficulty in model configuration

[0009] In complex scenarios, traditional dry bulk port operations have limited data collection, making it difficult to ensure the integrity and accuracy of data records, limiting the effectiveness of decision-making models based on data analysis. Leveraging the powerful natural language processing capabilities of large language models, we can overcome data limitations, achieve more refined and comprehensive model configuration, and improve simulation accuracy and generalization capabilities.

[0010] (2) Difficulty in optimizing operational efficiency and resources

[0011] When responding to emergencies, manual scheduling and planning cannot be dynamically adjusted, resulting in low operational efficiency. Furthermore, the lack of a decision support system based on big data analysis leads to a highly subjective decision-making process and weak risk control capabilities. By collecting relevant port operation data, such as equipment performance, operation time, and material attributes, a large language model is trained to perform multi-scenario resource allocation reasoning for port simulations. Based on this data, AnyLogic is used to construct a virtual port environment, simulating real-world port operation scenarios and implementing multi-scenario simulations. This simulation method provides the port with multi-scenario simulation preview results, enabling intelligent decision-making support for dry bulk ports, such as equipment scheduling and operation plan optimization, improving port operational efficiency, reducing operating costs, and ensuring operational safety. Summary of the Invention

[0012] In response to the above-mentioned defects or improvement needs of the existing technology, the present invention proposes a new multi-scenario simulation method for dry bulk ports based on a large language model to solve the production and operation management problems of the new "bulk-to-container" dry bulk ports.

[0013] To achieve the above objectives, the present invention provides a novel multi-scenario simulation method for dry bulk ports based on a large language model, comprising:

[0014] Based on the simulation preview requirements of the new type of bulk-to-container dry bulk port, combined with historical simulation data, port production scenarios and simulation data under multiple operation scenarios are collected as training samples;

[0015] The general knowledge of the large language model is used to expand the training samples, and the expanded training samples are divided into training sets, test sets, and validation sets.

[0016] Based on the Adam optimizer, the initial large language model is trained using the training set, and the initial large language model is verified using the validation set;

[0017] Based on the test results of the test set, the multi-scenario simulation results guided by the large language model are evaluated. Then, using the prompt learning mechanism, iterative expansion training is performed in combination with multi-scenario data to obtain the trained large language model.

[0018] Based on the current port operation scenario, the scenario feature corpus is input and the key semantics of the scenario features are extracted through the trained large language model;

[0019] Use the trained large language model to reason about simulation resource configuration based on the extracted key semantics, thereby obtaining the simulation resource configuration for the current job scenario;

[0020] Based on the simulation resource configuration provided by the trained large language model, AnyLogic simulation software was used to build a new dry bulk port simulation model for the corresponding scenario, providing the port with simulation preview results.

[0021] Update the large language model training corpus and training set based on simulation rehearsal results and actual port operation data.

[0022] In some optional implementation schemes, new dry bulk port operation scenarios include:

[0023] Loading system: It consists of forklifts and yard cranes, responsible for container loading operations;

[0024] Unloading system: It consists of a dumper and a belt conveyor, responsible for unloading bulk cargo from trains;

[0025] Belt conveyor unit: composed of multiple belt lines, responsible for bulk cargo transportation within the yard;

[0026] Stacking system: It consists of a stacker and a belt conveyor, responsible for stacking the unloaded bulk cargo according to demand;

[0027] Packing system: It consists of a belt conveyor and a packing crane and is responsible for the bulk cargo filling operation in containers;

[0028] The SkyRail Collection and Distribution System: It consists of six major systems: track, train, transportation control, power supply and communication, transfer, and information dispatching, and is responsible for the collection, distribution, and loading and unloading of containers within the yard;

[0029] Loading system: It consists of a quay crane and a belt conveyor, responsible for container loading operations;

[0030] Ship unloading system: It consists of a ship unloader and a belt conveyor, and is responsible for bulk cargo unloading operations.

[0031] In some optional implementation schemes, the new dry bulk port multi-scenario simulation targets six sub-operation scenarios and their mixed operation scenarios:

[0032] Train loading operation scenario: filling bulk cargo into containers according to the requirements of arriving trains, and then loading the containers onto trains;

[0033] Train unloading operation scenario: the operation scenario of unloading bulk cargo from the train at the port to the yard;

[0034] Car loading operation scenario: filling bulk cargo into containers according to the demand of arriving cars, and then loading the containers onto cars;

[0035] Truck unloading operation scenario: unloading bulk cargo from a container loaded by a truck at the port to the storage yard;

[0036] Ship loading operation scenario: filling bulk cargo into containers according to the needs of arriving ships and loading the containers onto ships;

[0037] Ship unloading operation scenario: The operation scenario of unloading bulk cargo from a ship in port to the yard.

[0038] In some optional implementation schemes, the port production scenarios and simulation data under multiple operation scenarios collected as training samples based on the dry bulk port simulation preview requirements and combined with historical simulation data include:

[0039] Resource allocation data includes: port structure information, including terminal layout and zoning, berth number and type, sea freight lines, rail freight lines, and truck freight lines; equipment information includes: the number, type, and performance parameters of loading and unloading equipment, belt conveyors, and horizontal transportation equipment; personnel configuration information includes: the number, skill level, and working hours of operators, managers, and maintenance personnel; storage capacity information includes: the storage capacity of the yard;

[0040] Simulation configuration data includes: time parameters including the total simulation run time and time step; input / output data including freight volume, cargo type, arrival and departure time, loading and unloading time; rule strategies including scheduling rules, job priority, resource allocation strategy; model parameters including queue length, storage area and height;

[0041] Simulation performance data includes: throughput, including the amount of cargo handled by the port within the preset time; operation time, including vehicle and ship waiting time, loading and unloading time, and transportation time; resource utilization, including equipment utilization and staffing utilization; congestion conditions, including vehicle / ship queue length and warehouse inventory.

[0042] In some optional implementation schemes, the expansion of training samples using the general knowledge of the large language model includes:

[0043] Leveraging the text generation capabilities of the large language model, descriptive text is generated based on defined operational scenarios and simulation data, covering port operations and simulation data under various operating weather conditions, operating states, and operating time periods.

[0044] Based on the historical simulation data format, descriptive text is converted into structured data suitable for training simulation models. Manual review and simulation model testing are performed to ensure that the generated samples conform to physical laws and port business logic.

[0045] Based on preliminary simulation results and feedback, the input prompts, generation parameters, and data conversion and synthesis rules of the large language model are continuously adjusted.

[0046] In some optional implementation schemes, the training of the initial large language model using the training set and the verification of the initial large language model using the verification set include:

[0047] Divide the training set into multiple subsets according to the set batch size;

[0048] Randomly extract a batch of data from the training set as the input data and corresponding target value of the current iteration;

[0049] Input the current batch of data into the initial large language model and calculate the predicted value through forward propagation;

[0050] Given a target text sequence Y(y1,y2,...y T ), calculate the loss function value based on the predicted value and the true target value: Among them, T is the length of the target text sequence, and the probability distribution predicted by the large language model at each step t is y t is the true value of the target sequence at time t;

[0051] Depend on Calculate the gradient of the loss function with respect to the parameters of the large language model, where is the gradient of the average loss function of the entire sequence with respect to the parameters. For a given target text sequence Y and the probability distribution predicted by the model, θ is the model parameter;

[0052] The update rule of the Adam optimizer is used to update the large language model parameters, where the updated first-order moment estimate is Update the second-order moment estimate to The deviation is corrected to Update parameters to β1 and β2 are hyperparameters, α is the learning rate, t is the current number of iterations, ∈ is a very small positive number, and m t is the first-order exponentially weighted moving average of the gradient, v t is the second-order exponentially weighted moving average of the squared gradient, is the bias-corrected first-order moment, is the bias-corrected second-order moment, θ t+1 To update the parameters;

[0053] After all batches have been iterated, the model is validated using the validation set. If the validation result fails, the model returns to the step of dividing the training set into multiple subsets according to the set batch size until the validation result passes.

[0054] In some optional implementations, the test results based on the test set are used to evaluate the multi-scenario simulation results under the guidance of the large language model, and a prompt learning mechanism is used to perform iterative expansion training in combination with multi-scenario data to obtain a trained large language model, including:

[0055] Use the test set to evaluate the performance of the initial large language model, including model accuracy, stability, and generalization ability;

[0056] Identify model deficiencies based on evaluation results and guide the model's behavior in appropriate scenarios by defining new prompt functions or modifying existing prompts.

[0057] By continuously collecting multi-scenario data, adjusting prompt templates and training strategies, the large language model is trained multiple times.

[0058] In some optional implementations, the large language model is trained multiple times by continuously collecting multi-scenario data, adjusting prompt templates and training strategies, including:

[0059] As iterative training progresses, new multi-simulation scenario data is continuously collected, preprocessed, and annotated;

[0060] Use the designed prompt template to convert the raw data into the form of prompt input;

[0061] Input the prompt input and the corresponding label into the pre-trained large language model for training;

[0062] Based on the difference between the output of the large language model and the true label, the loss function is calculated and the model parameters are updated through backpropagation;

[0063] After each iteration, the performance of the large language model is evaluated using the validation set or test set.

[0064] Adjust the design of the prompt template based on the evaluation results, and adjust the training strategy based on the performance of the large language model during training;

[0065] Continue to iterate, constantly collecting new data, adjusting prompts and training strategies, and iteratively training until the performance evaluation passes.

[0066] In some optional implementation schemes, the inputting of scene feature corpus based on the current port operation scene and the extraction of key semantics of scene features through the trained large language model include:

[0067] Select text data related to the current port operation scenario, remove invalid information from the text data, and convert unstructured data into a structured format;

[0068] The large language model uses advanced semantic analysis to identify and distinguish key words, phrases, and conceptual entities in textual materials, directly mapping them to key parameters and operating modes of port operations.

[0069] Mining the meaning of keywords and phrases in specific scenarios, using contextual information to analyze their functions and relationships, and extracting key semantic information of simulation configurations.

[0070] In some optional implementation schemes, the use of the trained large language model to perform simulation resource configuration reasoning on the extracted key semantics, thereby obtaining the simulation resource configuration in the current operation scenario, includes:

[0071] Based on key semantic information, the trained large language model is used to perform simulation resource configuration reasoning. By analyzing the current operation scenario and constructing the simulation scenario, a reasonable resource allocation plan is inferred, including the quantity allocation of resources such as loading and unloading equipment, transport vehicles, and storage space, operation path planning, and shift scheduling.

[0072] In some optional implementations, the simulation resource configuration provided by the trained large language model is used to construct a new dry bulk port simulation model for the corresponding scenario using AnyLogic simulation software, providing simulation preview results for the port, including:

[0073] Using the simulation resource configuration data provided by the large language model as input, AnyLogic simulation software was used to construct an operational system model of a new dry bulk port, encompassing port infrastructure layout, equipment configuration, and operational processes such as ship arrival, cargo loading and unloading, storage, and transportation.

[0074] Run the simulation model to observe the port's actual operation and collect simulation preview results of operation time, operation efficiency, equipment utilization, and terminal throughput to make production and operation decisions for the port.

[0075] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0076] (1) This paper uses the Adam optimizer to train the initial model and validates the model using a validation set, ensuring the model's training effect and generalization ability. At the same time, it combines the prompt learning mechanism for iterative expansion training, further improving the multi-scenario simulation performance under the guidance of the large language model.

[0077] (2) By combining historical simulation data with the general knowledge of large language models, the present invention effectively addresses the problem of limited port simulation data scale, achieves the expansion of training samples, and significantly improves the richness and accuracy of simulation data. At the same time, the present invention extracts the key semantics of port operation scenarios through large language models and performs simulation resource allocation reasoning based on these key semantics, realizing an intelligent and automated resource allocation process. This improves the efficiency and accuracy of resource allocation, reduces the need for human intervention, and enhances the intelligence level of port production operations.

[0078] (3) Based on the simulation resource configuration provided by the large language model, this paper uses AnyLogic simulation software to construct a new dry bulk port simulation model for the corresponding scenario, providing simulation preview results for the port. This accurately simulates actual port operations, optimizes production and operation decisions, reduces unnecessary resource waste and cost expenditures, and improves the overall operational efficiency of the port. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a flow chart of a novel multi-scenario simulation method for a dry bulk port based on a large language model provided by an embodiment of the present invention;

[0080] Figure 2 This is a schematic diagram of a new dry bulk cargo port operation scenario mainly used in a new dry bulk cargo port multi-scenario simulation method provided by an embodiment of the present invention;

[0081] Figure 3 This is a schematic diagram of initial model training and verification provided by an embodiment of the present invention;

[0082] Figure 4 This is a schematic diagram of an operational system model of a new dry bulk port constructed using AnyLogic simulation software, according to an embodiment of the present invention.

[0083] Figure 5 This is a schematic diagram of an iterative training process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0084] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0085] The technical solution of the present invention is: a new multi-scenario simulation method for dry bulk ports based on a large language model, such as Figure 1 As shown, the specific steps include:

[0086] S1. Aiming at the multi-scenario simulation problem of the new "bulk-to-container" dry bulk port, based on the simulation pre-run requirements of the dry bulk port and combined with historical simulation data, we collected port production scenarios and simulation data under multiple operation scenarios, including resource allocation data, simulation configuration data, and simulation performance data.

[0087] S2. Considering the scale limitations of port multi-scenario simulation data, we leveraged the general knowledge of the large language model to expand the training samples. The expanded training samples were then divided into training, test, and validation sets.

[0088] S3. Based on the Adam optimizer, the initial model is trained using the training set and verified using the validation set.

[0089] S4. Evaluate the multi-scenario simulation results guided by the large language model based on the test set results. Utilize a prompt learning mechanism and iteratively expand training with multi-scenario data to further improve the performance of the multi-scenario simulation under the guidance of the large language model.

[0090] S5. Based on the current port operation scenario, input the scenario feature corpus and extract the key semantics of the scenario features using the trained large language model;

[0091] S6. Use the trained large language model to reason about simulation resource configuration based on the extracted key semantics, thereby obtaining the simulation resource configuration for the current job scenario.

[0092] S7. Based on the simulation resource configuration provided by the large language model, AnyLogic simulation software was used to construct a new dry bulk port simulation model for the corresponding scenario. This provided the port with simulation preview results to support decision-making in port production and operations.

[0093] S8. Update the large language model training corpus and training set based on the simulation results and actual port operation data to conduct a new round of simulation rehearsals.

[0094] In the above example, if Figure 2 As shown in FIG, the multi-scenario simulation method of the new dry bulk cargo port is mainly applied to the new dry bulk cargo port operation scenario, wherein the new dry bulk cargo port operation scenario includes the following main components:

[0095] Loading system: It consists of forklifts and yard cranes, responsible for container loading operations;

[0096] Unloading system: It consists of a dumper and a belt conveyor, responsible for unloading bulk cargo from trains;

[0097] Belt conveyor unit: composed of multiple belt lines, responsible for bulk cargo transportation within the yard;

[0098] Stacking system: It consists of a stacker and a belt conveyor, responsible for stacking the unloaded bulk cargo according to demand;

[0099] Packing system: It consists of a belt conveyor and a packing crane and is responsible for the bulk cargo filling operation in containers;

[0100] The SkyRail Collection and Distribution System: It consists of six major systems: track, train, transportation control, power supply and communication, transfer, and information dispatching, and is responsible for the collection, distribution, and loading and unloading of containers within the yard;

[0101] Loading system: It consists of a quay crane and a belt conveyor, responsible for container loading operations;

[0102] Ship unloading system: It consists of a ship unloader and a belt conveyor, and is responsible for bulk cargo unloading operations.

[0103] In the above example, the multi-scenario simulation of the new dry bulk port mainly targets six sub-operation scenarios and their mixed operation scenarios:

[0104] Train loading operation scenario: filling bulk cargo into containers according to the requirements of arriving trains, and then loading the containers onto trains;

[0105] Train unloading operation scenario: the operation scenario of unloading bulk cargo from the train at the port to the yard;

[0106] Car loading operation scenario: filling bulk cargo into containers according to the demand of arriving cars, and then loading the containers onto cars;

[0107] Truck unloading operation scenario: unloading bulk cargo from a container loaded by a truck at the port to the storage yard;

[0108] Ship loading operation scenario: filling bulk cargo into containers according to the needs of arriving ships and loading the containers onto ships;

[0109] Ship unloading operation scenario: The operation scenario of unloading bulk cargo from a ship in port to the yard.

[0110] In the above example, in step S1, based on the dry bulk port simulation preview requirements and combined with historical simulation data, the port production scenarios and simulation data collected under multiple operation scenarios mainly include:

[0111] Resource allocation data: port structure information, including terminal layout and zoning, berth number and type, sea freight lines, railway freight lines, automobile freight lines, etc.; equipment information, including the number, type, and performance parameters of loading and unloading equipment, belt conveyors, and horizontal transportation equipment; personnel configuration information, including the number, skill level, and working hours of operators, managers, and maintenance personnel; storage capacity information, including the storage capacity of the yard.

[0112] Simulation configuration data: time parameters, including the total simulation run time and time step; input / output data, including freight volume, cargo type, arrival and departure times, loading and unloading times; rule strategies, including scheduling rules, job priority, resource allocation strategy, etc.; model parameters, including queue length, storage area and height, etc.

[0113] Simulation performance data: throughput, the amount of cargo handled by the port within a certain period of time; operation time, including vehicle and vessel waiting time, loading and unloading time, and transportation time; resource utilization, including equipment utilization and staffing utilization; congestion conditions, including vehicle / vessel queue length and warehouse inventory.

[0114] In the above example, in step S2, the general knowledge of the large language model is used to expand the training samples and divide the expanded training samples. The training sample expansion process is as follows:

[0115] S21. Leverage the text generation capabilities of the large language model to generate a large amount of descriptive text based on defined operational scenarios and simulation data, covering port operations and simulation data under various operational weather conditions, operational states, and operational time periods.

[0116] S22. Based on the historical simulation data format, convert descriptive text into structured data suitable for training the simulation model. Through manual review and simulation model testing, ensure that the generated samples conform to physical laws and port business logic, and have practical value and credibility.

[0117] S23. Based on preliminary simulation results and feedback, continuously adjust the input prompts, generation parameters, and data conversion and synthesis rules of the large language model to improve the quality and applicability of the expanded samples through multiple iterations.

[0118] In the above example, if Figure 3 As shown, in step S3, the steps of using the training set to train the initial model and using the validation set to validate the initial model are as follows:

[0119] S31. A training loop that divides the training dataset into multiple subsets according to a set batch size;

[0120] S32. Data sampling: randomly extract a batch of data from the training set as the input data and corresponding target value of the current iteration;

[0121] S33. Forward propagation: input the current batch of data into the model and calculate the predicted value through forward propagation;

[0122] S34. Calculate the loss, given the target text sequence Y(y1,y2,...y T), calculate the loss function value based on the predicted value and the true target value;

[0123]

[0124] Where T is the sequence length, and the probability distribution predicted by the model at each step t is

[0125] S35. Back propagation, calculating the gradient of the loss function with respect to the model parameters;

[0126]

[0127] S36. Update parameters. Use the update rule of the Adam optimizer to update the model parameters, including calculating the first-order moment estimate and the second-order moment estimate, correcting the deviation, and updating the model parameters accordingly.

[0128] Update first-order moment estimate

[0129] Updated second-order moment estimates

[0130] Bias Correction

[0131] Update parameters

[0132] Among them, β1 and β2 are hyperparameters, α is the learning rate, t is the current number of iterations, and ∈ is a very small positive number.

[0133] S37. Validation set verification: After all batches have been iterated, the model is verified using the validation set.

[0134] S38. Check the result. If the verification result fails, jump to step S31 and continue executing until the verification result passes.

[0135] In the above example, in step S4, the steps of evaluating the multi-scenario simulation results under the guidance of the large language model based on the test results of the test set and performing iterative expansion training using the prompt learning mechanism are as follows:

[0136] S41. Test result evaluation: using the test set, perform performance evaluation on the initial model, including model accuracy, stability, and generalization ability;

[0137] S42. Prompt learning mechanism: identifies model deficiencies based on evaluation results and guides the model's behavior in corresponding scenarios by defining new prompt functions or modifying existing prompts.

[0138] S43. Iterative training process: By continuously collecting multi-scenario data, adjusting prompt templates and training strategies, the large language model is trained multiple times to improve the multi-scenario simulation performance under the guidance of the large language model.

[0139] In the above example, in step S5, based on the current port operation scene, scene feature corpus closely related to the operation is collected and preprocessed, and the trained large language model is used to perform in-depth analysis and semantic understanding on the corpus to extract key semantic information of the scene features. The steps are as follows:

[0140] S51. Collect scenario-specific corpus and select textual materials related to the current port operation scenario, including detailed descriptions of cargo type, loading and unloading operations, equipment usage, staffing, environmental conditions, etc.

[0141] S52. Data preprocessing: removing invalid information from the corpus to ensure data quality and converting unstructured data into a structured format to facilitate model understanding.

[0142] S53. Semantic analysis: The model uses advanced semantic analysis to identify and distinguish key words, phrases, and conceptual entities in the text, directly mapping them to key parameters and operating modes of port operations.

[0143] S54. Key information extraction: Deeply explore the meaning of keywords and phrases in specific scenarios, use contextual information to analyze their functions and relationships, and extract key semantic information of simulation configurations.

[0144] In the above example, in step S6, the trained large language model is used to perform simulation resource allocation reasoning based on the key semantic information. Through in-depth analysis of the current operation scenario and the construction of the simulation scenario, a reasonable resource allocation plan is inferred, including the allocation of resources such as loading and unloading equipment, transportation vehicles, and storage space, as well as operation route planning and shift scheduling.

[0145] In the above example, if Figure 4 As shown, in step S7, the simulation resource configuration data provided by the large language model is used as input to construct an operational system model of a new dry bulk port using AnyLogic simulation software. This model includes the port's infrastructure layout, equipment configuration, and operational processes such as ship arrival, cargo loading and unloading, storage, and transportation. The simulation model is run to observe the port's actual operations and collect simulation results such as operating time, efficiency, equipment utilization, and terminal throughput. This allows for precise production and operational decisions, such as optimizing resource allocation and adjusting operational processes, to maximize port operational efficiency.

[0146] In the above example, if Figure 5 As shown, the iterative training process steps are as follows:

[0147] S431. Collect and annotate data. As iterative training progresses, new multi-simulation scenario data is continuously collected, preprocessed, and annotated.

[0148] S432. Construct prompt input, use the designed prompt template to convert the original data into the form of prompt input;

[0149] S433. Model training: input the prompt input and the corresponding label into the pre-trained model for training;

[0150] S434. Update parameters, calculate the loss function based on the difference between the model output and the true label, and update the model parameters through back propagation;

[0151] S435. Evaluate model performance. After each iteration, use the validation set or test set to evaluate the performance of the model.

[0152] S436. Adjust prompts and training strategies. Based on the evaluation results, adjust the design of the prompt template to better guide the model to focus on key information. Also, adjust the training strategies such as learning rate and batch size based on the performance of the model during training.

[0153] S437. Continue iterating and repeating the above steps, constantly collecting new data, adjusting prompts and training strategies, and performing iterative training until the performance evaluation passes.

[0154] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, or two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0155] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A new multi-scenario simulation method for dry bulk ports based on a large language model, characterized by: include: Based on the simulation preview requirements of the new type of bulk-to-container dry bulk port, combined with historical simulation data, port production scenarios and simulation data under multiple operation scenarios are collected as training samples; The general knowledge of the large language model is used to expand the training samples, and the expanded training samples are divided into training sets, test sets, and validation sets. Based on the Adam optimizer, the initial large language model is trained using the training set, and the initial large language model is verified using the validation set; Based on the test results of the test set, the multi-scenario simulation results guided by the large language model are evaluated. Then, using the prompt learning mechanism, iterative expansion training is performed in combination with multi-scenario data to obtain the trained large language model. Based on the current port operation scenario, the scenario feature corpus is input and the key semantics of the scenario features are extracted through the trained large language model; Use the trained large language model to reason about simulation resource configuration based on the extracted key semantics, thereby obtaining the simulation resource configuration for the current job scenario; Based on the simulation resource configuration provided by the trained large language model, AnyLogic simulation software was used to build a new dry bulk port simulation model for the corresponding scenario, providing the port with simulation preview results. Update the large language model training corpus and training set based on simulation rehearsal results and actual port operation data.

2. The method according to claim 1, characterized in that New dry bulk port operation scenarios include: Loading system: It consists of forklifts and yard cranes, responsible for container loading operations; Unloading system: It consists of a dumper and a belt conveyor, responsible for unloading bulk cargo from trains; Belt conveyor unit: composed of multiple belt lines, responsible for bulk cargo transportation within the yard; Stacking system: It consists of a stacker and a belt conveyor, responsible for stacking the unloaded bulk cargo according to demand; Packing system: It consists of a belt conveyor and a packing crane and is responsible for the bulk cargo filling operation in containers; The SkyRail Collection and Distribution System: It consists of six major systems: track, train, transportation control, power supply and communication, transfer, and information dispatching, and is responsible for the collection, distribution, and loading and unloading of containers within the yard; Loading system: It consists of a quay crane and a belt conveyor, responsible for container loading operations; Ship unloading system: It consists of a ship unloader and a belt conveyor, responsible for bulk cargo unloading operations; The new dry bulk port multi-scenario simulation targets six sub-operation scenarios and their mixed operation scenarios: Train loading operation scenario: filling bulk cargo into containers according to the requirements of arriving trains, and then loading the containers onto trains; Train unloading operation scenario: the operation scenario of unloading bulk cargo from the train at the port to the yard; Car loading operation scenario: filling bulk cargo into containers based on the demand of arriving cars, and then loading the containers onto cars; Truck unloading operation scenario: unloading bulk cargo from a truck-loaded container at the port to the storage yard; Ship loading operation scenario: filling bulk cargo into containers according to the needs of arriving ships and loading the containers onto ships; Ship unloading operation scenario: The operation scenario of unloading bulk cargo from a ship at port to the yard.

3. The method according to claim 1 or 2, characterized in that According to the simulation preview requirements of the new type of bulk-to-container dry bulk port, combined with historical simulation data, port production scenarios and simulation data under multiple operation scenarios are collected as training samples, including: Resource allocation data includes: port structure information, including terminal layout and zoning, berth number and type, sea freight lines, rail freight lines, and truck freight lines; equipment information includes: the number, type, and performance parameters of loading and unloading equipment, belt conveyors, and horizontal transportation equipment; personnel configuration information includes: the number, skill level, and working hours of operators, managers, and maintenance personnel; storage capacity information includes: the storage capacity of the yard; Simulation configuration data includes: time parameters including the total simulation run time and time step; input / output data including freight volume, cargo type, arrival and departure time, loading and unloading time; rule strategies including scheduling rules, job priority, resource allocation strategy; model parameters including queue length, storage area and height; Simulation performance data includes: throughput, including the amount of cargo handled by the port within the preset time; operation time, including vehicle and ship waiting time, loading and unloading time, and transportation time; resource utilization, including equipment utilization and staffing utilization; congestion conditions, including vehicle / ship queue length and warehouse inventory.

4. The method according to claim 3, characterized in that The method of expanding training samples by utilizing the general knowledge of the large language model includes: Leveraging the text generation capabilities of the large language model, descriptive text is generated based on defined operational scenarios and simulation data, covering port operations and simulation data under various operating weather conditions, operating states, and operating time periods. Based on the historical simulation data format, descriptive text is converted into structured data suitable for training simulation models. Manual review and simulation model testing are performed to ensure that the generated samples conform to physical laws and port business logic. Based on preliminary simulation results and feedback, the input prompts, generation parameters, and data conversion and synthesis rules of the large language model are continuously adjusted.

5. The method according to claim 4, characterized in that The training of the initial large language model using the training set and the verification of the initial large language model using the verification set include: Divide the training set into multiple subsets according to the set batch size; Randomly extract a batch of data from the training set as the input data and corresponding target value of the current iteration; Input the current batch of data into the initial large language model and calculate the predicted value through forward propagation; Given a target text sequence , calculate the loss function value based on the predicted value and the true target value: , where T is the target text sequence length, and the probability distribution predicted by the large language model at each step t is , is the true value of the target sequence at time t; Depend on Calculate the gradient of the loss function with respect to the parameters of the large language model, where is the gradient of the average loss function of the entire sequence with respect to the parameters. For a given target text sequence Y and the probability distribution predicted by the model, , are model parameters; The update rule of the Adam optimizer is used to update the large language model parameters, where the updated first-order moment estimate is , update the second-order moment estimate to , the deviation is corrected to , update the parameters to , is a hyperparameter, is the learning rate, t is the current number of iterations, is a very small positive number, is the first-order exponentially weighted moving average of the gradient, is the second-order exponentially weighted moving average of the squared gradient, is the bias-corrected first-order moment, is the bias-corrected second-order moment, To update the parameters; After all batches have been iterated, the model is validated using the validation set. If the validation result fails, the model returns to the step of dividing the training set into multiple subsets according to the set batch size until the validation result passes.

6. The method according to claim 5, characterized in that The test results based on the test set are used to evaluate the multi-scenario simulation results under the guidance of the large language model, and the prompted learning mechanism is used to perform iterative expansion training in combination with multi-scenario data to obtain the trained large language model, including: Use the test set to evaluate the performance of the initial large language model, including model accuracy, stability, and generalization ability; Identify model deficiencies based on evaluation results and guide the model's behavior in appropriate scenarios by defining new prompt functions or modifying existing prompts. By continuously collecting multi-scenario data, adjusting prompt templates and training strategies, the large language model is trained multiple times.

7. The method according to claim 6, characterized in that The large language model is trained multiple times by continuously collecting multi-scenario data, adjusting prompt templates and training strategies, including: During each iterative training process, new multi-simulation scenario data is continuously collected, pre-processed and annotated; Use the designed prompt template to convert the raw data into the form of prompt input; Input the prompt input and the corresponding label into the pre-trained large language model for training; Based on the difference between the output of the large language model and the true label, the loss function is calculated and the model parameters are updated through backpropagation; After each iteration, the performance of the large language model is evaluated using the validation set or test set. Adjust the design of the prompt template based on the evaluation results, and adjust the training strategy based on the performance of the large language model during training; Continue to iterate, constantly collecting new data, adjusting prompts and training strategies, and iteratively training until the performance evaluation passes.

8. The method according to claim 7, characterized in that According to the current port operation scenario, the scenario feature corpus is input and the key semantics of the scenario features are extracted through the trained large language model, including: Select text data related to the current port operation scenario, remove invalid information from the text data, and convert unstructured data into a structured format; The large language model uses advanced semantic analysis to identify and distinguish key words, phrases, and conceptual entities in textual materials, directly mapping them to key parameters and operating modes of port operations. Mining the meaning of keywords and phrases in specific scenarios, using contextual information to analyze their functions and relationships, and extracting key semantic information of simulation configurations.

9. The method according to claim 8, characterized in that The trained large language model is used to perform simulation resource configuration reasoning on the extracted key semantics, thereby obtaining the simulation resource configuration for the current operation scenario, including: Based on key semantic information, the trained large language model is used to perform simulation resource configuration reasoning. By analyzing the current operation scenario and constructing the simulation scenario, a reasonable resource allocation plan is inferred, including the quantity allocation of resources such as loading and unloading equipment, transport vehicles, and storage space, operation path planning, and shift scheduling.

10. The method according to claim 9, characterized in that Based on the simulation resource configuration provided by the trained large language model, AnyLogic simulation software was used to construct a new dry bulk port simulation model for the corresponding scenario, providing the port with simulation preview results, including: Using the simulation resource configuration data provided by the large language model as input, AnyLogic simulation software was used to construct an operational system model of a new dry bulk port, encompassing port infrastructure layout, equipment configuration, and operational processes such as ship arrival, cargo loading and unloading, storage, and transportation. Run the simulation model to observe the port's actual operation and collect simulation preview results of operation time, operation efficiency, equipment utilization, and terminal throughput to make production and operation decisions for the port.

Images