Transport vehicle scheduling method and system based on large language model, and storage medium

Abstract

The invention belongs to the field of transport scheduling, and particularly discloses a transport vehicle scheduling method and system based on a large language model and a storage medium, and the method comprises the steps: obtaining order information through the large language model based on an adjusted order text; acquiring required vehicle information based on the order information, and packaging the order information and the required vehicle information into a transportation application; a plurality of current transportation applications are sequenced, transportation files are formed according to vehicle conditions, and one transportation file corresponds to one vehicle; screening the transportation files to obtain the transportation files meeting the transportation conditions; combining the screened transportation files to obtain a plurality of transportation schemes, and screening the transportation schemes according to a preset target to obtain K transportation schemes; and according to the real-time text indication, determining a final transportation scheme from the K transportation schemes through a large language model. According to the invention, the transport plan of the transport vehicle can be adjusted in real time according to the change of order information in the supply chain, and the influence caused by plan adjustment is minimized.

Description

Technical Field

[0001] The present invention belongs to the field of transportation scheduling, and more specifically, relates to a transportation vehicle scheduling method, system and storage medium based on a large language model. Background Art

[0002] Logistics and transportation management is a major challenge for modular integrated buildings. This includes ensuring the timely and cost-effective delivery of prefabricated components and modules, optimizing routes to reduce transportation costs and environmental impact, coordinating large prefabricated modules and effectively managing their transportation. In this supply chain, there are complex interactions between all participants and stakeholders, including component suppliers, MIC factories, carriers, and construction parties. On this basis, the inherent uncertainty and variability in building logistics make it necessary to adjust logistics and transportation plans in real time to improve the flexibility of the supply chain.

[0003] Many studies have explored the challenges of off-site construction logistics planning, including MIC. These studies mainly focus on a single supply chain stage, involving problems such as vehicle routing and integrated truck planning. Some researchers have proposed optimization models that integrate different supply chain stages, such as combining production with logistics and combining logistics with installation. These integrated approaches aim to provide more comprehensive solutions, but often ignore uncertainty and dynamic interactions, which are crucial in practical applications. To address these limitations, some researchers have adopted the optimization of MIC transportation logistics by using AnyLogic for multi-agent systems, which simulates the interactions between agents involved in the MIC process to improve transportation efficiency and reduce costs. Most of the current methods are only used for planning before the start of transportation logistics, and cannot discuss and make decisions on the impact of unexpected events during the transportation process on the plan. Even the method of using AnyLogic for multi-agent systems to optimize MIC transportation logistics mainly relies on heuristic rules in its negotiation mechanism, which shows that it is not smart enough in selecting the optimal or suitable machine based on various information about orders and transportation vehicles. Large language models (LLMs) provide a feasible way to enhance the capabilities of agents in multi-agent manufacturing systems. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention provides a transport vehicle scheduling method, system and storage medium based on a large language model, the purpose of which is to achieve real-time updating of the transport vehicle scheduling plan based on order information adjustment and reduce operating costs.

[0005] To achieve the above object, according to a first aspect of the present invention, a transportation vehicle scheduling method based on a large language model is proposed, comprising the following steps:

[0006] Based on the adjusted order text, the order information is obtained through a large language model;

[0007] Obtain required vehicle information based on order information, and package the order information and required vehicle information into a transportation application;

[0008] Sort the current several transport applications and generate transport documents according to the vehicle conditions. One transport document corresponds to one vehicle.

[0009] Screen the transportation documents to obtain all transportation documents that meet the transportation conditions; combine the screened transportation documents to obtain multiple transportation plans, and screen the transportation plans according to the preset goals to obtain K transportation plans;

[0010] According to the textual indication of real-time demand, the final transportation plan is determined from K transportation plans through a large language model.

[0011] As a further preferred embodiment, before using the large language model, a target instruction set including a task description text and a corresponding vehicle transportation plan is pre-built; then the large language model is fine-tuned using the target instruction set, that is, the large language model is trained with the task description text as input and the vehicle transportation plan as output.

[0012] As a further preferred embodiment, the target instruction set is constructed as follows:

[0013] Collect the task description text and the corresponding vehicle transportation plan, and use the natural language task description to format the collected data. Then clean the formatted data to obtain the target instruction set.

[0014] As a further preferred method, the formatted data is cleaned, including deduplication, correction and complication, specifically:

[0015] The cosine similarity algorithm is used to calculate the semantic similarity of the task description text in the target instruction set, and some instances with high repetition are deleted or semantically similar instances are merged into a new instance. Then, errors in the data are detected and corrected through syntax analysis models, manual spot checks and script verification. Finally, some simple tasks are rewritten into complex tasks.

[0016] As a further preference, the transportation document includes vehicle number, vehicle capacity, vehicle cost, estimated time, target order time, and start and end times corresponding to two adjacent orders.

[0017] As further preferred, the transportation conditions include a vehicle departure time constraint, a vehicle estimated arrival time constraint, and a latest end time constraint, wherein:

[0018] The vehicle departure time constraint is:

[0019] t ei <t O(i-1) <t li

[0020] The vehicle estimated arrival time constraint is:

[0021] t Ei <t O(i-1) +T i <t Li

[0022] The latest end time constraint is:

[0023] t O(i-1) +T i <t S(i+1)

[0024] Among them, t ei ,t li are the earliest delivery time and the latest delivery time of the i-th order respectively; t Ei ,t Li are the earliest entry time and the latest entry time of the i-th order respectively; t O(i-1) is the estimated end time of the i-1th order, T i is the estimated time it takes for the vehicle to complete the i-th order, t S(i+1) is the start time of the i+1th order.

[0025] As a further preferred method, the screened transport documents are combined to obtain multiple transport solutions, and the transport solutions are screened according to a preset target to obtain K transport solutions, including:

[0026] Based on the transportation cost, a plurality of transportation plans obtained by combining the transportation documents are preliminarily screened to obtain L transportation plans with lower transportation costs;

[0027] Then the transportation time of the L transportation plans is calculated respectively, and K transportation plans with less transportation time are further screened.

[0028] As a further preferred method, a plurality of transportation schemes obtained by combining the transportation documents are preliminarily screened based on the transportation cost, specifically:

[0029] According to the cargo volume, the selected transport documents are combined through the DP matrix, and the loss of each transport plan is calculated. The top N transport plans with the smallest loss value are initially screened;

[0030] The values ​​in the DP matrix represent the loss, and its 0th row is initialized to the maximum cost, that is, DP 0,j =P, the maximum cost P is defined as the value greater than the sum of the transportation costs of all vehicles; when l≥1, the value DP in the lth row and jth columnl,j It represents the minimum loss when the first l vehicles are included in the combination and the vehicle combination transports j tons of goods. The calculation formula is:

[0031] DP l,j =min(DP l-1,j-pl +C l -P×min(p l ,j),DP l-1,j )

[0032] Among them, p l represents the capacity of the lth vehicle, C l Represents the transportation cost of the lth vehicle.

[0033] According to a second aspect of the present invention, a system for implementing the above-mentioned transportation vehicle scheduling method based on a large language model is provided, comprising a decision agent, an order agent, a scheduling agent and a vehicle agent, wherein:

[0034] The decision agent is implemented based on a large language model, and is used to form order information based on the adjusted order text and send it to the order agent;

[0035] The order agent is used to obtain the required vehicle information based on the order information, and package the order information and the required vehicle information into a transportation application and send it to the dispatch agent;

[0036] The dispatch agent is used to sort the transport applications from the order agents and form transport invitations to be sent to all vehicle agents;

[0037] The vehicle agent is used to generate a transportation document based on the vehicle situation after receiving the transportation invitation and send it to the dispatch agent, where one transportation document corresponds to one vehicle;

[0038] The dispatch agent is also used to screen the transport documents to obtain all transport documents that meet the transport conditions; combine the screened transport documents to obtain multiple transport plans, and screen the transport plans according to the preset goals to obtain K transport plans, which are sent to the decision agent;

[0039] The decision agent is also used to determine the final transportation plan from K transportation plans according to the text indication of real-time demand, so as to form a decision and send it to the scheduling agent and the order agent. The order agent completes its own information update, and the scheduling agent sends the decision to the vehicle agent for execution.

[0040] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a transportation vehicle scheduling method based on a large language model.

[0041] In general, the above technical solution conceived by the present invention has the following technical advantages compared with the prior art:

[0042] 1. When the existing methods are used for transport vehicle scheduling management, it is assumed that the transport demand for modules and components is fixed, which cannot fully reflect the changes and uncertainties in the real world. The present invention focuses on the need for real-time adjustment of transport plans due to order delivery delays caused by variable module sizes and cost fluctuations required by suppliers and MIC factory projects. The vehicle scheduling method is designed based on a large language model. During the entire transport vehicle scheduling management process, the transport plan of the transport vehicle can be adjusted in real time according to changes in order information in the supply chain to minimize the impact of plan adjustments; that is, the transport arrangement can be adjusted in real time by changing the order information at any time to reduce the operating costs of the carrier.

[0043] 2. The present invention designs a method for screening transportation plans based on the DP matrix, which can obtain the optimal vehicle combination loss under the same cargo carrying capacity, different transport capacities and conditions, thereby screening out the top L transportation plans with the lowest loss values; and further combine the screening of the top K transportation plans with the lowest time consumption, so as to quickly obtain a transportation plan with lower transportation cost and time consumption.

[0044] 3. The present invention designs a multi-agent transport vehicle scheduling system, which facilitates the entire process from control to decision-making by defining agents for each object of logistics scheduling. At the same time, these agents also act as a channel between the multi-agent scheduling system and the large language model LLM, greatly reducing the complexity of modifying other conditions while improving system performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram of a transportation vehicle scheduling process based on a large language model according to an embodiment of the present invention;

[0046] Figure 2 It is a physical implementation framework diagram of a transportation vehicle scheduling system based on a large language model according to an embodiment of the present invention;

[0047] Figure 3 This is a flow chart of a method for transport vehicle scheduling based on a large language model according to an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of LLM text indication according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain 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 can be combined with each other as long as they do not conflict with each other.

[0050] The embodiment of the present invention provides a transport vehicle scheduling system based on a large language model, which determines the transport vehicles for the completed order components or modules. In this regard, the earliest arrival time and the latest arrival time of the components and modules are the requirements of the supplier and the MIC factory, and the minimum cost and the shortest transportation time of the transport plan are the embodiment of the interests of the carrier. In addition, for transportation, it is necessary to clarify that all orders must be completed before they can be transported in batches, and some components cannot be transported during the production of the order.

[0051] The multi-agent-based system of the present invention includes various entities involved in the MIC supply chain process, and each entity is simulated as an agent with specific behaviors and interactions. In this embodiment, for the convenience of explanation, it is assumed that there are N suppliers and M MIC factories. The physical process first starts with suppliers who produce prefabricated components (such as beams, exterior walls and laminates). These components are produced in the factories of several suppliers, and the design of these factories can handle large-scale production with high precision. After production is completed, these components are transported to several MIC factories by transport vehicles arranged by carriers. In these MIC factories, the prefabricated components are assembled into integrated building modules. The system is mainly aimed at the transportation stage of components from suppliers to MIC assembly factories, focusing on solving the problem of combining different transport vehicles with reasonable vehicle costs and transportation time when changing the existing transportation arrangement plan under the premise of meeting the delivery time of MIC factories and construction sites, so as to meet the minimum operating cost of the carrier.

[0052] The transportation vehicle scheduling system based on the large language model includes decision-making agent, order agent, scheduling agent, and vehicle agent. The functions of each agent are as follows:

[0053] 1. Decision-making Agent

[0054] This agent is an instance of a large language model. It is trained after receiving prompt words and has a certain degree of human intelligence. It can receive text information, process information, and output dialogues through text. Its inherent intelligence comes from the large language model, which is called by prompts. The role of decision-making is to make decisions based on the transportation plans received from the dispatcher, select the most appropriate transportation plan, and send it to the dispatcher for execution. It includes the following functions:

[0055] (1) Accept natural language input, convert it into order information data, send it to the order agent to update the order information, and trigger the system operation trigger;

[0056] (2) Accept the feasible transport vehicle set from the scheduling agent, and judge the value of each transportation plan according to the minimum vehicle idle time logic, and select the one with the highest value as the decision and send it to the scheduling agent.

[0057] 2. Order Agent

[0058] The following parameters are included:

[0059] Component order information: order number i, starting point (supplier) number N, end point (MIC factory) number M, earliest delivery time t ei , Latest delivery time t li , the earliest order entry time t Ei , the latest order entry time t Li , Transportation time T N , Estimated start time t SN , Estimated end time t ON 、Cargo volume M i , and the corresponding transport vehicle number Q.

[0060] Order information summary table: order number i, transported goods quantity m i And the corresponding budget vehicle number Q.

[0061] The features include:

[0062] (1) Receive the latest component order information sent by the decision agent and update the component order information list;

[0063] (2) Select the component orders that have changed, sort out the packaging and transportation requirements, and then send the transportation request to the dispatcher;

[0064] 3. Scheduling Agent

[0065] The features include:

[0066] (1) Accepting transport applications, sorting out the transport requirements therein, and sending a transport invitation containing transport requirement information to the vehicle agent.

[0067] (2) Arrange the transport documents sent by the vehicle agents, filter them, form a feasible transport vehicle set consisting of several transport documents, and send this information set to the decision agent. After receiving the decision sent by the decision agent, generate a transport work order and send the work order to all corresponding vehicle agents.

[0068] 4. Vehicle agency

[0069] It consists of all vehicles owned by the carrier, and each vehicle is an agent individual.

[0070] Parameters: vehicle number Q, vehicle capacity mQ , vehicle expenses (i.e. transportation costs) C Q , Q vehicle work order task list (work order number i, starting number N, end number M, start time t Si , estimated time T i , expected end time t Oi ).

[0071] The features include:

[0072] (1) Based on the order information, clear the orders with the same order number from the vehicle work order task list and send the transportation documents to the dispatch agent.

[0073] (2) After accepting the dispatch agent's transport work order, update the work order task list.

[0074] (3) Based on the starting point number, end point number and own location, the SUMO software can be used to calculate in real time the estimated time T for the corresponding starting point N and end point MD of the own work order task i , expected end time t Oi .

[0075] The embodiment of the present invention provides a transportation vehicle scheduling method based on a large language model, such as Figure 3 As shown in the figure, the agent negotiation process is as follows Figure 1 As shown, the specific steps include:

[0076] S1: Fine-tune the large-scale language model in advance and use the fine-tuned language model as the decision agent. It includes the following steps:

[0077] S11: Collect data, including task description text and corresponding vehicle transportation plan. The task description text is specifically a component order description text and corresponding component order information parameters, and an information set of different transportation vehicle plans.

[0078] Specifically, data can be crawled on the Internet, different vehicle transportation plans can be obtained from papers related to transportation vehicle scheduling management, and corresponding plan description texts can be written or modified. The user input, i.e., the task description text, and the expected output, i.e., the vehicle transportation plan, can be paired for subsequent training.

[0079] S12: Process the collected data to form a target instruction set, specifically formatting and cleaning the collected data.

[0080] Specifically, the data crawled from the web and the data filtered from the data set are formatted using natural language task descriptions, fine-tuning instances and instruction sets are constructed, and then data cleaning is performed. Data cleaning is based on the existing fine-tuning data and performs multi-stage data deduplication, correction, and complication, including: first, using the cosine similarity algorithm to calculate the semantic similarity of the task description text in the instruction set, deleting instances with high repetition or merging semantically similar instances into a new fine-tuning instance; then, errors in the code are detected and corrected through syntax analysis models, manual spot checks, and script verification; finally, a small number of simple tasks are rewritten into complex tasks to improve the knowledge understanding ability and spatial reasoning ability of large-scale language models in the field of building assembly.

[0081] S13: Fine-tune the instructions of the large language model using the target instruction set.

[0082] S2: Based on the adjustment of order information by component suppliers and MIC factories, the decision agent sends order information to the order agent, and the order agent sends the transportation application to the dispatch agent. It includes the following steps:

[0083] S21: The decision agent converts the natural language about order adjustment into component order information (including order number i, earliest delivery time t, etc.) through the fine-tuned large language model. ei , Latest delivery time t li , the earliest order entry time t Ei , the latest order entry time t Li 、Cargo volume M i ), the changed information objects are the earliest delivery time, the latest delivery time, the earliest entry time, the latest entry time, and the transported goods quantity. After sending the changed component order information to the order agent with the change information, the order agent updates the order information and generates the transport application.

[0084] Specifically, in this step, information about the changed component orders is received. The orders are placed in a queue, indicating that multiple orders can be managed simultaneously. The purpose of this step is to update the changed orders and sort them in the order list (determine the priority by the order delivery time) so that the components can be produced and delivered.

[0085] S22: The order agent will index the order information table based on the received change order and obtain the original component order information of the order; then the change component order information and the required vehicle information will be integrated and packaged as a transportation application and sent to the dispatch agent.

[0086] Specifically, in this step, the order number of the changed order is indexed in the order information table, the corresponding transport vehicle number Q is obtained, and it is integrated with the changed order information to package the order's transportation needs (order number i, starting point number N, destination number M, order earliest delivery time t ei , Latest delivery time t li , the earliest order entry time t Ei , the latest order entry time t Li 、Cargo volume M i ) and sent to the dispatch agent as a transportation request.

[0087] S3: After the dispatch agent sorts out the transport application information, it sends a transport invitation to the vehicle agents. The vehicle agents package the vehicle information according to the transport invitation and send the transport documents to the dispatch agent. It includes the following steps:

[0088] S31: After receiving the transport application from the order agent, the dispatch agent will process the transport application in a first-come-first-served order, that is, package the information as a transport invitation and send it to all vehicle agents.

[0089] Specifically, in this step, the transport request is placed in a queue, indicating that several transport requests can be managed at the same time. The purpose of this step is to sort the received requests in a first-come-first-served order, and then send the order information in the transport request as the transport condition to the vehicle agent, inviting each vehicle agent to send the transport documents.

[0090] S32: The vehicle agent who receives the transport invitation will first check whether there is a task in his work order task list that matches the order number of this transport application. If it matches, the task will be eliminated to free up working time. Then the vehicle agent will determine whether the required vehicle type is met. The vehicle agent that meets the conditions responds to the transport invitation, that is, the vehicle information is packaged as a transport document (one transport document corresponds to one vehicle) and sent to the dispatch agent.

[0091] Specifically, in this step, the transport invitations sent to the vehicle agent are placed in a queue, which means that several transport invitations can be managed at the same time. The agent will sort the received invitations in the order of first come first served. Next, it will check the order information in the transport invitation sent, clear the work order tasks with the same order number i in its own work order task list, and restore itself to idle time. Then it will integrate the vehicle information, and put the vehicle number Q, vehicle capacity m Q 、Vehicle expenses C Q , Estimated time T i , the starting time t of two adjacent orders corresponding to the target order i O(i-1) , end time t S(i+1)Pack as shipping documents and send shipping documents to dispatch representative.

[0092] S4: The dispatch agent screens the transport documents according to the preset screening criteria, forms a feasible transport vehicle set, and sends it to the decision agent. It includes the following steps:

[0093] S41: Based on the transportation document containing the parameters of the corresponding vehicle: vehicle number Q, estimated time T i , the corresponding starting time t of the two adjacent orders of target order i O(i-1) , end time t S(i+1) , conduct preliminary screening of transport documents. The screening criteria are as follows:

[0094] (1) Time when the vehicle can depart:

[0095] t ei <t O(i-1) <t li

[0096] (2) Estimated time of arrival of the vehicle:

[0097] t Ei <t O(i-1) +T i <t Li

[0098] (3)Latest end time:

[0099] t O(i-1) +T i <t S(i+1)

[0100] Only when the above three conditions are met at the same time, this transport vehicle can pass the preliminary screening and form the candidate transport vehicle set X.

[0101] S42: In the candidate transport vehicle set X, the transport documents are combined, and the obtained transport solutions are preliminarily screened based on the transport costs to obtain L transport solutions.

[0102] Specifically, the transportation documents can be directly combined to obtain a variety of transportation plans that meet the transportation needs, and then L transportation plans with lower transportation costs can be screened out.

[0103] Or a preliminary screening is performed based on the DP matrix, which is a dynamic programming algorithm obtained by adjusting and modifying the traditional 01 knapsack problem in the present invention. The core of the algorithm is the DP matrix obtained after calculation (i.e., Dynamic programming). The calculation idea of ​​the algorithm is: facing a fixed amount of cargo and transport vehicles with different transport capacities, different vehicle combinations are regarded as "states", represented by the row l and column j of the matrix, l represents the first l vehicles in the set X in the order of arrangement included in the combination range, and j represents the situation where the vehicle combination is loaded with j tons of cargo (j can be greater than the cargo volume M of the order number i). i ), and the corresponding matrix element value is the loss in this state, which represents the loss caused by using this vehicle combination. From the "state" of only the first vehicle, the algorithm can calculate several "states" to be selected, and the algorithm will select the one with the lowest loss as the next "state". Through continuous "state" recursion, the optimal vehicle combination loss under different transport capacities and situations with the same cargo volume can be obtained, thereby screening out the top L transportation plans with the lowest loss values.

[0104] Specifically, suppose that the vehicle set X has a total of n vehicles, Q is the vehicle number, C Q is the vehicle cost (i.e., transportation cost), m Q is the vehicle capacity (in tons and an integer), that is:

[0105] X={(C 1 ,m 1 ),(C 2 ,m 2 ),…,(C i ,m i ),…,(C n ,m n )}

[0106] The target transport volume is M i Tons, by combining the vehicles in set X, we can get the sum of the transport capacity corresponding to the combination ∑m. If M i >∑m, for every ton of cargo remaining, a penalty of P is imposed, where P is a value greater than the sum of all vehicle costs, which makes the loss much higher than the cost of all vehicles even if there is only 1 ton of cargo remaining. The loss is the sum of vehicle cost and cargo remaining penalty.

[0107] Initialize the DP matrix, and initialize the 0th row to the maximum cost. l means that there are l vehicles in the current state, and j refers to the amount of cargo carried by the current vehicle combination (j can be greater than M i ), DP l,j It represents the minimum loss incurred when the first l vehicle combinations and vehicle combinations transport j tons of goods.

[0108]

[0109] DP 0,j =P

[0110] For the lth vehicle, when l≥0:

[0111]

[0112] It can be understood that for the lth car, there are two decisions: use or not use; when using, the loss is The loss when not in use is DP l-1,j , take the minimum value of the two.

[0113] Iterative calculation can obtain the carrying capacity M i When , the loss values ​​of various transportation plans are obtained; the loss is arranged from small to large, and the top L transportation plans are obtained.

[0114] S43: Based on the earliest and latest transportation time of the transport vehicles in each of the L transportation plans, the transportation time T of each plan can be calculated. N , Estimated start time t SN , Estimated end time t ON Through permutation, the first K transportation plans with the least transportation time can be obtained, forming a set of feasible transportation vehicles.

[0115] After a feasible transport vehicle set is formed, it will be sent to the decision agent. If no transport plan can be formed, the "unable to dispatch" opinion will be fed back to the decision agent.

[0116] S5: The decision agent makes value judgments on each transportation plan, forms a decision and sends it to the dispatch agent and order agent. The order agent completes its own information update, and the dispatch agent sends the decision to the vehicle agent for execution. It includes the following steps:

[0117] S51: The decision agent makes the final decision by calling the LLM through prompts. During the calling process, the text instructions are used to determine the LLM's solution selection, and the instruction constraint text is used in real time during the decision-making process to influence the selection of LLMs from multiple perspectives, including the estimated time spent, loss, estimated start time, and estimated end time of the transportation vehicle solution, such as Figure 4 As shown. Use the prompt text to make a value judgment on the transportation plan and select the best transportation plan as the decision.

[0118] Specifically, in order to make full use of the functions of LLM, the Markdown format is used to describe the behavior of LLM from multiple perspectives, including the agent's target priority, output information, and instruction constraints. The agent's target priority is the direction in which the agent chooses the transportation plan, which is explained in natural language to facilitate LLM's understanding. The purpose of the output information is to define the information related to the agent's response. For example, in order to reduce costs, the decision-making agent can only output the estimated time and loss of the road section of the plan. In order to better control the time nodes, the estimated start time and the estimated end time can be output; however, in order to make full use of the agent's cognitive ability, the present invention adopts the widely used "thinking chain", that is, guiding the agent to gradually consider the decision-making process, using multiple instruction constraints to gradually narrow the range of options, and select high-value options that meet management requirements.

[0119] S52: After the decision is made, the decision will be sent to the order agent. The order agent that receives the decision will update its own order information table in preparation for the information call of the next event.

[0120] S53: After the decision is made, the decision will be sent to the dispatch agent, and the dispatch agent will send the decision to the vehicle agent with the corresponding vehicle number. The vehicle agent that receives the decision will add the transportation task to its own work order task list, thus completing the decision execution.

[0121] The physical implementation framework of the above scheduling method and system is as follows: Figure 2 As shown, it includes the physical layer, data layer, proxy negotiation layer and engine layer, where:

[0122] 1. Physical layer: The physical layer includes all the transport vehicles dispatched by the carrier for this project. It mainly relies on intelligent networking to detect the status of the transport vehicles in real time, and uses its data information to determine when the transport vehicles will deliver the goods, that is, to calculate the time it takes for the vehicle to complete the current transport work order. In the intelligent networking environment, the vehicle's transport work order completion time is measured by the following methods:

[0123] GPS positioning: The GPS positioning system records the vehicle's location information and driving speed in real time, and can accurately calculate the vehicle's delay based on the vehicle's actual driving conditions on the road.

[0124] Vehicle-mounted sensors: Vehicle-mounted sensors can be used to monitor the vehicle’s acceleration, speed, steering angle and other data in real time to determine whether the vehicle is in a congested state and further calculate the vehicle’s delay.

[0125] Internet of Vehicles data: Internet of Vehicles technology can be used to exchange information between vehicles, including road conditions, vehicle driving conditions, etc., so as to calculate vehicle delays in real time.

[0126] 2. Data layer: The data layer divides the data obtained by the physical layer into vehicle physical information and vehicle semantic information. With this information, the "Urban Traffic Simulation" (SUMO) is used to simulate the transportation situation. SUMO is an open source, highly portable, microscopic and continuous traffic simulation software package designed to handle large networks. The above method can complete the real-time calculation of the current work order transportation situation of the vehicle. Finally, the information of the vehicle agent representing each vehicle is sorted out to circulate in the entire multi-agent system as vehicle transportation information.

[0127] 3. Agent negotiation layer: The agent negotiation layer is mainly composed of order agents and dispatch agents. It is the core component of this system. Its function is to organize the information of the data layer to form an information set that the decision agent can identify and make value judgments. In this layer, after the order agent obtains the changed component order information, it drives the dispatch agent to collect vehicle transportation information from each vehicle agent. With the changed order information and vehicle transportation information, the dispatch agent forms a feasible transportation vehicle set composed of several transportation plans for the decision agent to weigh and select.

[0128] 4. Engine layer: The engine layer provides the LLM engine. The reasoning and training of LLM require large graphics processing units (GPUs). Through the application programming interface (API), the system establishes communication with the LLM deployed in a computer center enhanced by multiple GPUs, so that the agents in the system can run more efficiently. However, this will definitely affect the performance of the agent. Alibaba's Tongyi Qianwen Dayan model can be used for fine-tuning here. The first step is to select a dataset related to the task and pre-process the data, including cleaning, word segmentation, encoding, etc. The second step is to set the fine-tuning parameters, set hyperparameters such as learning rate, training rounds (epochs), batch size (batch size), and set other hyperparameters as needed, such as weight decay and gradient clipping. The third step is to train the reward model, which is used to evaluate the quality of text sequences. The training data usually consists of text sequences generated by multiple language models, which are manually evaluated or scored using other large models. The fourth step is to fine-tune the large model using supervised data. The process is to load the pre-trained model and weights; make necessary modifications to the model according to task requirements, such as changing the output layer; select a suitable loss function and optimizer; and use the selected dataset for fine-tuning training, including forward propagation, loss calculation, back propagation, and weight update.

[0129] It will be easily understood by those skilled in the art that the above description is only 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 protection scope of the present invention.

Claims

1. A transportation vehicle scheduling method based on a large language model, characterized in that: The steps include: Based on the adjusted order text, the order information is obtained through a large language model; Obtain required vehicle information based on order information, and package the order information and required vehicle information into a transportation application; Sort the current several transport applications and generate transport documents according to the vehicle conditions. One transport document corresponds to one vehicle. Screen the transportation documents to obtain all transportation documents that meet the transportation conditions; combine the screened transportation documents to obtain multiple transportation plans, and screen the transportation plans according to the preset goals to obtain K transportation plans; According to the textual indication of real-time demand, the final transportation plan is determined from K transportation plans through a large language model.

2. The method for dispatching transport vehicles based on a large language model as claimed in claim 1, characterized in that: Before using the large language model, a target instruction set including the task description text and the corresponding vehicle transportation plan is pre-built; then the large language model is fine-tuned using the target instruction set, that is, the large language model is trained with the task description text as input and the vehicle transportation plan as output.

3. The method for dispatching transport vehicles based on a large language model as claimed in claim 2, characterized in that: The target instruction set is constructed as follows: Collect the task description text and the corresponding vehicle transportation plan, and use the natural language task description to format the collected data. Then clean the formatted data to obtain the target instruction set.

4. The method for dispatching transport vehicles based on a large language model as claimed in claim 3, characterized in that: Clean the formatted data, including deduplication, correction and complication, specifically: The cosine similarity algorithm is used to calculate the semantic similarity of the task description text in the target instruction set, and some instances with high repetition are deleted or semantically similar instances are merged into a new instance. Then, errors in the data are detected and corrected through syntax analysis models, manual spot checks and script verification. Finally, some simple tasks are rewritten into complex tasks.

5. The method for dispatching transport vehicles based on a large language model as claimed in claim 1, characterized in that: The transportation document includes the vehicle number, vehicle capacity, vehicle transportation cost, estimated time, target order time, and the start and end times of two adjacent orders.

6. The method for dispatching transport vehicles based on a large language model as claimed in claim 5, characterized in that: The transportation conditions include vehicle departure time constraints, vehicle estimated arrival time constraints, and latest end time constraints, where: The vehicle departure time constraint is: t ei <t O(i-1) <t li The vehicle estimated arrival time constraint is: t Ei <t O(i-1) +T i <t Li The latest end time constraint is: t O(i-1) +T i <t S(i+1) Among them, t ei ,t li are the earliest delivery time and the latest delivery time of the i-th order respectively; t Ei ,t Li are the earliest entry time and the latest entry time of the i-th order respectively; t O(i-1) is the estimated end time of the i-1th order, T i is the estimated time it takes for the vehicle to complete the i-th order, t S(i+1) is the start time of the i+1th order.

7. The transportation vehicle scheduling method based on a large language model according to any one of claims 1 to 6, characterized in that: Combine the selected transportation documents to obtain multiple transportation plans, and filter the transportation plans according to the preset goals to obtain K transportation plans, including: Based on the transportation cost, a plurality of transportation plans obtained by combining the transportation documents are preliminarily screened to obtain L transportation plans with lower transportation costs; Then the transportation time of the L transportation plans is calculated respectively, and K transportation plans with less transportation time are further screened.

8. The method for dispatching transport vehicles based on a large language model as claimed in claim 7, characterized in that: Based on the transportation cost, a preliminary screening is conducted on the various transportation solutions obtained by combining the transportation documents, specifically: According to the cargo volume, the selected transport documents are combined through the DP matrix, and the loss of each transport plan is calculated. The top N transport plans with the smallest loss value are initially screened; The values ​​in the DP matrix represent the loss, and its 0th row is initialized to the maximum cost, that is, DP 0,j =P, the maximum cost P is defined as the value greater than the sum of the transportation costs of all vehicles; when l≥1, the value DP in the lth row and jth column l,j It represents the minimum loss when the first l vehicles are included in the combination and the vehicle combination transports j tons of goods. The calculation formula is: DP l,j =min(DP l-1,j-pl +C l -P×min(p l ,j),DP l-1,j ) Among them, p l represents the capacity of the lth vehicle, C l Represents the transportation cost of the lth vehicle.

9. A system for implementing the transportation vehicle scheduling method based on a large language model as described in any one of claims 1 to 8, characterized in that: It includes decision-making agent, order agent, dispatch agent and vehicle agent, among which: The decision agent is implemented based on a large language model, and is used to form order information based on the adjusted order text and send it to the order agent; The order agent is used to obtain the required vehicle information based on the order information, and package the order information and the required vehicle information into a transportation application and send it to the dispatch agent; The dispatch agent is used to sort the transport applications from the order agents and form transport invitations to be sent to all vehicle agents; The vehicle agent is used to generate a transportation document based on the vehicle situation after receiving the transportation invitation and send it to the dispatch agent, where one transportation document corresponds to one vehicle; The dispatch agent is also used to screen the transport documents to obtain all transport documents that meet the transport conditions; combine the screened transport documents to obtain multiple transport plans, and screen the transport plans according to the preset goals to obtain K transport plans, which are sent to the decision agent; The decision agent is also used to determine the final transportation plan from K transportation plans according to the text indication of real-time demand, so as to form a decision and send it to the scheduling agent and the order agent. The order agent completes its own information update, and the scheduling agent sends the decision to the vehicle agent for execution.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the transportation vehicle scheduling method based on a large language model as described in any one of claims 1 to 8 is implemented.

Images