An artificial intelligence-based logistics transportation management method and system

Abstract

The application provides a logistics transportation management method and system based on artificial intelligence, relates to the field of logistics transportation, and solves the technical problems that the prior art is difficult to balance real-time performance, distribution feasibility and resource utilization efficiency in the sudden interruption scene of a logistics vehicle, resulting in slow scheduling response, frequent secondary conflicts or failure of high-value order fulfillment. The method comprises the following steps: obtaining a set of orders to be delivered by a target vehicle at present and real-time position information; based on a preset hard constraint condition, screening candidate vehicles and constructing a task migration feasibility graph; inputting the task migration feasibility graph into a pre-trained lightweight graph neural network model to output the compatibility scores of each candidate vehicle for each order to be delivered in the order set; based on the compatibility scores, evaluating the urgency of the order to be delivered; and based on the urgency of the order to be delivered, updating the delivery task scheme of each candidate vehicle. The application is used in the process of logistics transportation.

Description

Technical Field

[0001] This application relates to the field of logistics and transportation, and in particular to a logistics and transportation management method and system based on artificial intelligence. Background Technology

[0002] In modern urban logistics, on-demand delivery, and same-city freight multi-vehicle collaborative delivery systems, when a target vehicle is unable to continue its original delivery task due to unforeseen circumstances (such as mechanical failure, traffic accident, temporary driver absence, or traffic control), the system must quickly reassign some or all of the orders to be delivered carried by that vehicle to other available vehicles to ensure overall service continuity and customer experience.

[0003] Currently, mainstream scheduling systems primarily employ two strategies to address such interruption events: one is the global rescheduling method, which triggers network-wide vehicle route re-optimization, resolving all incomplete orders (including tasks for non-interrupted vehicles); the other is the local rule-based allocation method, such as simply assigning interrupted orders to geographically nearest or lowest-loaded neighboring vehicles. However, while the former can theoretically achieve a globally optimal solution, its computational complexity increases exponentially with the number of vehicles and orders, making it difficult to meet second-level response requirements in real-world high-concurrency scenarios and prone to causing drastic adjustments to network-wide tasks, resulting in scheduling oscillations. The latter, while computationally efficient, lacks comprehensive consideration of time window constraints, customer priorities, vehicle compatibility, and path coupling effects, often leading to new allocation schemes violating hard constraints (such as overload and timeout), thereby triggering secondary conflicts or defaults on high-value orders.

[0004] Especially in large-scale real-time scheduling scenarios, existing technologies lack an intelligent mechanism that can quickly generate highly feasible redistribution schemes within a limited neighborhood, avoiding the high overhead of global replanning and overcoming the blindness of simple rules. Therefore, this application provides a logistics transportation management method and system based on artificial intelligence. Summary of the Invention

[0005] This application provides an artificial intelligence-based logistics transportation management method and system, which solves the technical problem that existing technologies struggle to balance real-time performance, allocation feasibility, and resource utilization efficiency in the event of sudden disruptions in logistics vehicles, leading to slow dispatch response, frequent secondary conflicts, or failure to fulfill high-value orders.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, an artificial intelligence-based logistics transportation management method is provided, including: Obtain the current set of orders awaiting delivery and their real-time location information for the target vehicle; wherein, the target vehicle is the vehicle whose transportation task has been interrupted; Based on preset hard constraints, candidate vehicles that are geographically close to the target vehicle and meet the carrying capacity are selected, and a task migration feasibility map is constructed. The nodes of the task migration feasibility map represent the candidate vehicles and the orders to be delivered. The edge weights are determined based on the carrying cost of the orders carried by the candidate vehicles. The carrying cost is obtained by weighted summation of detour cost, timeliness breach cost, and service priority compensation. The task migration feasibility map is input into a pre-trained lightweight graph neural network model, which outputs the compatibility score of each candidate vehicle for each order to be delivered in the order set. Based on the compatibility score, the urgency of the orders to be delivered is assessed; based on the urgency of the orders to be delivered, the delivery task plan for each candidate vehicle is updated.

[0007] Based on the above technical solutions, in the AI-based logistics transportation management method provided in this application, the key challenge in ensuring service continuity is how to quickly and reliably reassign the original orders to other vehicles when a vehicle's task is interrupted due to sudden situations such as malfunctions or accidents. Traditional methods either rely on global rescheduling or use simple rule allocation, making it difficult to generate high-quality reassignment plans within seconds. To address this, this solution proposes an intelligent reassignment mechanism that integrates artificial intelligence and operations research optimization: it filters neighboring candidate vehicles based on hard constraints, constructs a task migration feasibility graph, and explicitly filters infeasible matches; it designs a multi-dimensional acceptance cost model that includes detour costs, timeliness default costs, and service priority compensation items, assigning weights to edges; it then learns high-order compatibility between orders and vehicles through a lightweight graph neural network, outputting interpretable matching scores; and it feeds compatibility back to urgency assessment to achieve dynamic priority ranking based on "environmental awareness" and update task plans. This solution, without triggering a full-network replanning, balances real-time performance, feasibility, and commercial value, significantly improving the fulfillment rate of high-priority orders and system robustness, providing an efficient and intelligent solution for large-scale dynamic logistics scheduling.

[0008] In conjunction with the first aspect above, in one possible implementation, the method for constructing the task migration feasibility map includes: Each order in the set of orders to be delivered to the target vehicle is taken as an order node, and each candidate vehicle in the set of candidate vehicles is taken as a vehicle node, together forming a node set of the graph; For each order node and each vehicle node, determine whether the preset soft constraint compatibility conditions are met; if yes, then establish a directed edge between the order node and the vehicle node; otherwise, do not establish a directed edge between them; the soft constraint compatibility conditions include: there is at least one path insertion position such that the time window of the order is not violated and the vehicle is not blacklisted by the customer corresponding to the order; if satisfied, then establish a directed edge between them. For each established directed edge, calculate its edge weight; Configure a node feature vector for each order node, which includes order weight, volume, start and end time of time window, and customer geographical location; configure a node feature vector for each vehicle node, which includes vehicle current location, remaining load capacity, remaining cargo hold volume, current cumulative driving time, and historical order fulfillment rate; Based on the node set, directed edges, edge weights, and node feature vectors, a feasibility map for task migration is generated.

[0009] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the edge weights includes: Calculate the detour distance increment caused by inserting the order into the current delivery route of the candidate vehicle, and use the ratio of the detour distance increment to a preset baseline distance as the detour cost; The estimated delivery time of the order is calculated based on the inserted path. If the estimated delivery time exceeds the deadline of the time window specified by the customer, the ratio of the timeout amount to the width of the time window of the order is used as the time-out penalty; otherwise, the time-out penalty is set to zero. Obtain the normalized priority coefficient of the customer corresponding to the order. and will As a service priority compensation item; The detour cost, time-sensitive breach of contract cost, and service priority compensation item are weighted and summed to obtain the acceptance cost, and the acceptance cost is assigned as the weight of the edge.

[0010] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the compatibility score includes: The graph neural network model is used to perform multi-round message passing on the order nodes and vehicle nodes in the task migration feasibility map to generate the embedding vectors of each order node and each vehicle node. Based on the embedding vector of any vehicle node, the embedding vector of the corresponding order node, and the features of the edge between them, the compatibility score of the vehicle for the order is calculated through a preset compatibility prediction module. The compatibility prediction module is a multilayer perceptron containing at least one fully connected layer, and its output, after being processed by an activation function, represents the probability of matching feasibility.

[0011] In conjunction with the first aspect above, in one possible implementation, updating the delivery task plan for each candidate vehicle includes: After sorting the orders to be assigned in descending order of urgency, they are assigned sequentially to the candidate vehicles with the highest current load capacity score that still meet all soft and hard constraints after the order is assigned; wherein, the load capacity score is used to characterize the resource utilization efficiency of the candidate vehicle after accepting the order; After the initial allocation is completed, it is checked whether any candidate vehicle has a time window violation or load overload conflict. If so, then in the local subgraph formed by the conflicting candidate vehicle and other candidate vehicles within its one-hop or two-hop adjacency range, at most two rounds of order exchange are performed within the local subgraph to eliminate the local conflict and generate a feasible allocation scheme. Here, the one-hop adjacency candidate vehicle is the candidate vehicle that is directly connected to the conflicting vehicle in the task migration feasibility graph; the two-hop adjacency candidate vehicle is the candidate vehicle that is not directly connected to the conflicting vehicle but has at least one common adjacency vehicle.

[0012] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the load factor score includes: Based on the total weight and volume of the cargo currently loaded on the vehicle, divide them by the maximum allowable load capacity and the maximum cargo hold volume respectively to obtain the weight utilization rate and volume utilization rate. Then, sum them according to the preset weights to obtain the full load rating of the candidate vehicle.

[0013] In conjunction with the first aspect above, in one possible implementation, the assessment of the urgency of orders to be delivered includes: Given the time urgency and customer importance, for any order to be reassigned Its urgency Defined as: ; in, Customer level weighting; This is the current timestamp of the scheduling system; For orders The deadline for the time window specified by the customer; The shortest possible travel time from the current location of the target vehicle or candidate vehicle to the order's destination; It is a very small positive number, used to prevent the denominator from being zero and causing numerical anomalies.

[0014] In conjunction with the first aspect above, in one possible implementation, the method for determining the customer level weight includes: Get an order The historical performance value score corresponds to the customer. The historical performance value score is calculated based on the customer's order gross profit within a preset number of days in the past and is weighted by time decay. Calculate the current scheduling sensitivity score for the customer, where the current scheduling sensitivity score is the ratio of the order price to the minimum cost required to accept the order among all candidate vehicles; Obtain the customer's service stability coefficient, which is generated based on the normalization of its historical cancellation rate, complaint rate, and on-time delivery rate. After normalizing the historical performance value score, current scheduling sensitivity score, and service stability coefficient, the scores are weighted and summed according to preset weights, and then mapped to the (0,1) interval using the Sigmoid function to obtain the customer level weight. .

[0015] In conjunction with the first aspect above, in one possible implementation, the method for screening candidate vehicles includes: Obtain the real-time location of all other vehicles in the dispatch system that are not in a faulty state; Calculate the road network distance from each vehicle to the target vehicle, and retain vehicles whose road network distance does not exceed a preset neighborhood radius threshold as the initial screening vehicle set; For each vehicle in the initial screening vehicle set, the following hard constraints are checked sequentially: The remaining load capacity of the vehicles initially screened is not less than the total weight of the set of orders to be delivered; The remaining cargo hold volume of the initially screened vehicles shall not be less than the total volume of the set of orders to be delivered; After inserting any order from the set of orders to be delivered into the current delivery route of the initial screening vehicle, there is at least one insertion position such that the estimated delivery time of the order falls within the time window specified by its customer. The cumulative driving time of the vehicles screened in the initial screening after accepting new orders shall not exceed the legally mandated continuous driving time limit; The initial screening process ensures that the vehicle model and onboard equipment configuration match the transportation requirements of all orders in the set of orders to be delivered. When all hard constraints are met, the vehicles initially screened are considered candidate vehicles.

[0016] Secondly, this application provides an artificial intelligence-based logistics transportation management system, comprising: an acquisition module, a construction module, and an allocation module; wherein, the acquisition module is used to acquire the current set of orders to be delivered by the target vehicle and its real-time location information; the construction module is used to filter out candidate vehicles that are geographically close to the target vehicle and meet the carrying capacity requirements based on preset hard constraints, and construct a task migration feasibility map; the allocation module is used to input the task migration feasibility map into a pre-trained lightweight graph neural network model, and output the compatibility score of each candidate vehicle to each order to be delivered in the order set; based on the compatibility score, the urgency of the orders to be delivered is evaluated; and based on the urgency of the orders to be delivered, the delivery task plan of each candidate vehicle is updated.

[0017] This application provides an AI-based logistics transportation management method and system that effectively addresses the challenges of dynamic order reallocation in scenarios of sudden vehicle disruptions, significantly improving the real-time performance, feasibility, and service quality of the scheduling system. The method constructs a task migration feasibility graph, modeling orders to be delivered and neighboring candidate vehicles that meet hard constraints (such as load capacity, volume, time window, and vehicle type matching) as a heterogeneous graph structure. It introduces a multi-dimensional acceptance cost that integrates detour costs, timeliness default risks, and customer priority as edge weights, providing structured input for intelligent decision-making. Based on this, a pre-trained lightweight graph neural network is used to learn high-order interaction relationships between nodes, outputting interpretable compatibility scores to avoid the problem of traditional black-box models ignoring business constraints. Furthermore, the system feeds back compatibility information to an urgency assessment mechanism, achieving dynamic priority ranking based on "environmental awareness," and efficiently updating the delivery task plans for each vehicle accordingly. The entire process does not require full network replanning; it only completes millisecond-level reallocation within a locally feasible range, significantly reducing computational overhead while ensuring that the plan meets all soft and hard constraints. In practical applications, the system can significantly improve the on-time delivery rate of high-value orders, reduce secondary conflicts and resource waste, enhance the robustness and self-healing ability of the logistics network under emergencies, and provide technical support for large-scale, high-concurrency intelligent logistics scheduling that combines efficiency, fairness and commercial value.

[0018] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0019] Figure 1 A system architecture diagram of an artificial intelligence-based logistics transportation management system provided for embodiments of this application; Figure 2 A flowchart illustrating an artificial intelligence-based logistics transportation management method provided in this application embodiment; Figure 3This is a flowchart illustrating the method for constructing a task migration feasibility map provided in an embodiment of this application. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The AI-based logistics transportation management method provided in this application embodiment can be applied to an AI-based logistics transportation management system, such as... Figure 1 As shown, the system includes: an acquisition module, a construction module, and an allocation module; The acquisition module is used to acquire the current set of orders to be delivered and the real-time location information of the target vehicle; The construction module is used to filter out candidate vehicles that are geographically close to the target vehicle and meet the carrying capacity based on preset hard constraints, and to construct a task migration feasibility map. The allocation module is used to input the task migration feasibility map into a pre-trained lightweight graph neural network model and output the compatibility score of each candidate vehicle with each order to be delivered in the order set. Based on compatibility scores, the urgency of orders to be delivered is assessed; based on the urgency of orders to be delivered, the delivery task plan for each candidate vehicle is updated.

[0022] To address the technical challenges of existing technologies in balancing real-time performance, allocation feasibility, and resource utilization efficiency during sudden disruptions in logistics vehicle operations, leading to slow dispatch response, frequent secondary conflicts, or failure to fulfill high-value orders, this application provides an AI-based logistics transportation management method. This method includes: acquiring the current set of orders awaiting delivery for the target vehicle and its real-time location information; based on preset hard constraints, selecting candidate vehicles geographically close to the target vehicle and meeting its carrying capacity requirements, and constructing a task migration feasibility map; inputting the task migration feasibility map into a pre-trained lightweight graph neural network model, and outputting the order set for each candidate vehicle. This application proposes an intelligent reallocation mechanism that integrates graph neural networks and business logic. This mechanism involves: obtaining the orders and locations of interrupted vehicles; filtering nearby feasible candidate vehicles based on hard constraints; constructing a task migration feasibility graph to filter invalid matches at the source; semantically modeling the graph structure using a pre-trained lightweight GNN model to output order-vehicle compatibility scores and capture high-order interaction patterns; and then dynamically evaluating order urgency based on these scores and updating the task plans for each vehicle accordingly. This technical approach achieves efficient, feasible, and value-oriented local reallocation without triggering a full-network replanning, ensuring the fulfillment quality of high-priority orders while significantly improving system robustness and resource utilization efficiency, providing a feasible AI-enabled paradigm for intelligent logistics.

[0023] like Figure 2 As shown in the embodiment of this application, a logistics transportation management method based on artificial intelligence is provided, including: S201. Obtain the current set of orders awaiting delivery and the real-time location information of the target vehicle.

[0024] The target vehicle is the vehicle whose transportation task has been interrupted; the set of orders to be delivered includes all orders that the target vehicle has accepted but not yet delivered. Each order includes destination coordinates, customer time window, cargo volume and weight, service priority and order generation time; real-time location information is reported to the dispatch server at a preset sampling frequency (e.g., every 5 seconds) through the vehicle-mounted GPS terminal or mobile terminal positioning module, and is mapped to the precise road segment node in the road network topology after being corrected by the map matching algorithm.

[0025] In some implementations, if the target vehicle loses its location information due to communication interruption, its current location is predicted using Kalman filtering or particle filtering algorithms based on its most recent valid location, historical average driving speed, and current traffic flow status. At the same time, the set of orders to be delivered can be read synchronously from the task queue cache of the scheduling system to ensure that a complete set of unfulfilled tasks can still be obtained even when the vehicle is offline.

[0026] It should be noted that obtaining the order set and real-time location information is a prerequisite for triggering the feasibility analysis of subsequent task migration. The two must be synchronized in terms of timestamps (with an error not exceeding a preset threshold, such as 30 seconds) to avoid distortion of the undertaking cost calculation or failure of path planning due to inconsistent status. In addition, to ensure data security, all location and order information is encrypted during transmission and storage, and the access scope is restricted according to the principle of least privilege.

[0027] S202. Based on the preset hard constraints, select candidate vehicles that are geographically close to the target vehicle and meet the carrying capacity, and construct a task migration feasibility map.

[0028] In the task migration feasibility map, the nodes represent candidate vehicles and orders to be delivered; the edge weights are determined based on the acceptance cost of the candidate vehicle to accept the order, and the acceptance cost is obtained by weighted summation of detour cost, timeliness breach cost and service priority compensation.

[0029] S203. Input the task migration feasibility map into the pre-trained lightweight graph neural network model and output the compatibility score of each candidate vehicle with each order to be delivered in the order set.

[0030] The compatibility scores are assigned to the orders to be delivered, and a graph neural network model is obtained. The lightweight graph neural network model includes 2 to 3 graph convolutional layers (such as GraphSAGE or GAT) to aggregate neighborhood information to generate node embeddings, and is connected to an edge-level compatibility prediction head. This prediction head is fed into a multilayer perceptron (MLP) after splicing vehicle node embeddings, order node embeddings and original edge features, and finally outputs a compatibility score normalized to the [0,1] interval. The higher the score, the stronger the matching feasibility.

[0031] In some implementations, the graph neural network model is trained end-to-end based on historical scheduling logs during the offline phase, with labels as binary variables: if a vehicle successfully fulfills an order in the historical record, the corresponding edge label is 1, otherwise it is 0; during training, a binary cross-entropy loss function is used, and a negative sampling strategy is introduced to alleviate the problem of imbalance between positive and negative samples; during online inference, the model is deployed on edge computing nodes or scheduling servers, and the time for a single inference is controlled within 50 milliseconds to meet the requirements of real-time scheduling response.

[0032] It should be noted that the compatibility score is not directly used for the final allocation decision, but rather serves as the ranking basis or constraint filtering threshold in the greedy-backtracking hybrid strategy. In addition, to prevent the model from overfitting to unseen order-vehicle combinations or having inflated confidence, the system sets a minimum score threshold (e.g., 0.3). Edges below this threshold will be removed from the feasible allocation candidate set, thereby ensuring the reliability and executability of the allocation results.

[0033] S204. Based on the compatibility score, assess the urgency of the orders to be delivered; based on the urgency of the orders to be delivered, update the delivery task plan for each candidate vehicle.

[0034] Based on the above technical solutions, this application provides an AI-based logistics transportation management method. In the logistics transportation process, when a vehicle is unable to continue its task due to sudden interruptions such as malfunctions or accidents, a technical means is urgently needed to generate a high-quality redistribution solution in a very short time. Traditional methods either rely on global re-optimization, leading to high computational latency and large system fluctuations; or they use simple proximity or static rule allocation, ignoring multi-dimensional constraints such as time windows, load capacity, and customer priority, easily causing secondary conflicts or timeouts for high-value orders. To address this problem, this solution innovatively constructs an intelligent redistribution mechanism that integrates artificial intelligence and business rules: based on hard constraints, it filters geographically proximate and capability-matching candidate vehicles, forming a task migration feasibility map, effectively narrowing the search space and eliminating infeasible combinations; it designs a structured acceptance cost as edge weight by comprehensively considering detour costs, timeliness default risks, and customer priority; then, it uses a lightweight graph neural network to perform semantic perception on the graph structure, outputting a compatibility score with physical meaning; this score is fed back to the dynamic assessment of urgency, realizing "adaptive scheduling environment" order priority ranking, and updating the tasks of each vehicle accordingly. This method achieves a balance between millisecond-level response, highly feasible allocation, and efficient resource utilization while avoiding network-wide replanning, significantly improving the system's stability, service quality, and operational efficiency under emergencies.

[0035] In one possible implementation of the embodiments of this application, such as Figure 3 As shown, the above S202 can be specifically implemented through the following S301-S306, which are explained in detail below: S301. Screening candidate vehicles. The specific screening process is as follows: S3011. Obtain the real-time location of all other vehicles in the dispatch system that are not in a fault state; calculate the road network distance from each vehicle to the target vehicle, and retain vehicles whose road network distance does not exceed the preset neighborhood radius threshold as the initial screening vehicle set.

[0036] "Non-faulty state" refers to a state where the vehicle is not experiencing mechanical failure, communication interruption, driver absence, or is marked as unschedulable by the system. Real-time location information is reported to the dispatch server via the onboard terminal at a fixed frequency (e.g., every 5 seconds) and mapped to specific road segment nodes in the road network topology using a high-precision map matching algorithm. The road network distance is not a Euclidean straight-line distance, but rather the shortest drivable path length dynamically calculated based on current traffic conditions (e.g., congestion, traffic restrictions), determined by an integrated map path planning engine (such as Dijkstra or A). The algorithm is generated by combining real-time traffic conditions.

[0037] In some implementations, if some vehicles lose location information due to brief communication delays, their most recent valid location point is used in conjunction with their historical average speed for linear extrapolation to estimate their current location. Meanwhile, the preset neighborhood radius threshold can be dynamically adjusted according to the business scenario; for example, it can be set to 3 kilometers in densely populated urban areas and extended to 10 kilometers in suburban or highway scenarios. In addition, to improve computational efficiency, the system can pre-build a spatial index (such as an R-tree or GeoHash grid) to quickly filter out vehicles within the geographical coarse screening range, and then perform precise road network distance calculations on the candidate set.

[0038] It should be noted that the initial vehicle set is only used as the input basis for subsequent hard constraint verification and does not guarantee the ability to accept orders. The calculation of road network distance must be based on real-time or near real-time road network data. If a static road network model is used (such as ignoring current congestion), it may lead to distortion of proximity judgment and thus miss truly efficient candidate vehicles. Therefore, the dispatching system needs to keep data synchronized with the traffic information service platform to ensure that the road network distance reflects the true accessibility.

[0039] S3012. For each vehicle in the initial screening vehicle group, verify the following hard constraints in sequence: The remaining load capacity of the vehicles initially screened shall not be less than the total weight of the order collection to be delivered; The remaining cargo hold volume of the vehicles initially screened shall not be less than the total volume of the order set to be delivered; After inserting any order from the set of orders to be delivered into the current delivery route of the initial screening vehicle, there is at least one insertion position such that the estimated delivery time of the order falls within the time window specified by its customer. The cumulative driving time of the vehicles screened in the initial screening after accepting new orders shall not exceed the legally mandated continuous driving time limit; The initial screening process ensures that the vehicle model and onboard equipment configuration match the transportation requirements of all orders in the order set to be delivered; S3013. When all hard constraints are met, the vehicles initially screened are considered candidate vehicles.

[0040] Based on the above technical solutions, in a logistics scheduling system, when a target vehicle experiences a sudden interruption (such as a breakdown or accident), the original order needs to be quickly reassigned to other vehicles. Blindly trying all vehicles in the network results in high computational complexity and significant response latency, making it difficult to meet real-time requirements. On the other hand, simple nearest-neighbor allocation easily overlooks hard constraints such as load capacity, time window, and regulations, leading to secondary conflicts or service breaches. Therefore, this solution adopts a two-stage filtering mechanism: "geographical neighborhood initial screening + multi-dimensional hard constraint verification." The candidate range is limited based on road network distance, significantly narrowing the search space. Each initially screened vehicle is verified for load capacity, volume, time window feasibility, driving time compliance, and vehicle type matching, ensuring that only truly feasible candidate vehicles are retained. This method avoids the high overhead of global replanning and eliminates system oscillations caused by infeasible allocations, significantly improving redistribution efficiency and solution reliability, providing efficient and robust technical support for high-concurrency, strongly constrained real-time logistics scheduling.

[0041] S302. Treat each order in the target vehicle's set of orders to be delivered as an order node, and treat each candidate vehicle in the candidate vehicle set as a vehicle node, together forming a node set of the graph.

[0042] S303. For each order node and each vehicle node, determine whether the preset soft constraint compatibility condition is met; if yes, then establish a directed edge between the order node and the vehicle node; otherwise, do not establish a directed edge between them. The soft constraint compatibility conditions include: (1) there is at least one feasible path insertion position, such that after the order is inserted into the current delivery route of the candidate vehicle, its expected delivery time still falls within the time window specified by the customer; (2) the candidate vehicle is not included in the service blacklist by the customer corresponding to the order. The blacklist is generated by business rules such as customer history complaints, vehicle model incompatibility or special service restrictions, and stored in the customer profile database. Only when the above two conditions are met at the same time, a directed edge from the order to the vehicle is established between the order node and the vehicle node, indicating that the vehicle is a potential feasible recipient of the order.

[0043] In some implementations, the feasibility of a path insertion location is quickly verified using an incremental path evaluation algorithm: traversing the insertion gaps between all adjacent task pairs in the vehicle's current path (including before the start point and after the end point), calculating the driving time sequence after insertion, and using the triangle inequality pruning to terminate obviously infeasible branches in advance; if any insertion point satisfies the time window constraint and does not cause subsequent order chain timeouts, it is determined that there is a feasible insertion location; in addition, blacklist checks can be implemented in conjunction with real-time API calls or local caching mechanisms to balance response speed and data consistency.

[0044] It should be noted that the establishment of directed edges is a key preliminary step in constructing the feasibility graph for task migration. Its purpose is not to guarantee global optimality, but to efficiently filter out order-vehicle matching pairs with basic fulfillment probability. If this soft constraint verification is skipped and the subsequent graph neural network or backtracking search is used to handle infeasible combinations, the amount of invalid computation will be significantly increased and the system stability will be reduced. Therefore, the edge construction logic is essentially a lightweight feasibility pre-screening mechanism, providing a high-quality, high signal-to-noise ratio input graph structure for subsequent intelligent allocation.

[0045] S304. For each established directed edge, calculate its edge weight.

[0046] S3041. Calculate the detour distance increment caused by inserting the order into the current delivery route of the candidate vehicle, and use the ratio of the detour distance increment to the preset baseline distance as the detour cost.

[0047] The detour distance increment is the difference between the total distance the vehicle travels after inserting the order and the total distance of the original route. The preset baseline distance is the typical one-way delivery distance in the urban area (such as 5 kilometers), which is used to normalize the detour impact under different regions or scales.

[0048] In some implementations, to improve computational efficiency, the detour distance increment is calculated only at the optimal insertion position (which is the insertion point with the shortest total path); if there are multiple feasible insertion points, the one that minimizes the detour is selected as the evaluation benchmark.

[0049] S3042. Calculate the estimated delivery time of the order based on the inserted path. If the estimated delivery time exceeds the deadline of the time window specified by the customer, the ratio of the timeout amount to the width of the time window of the order shall be used as the time-out penalty. Otherwise, the time-out penalty shall be set to zero.

[0050] The time window width is defined as the difference between the earliest and latest delivery times allowed by the customer, and the timeout is the positive value of the estimated delivery time minus the time window deadline.

[0051] In some implementations, the cost of time-limited breach can be further introduced into a segmented penalty function; for example, an additional penalty factor is applied when the timeout exceeds 50% of the time window width.

[0052] S3043. Obtain the normalized priority coefficient of the customer corresponding to the order. and will As a service priority compensation item; Among them, the normalization priority coefficient Determined based on customer level, historical performance value, or contractual service level agreement (SLA), and mapped using Min-Max or Z-score methods. Interval.

[0053] S3044. The detour cost, time-related breach of contract cost, and service priority compensation are weighted and summed to obtain the acceptance cost, and the acceptance cost is assigned as the weight of the edge.

[0054] Among them, the three costs are weighted and integrated through non-negative weights, and the final acceptance cost is directly used as the weight of the corresponding edge in the task migration feasibility graph. The lower the value, the better the match.

[0055] Based on the aforementioned technical solutions, in dynamic order reallocation scenarios, allocating interrupted orders solely based on simple rules (such as proximity) can easily lead to resource waste, customer delays, or the neglect of high-value orders, making it difficult to balance operational efficiency and service quality. To address this, this solution proposes a multi-dimensional integrated acceptance cost model. This model uniformly weights detour costs (reflecting resource consumption), timeliness breach costs (quantifying service risk), and service priority compensation (reflecting customer value), generating comparable edge weights. Specifically, detour costs are normalized to eliminate regional differences; timeliness breach costs are measured by the proportion of delays within the time window to assess breach severity; and priority compensation ensures implicit preferential treatment for high-value customers. This acceptance cost model not only provides semantically rich edge features for graph neural networks but also provides a quantitative basis for subsequent greedy allocation and backtracking search. Compared to traditional single-objective optimization methods, this solution achieves synergistic optimization of cost, timeliness, and customer value without significantly increasing computational overhead, significantly improving the feasibility, fairness, and commercial benefits of scheduling schemes. It represents a key innovation for achieving refined decision-making in intelligent logistics systems.

[0056] S305. Configure a node feature vector for each order node.

[0057] The node feature vector includes order weight, volume, start and end time of time window, and customer geographical location; each vehicle node is configured with a node feature vector, which includes the vehicle's current location, remaining load capacity, remaining cargo hold volume, current cumulative driving time, and historical order fulfillment rate.

[0058] S306. Based on the node set, directed edges, edge weights, and node feature vectors, generate a task migration feasibility map.

[0059] Based on the above technical solutions, in scenarios of sudden vehicle outages, traditional scheduling methods struggle to efficiently and reliably reassign orders to other vehicles. This is because order-vehicle matching involves multi-dimensional heterogeneous constraints (such as time windows, blacklists, capacity, and location), and direct brute-force enumeration is computationally expensive and prone to generating infeasible solutions. To address this, this solution innovatively constructs a task migration feasibility graph: using orders and vehicles as heterogeneous nodes, directed edges are established only when soft constraints (feasible time window, non-blacklist) are met, and edge weights reflecting the acceptance cost are assigned, while rich structured features are embedded. This graph structure not only filters infeasible matches, significantly reducing the search space, but also provides semantically clear and constraint-aware input for subsequent graph neural networks. Compared to end-to-end black-box models or rule engines, this solution combines interpretability, scalability, and real-time performance, avoiding invalid allocations while supporting intelligent algorithms to accurately evaluate compatibility. Thus, it achieves high-efficiency, high-quality dynamic reallocation while ensuring hard constraints, serving as a crucial bridge connecting operations research optimization and deep learning.

[0060] In one possible implementation of this application embodiment, the above-mentioned S203 can be specifically implemented by the following S401-S402, which will be described in detail below: S401. Use a graph neural network model to perform multi-round message passing on the order nodes and vehicle nodes in the task migration feasibility map, and generate the embedding vectors of each order node and each vehicle node.

[0061] The graph neural network model is a lightweight architecture designed for heterogeneous graphs, containing 2 to 3 graph convolutional layers. In each round of message passing, each node aggregates the feature information of heterogeneous nodes in its first-order neighborhood. For example, an order node receives messages from connected vehicle nodes, and a vehicle node collects information from its associated order node. The aggregation function uses a weighted average or attention mechanism, with the edge weight (i.e., the aforementioned acceptance cost) serving as the normalization coefficient or attention score during aggregation to reflect the difference in credibility between different matching relationships. After multiple iterations, the hidden state of each node is updated to an embedding vector that integrates local structure and semantic features, denoted as the order embedding and vehicle embedding, respectively. The dimension is usually set to 64 or 128 to balance expressive power and computational cost.

[0062] In some implementations, to distinguish node types, the system introduces learnable type embeddings and concatenates them to the front end of the initial node features; during message passing, skip connections are used to concatenate or add the outputs of each layer to alleviate the gradient vanishing problem in deep networks; in addition, if the graph is large, sparse matrix operations and batch processing strategies can be enabled to perform inference only on active subgraphs, ensuring that the time taken to generate a single embedding is controlled within 50 milliseconds, meeting the real-time scheduling requirements.

[0063] It should be noted that the embedded vector is not the final decision result, but an intermediate representation used for subsequent compatibility prediction. Its core value lies in transforming the discrete order-vehicle matching relationship into a continuous, differentiable vector space expression, enabling the model to capture high-order interaction patterns (such as "a certain type of vehicle prefers orders from a certain region"). At the same time, since the message passing process explicitly utilizes edge weights (acceptance costs), the generated embedding naturally encodes business constraint information, thereby avoiding the "missing feasible solution" problem common in pure data-driven models and significantly improving the executability and robustness of downstream allocation schemes.

[0064] S402. Based on the embedding vector of any vehicle node, the embedding vector of the corresponding order node, and the features of the edge between them, calculate the compatibility score of the vehicle for the order through a preset compatibility prediction module. The compatibility prediction module is a multilayer perceptron (MLP) containing at least one fully connected layer, and its input is the concatenated joint feature vector. ;in, and These are the embedding vectors generated by the graph neural network for the vehicle node and the order node, respectively. The edge features include structured information such as the aforementioned acceptance cost, whether the time window is met, and blacklist status. The MLP sequentially fuses features through linear transformations and nonlinear activation functions (such as ReLU), and the final output layer uses the Sigmoid activation function to map the result to an interval. As a compatibility score for the vehicle-order pair, a higher score indicates a stronger matching feasibility.

[0065] In some implementations, MLP can be configured as a two-layer structure (e.g., input dimension → 64-dimensional hidden layer → 1-dimensional output). Dropout is introduced after the hidden layer to prevent overfitting. During the training phase, a binary cross-entropy loss function is used, and the labels are derived from historical scheduling logs. If a vehicle successfully fulfills an order in the historical record, the corresponding sample label is 1; otherwise, it is 0. During online inference, the compatibility score is not only used for ranking but also as a threshold filtering criterion (e.g., only retaining candidate matches with a score ≥ 0.3) to improve the solution efficiency and solution quality in the subsequent allocation phase.

[0066] It should be noted that the compatibility prediction module is not used in isolation, but rather forms an end-to-end trainable overall model with the graph neural network encoder. The key to its design lies in the explicit fusion of node semantic embeddings and original edge features, which not only preserves the high-order graph structure information captured by the GNN, but also does not lose the underlying business rules (such as whether hard constraints are satisfied). In addition, interpreting the output as "matching feasibility probability" helps to naturally connect with the confidence threshold mechanism in the downstream greedy-backtracking strategy, thereby maintaining the system's controllability and interpretability while ensuring intelligence.

[0067] Based on the above technical solutions, in dynamic order reallocation, the matching of orders and vehicles is not only constrained by explicit rules, but also implicitly contains complex high-order interaction patterns (such as certain types of vehicles being more suitable for specific regions or customers). Traditional rule engines struggle to capture such non-linear relationships, while pure data-driven models easily overlook business constraints, leading to infeasible outputs. Therefore, this solution adopts a collaborative architecture of Graph Neural Networks (GNNs) and a compatibility prediction module: multi-round message passing is performed on the task migration feasibility graph, embedding order and vehicle nodes with integrated local structure and neighbor information; combined with original edge features, a lightweight multilayer perceptron (MLP) is used to predict the matching feasibility probability. This design utilizes GNNs to capture the contextual semantics in the graph topology, while retaining underlying business rules (such as time windows and costs) through explicit edge features, achieving an organic integration of "data-driven" and "rule-constrained" approaches. Compared to end-to-end black-box models, the compatibility scores generated by this solution have strong interpretability and high feasibility, significantly improving allocation quality; compared to pure rule methods, it can discover potential high-quality matches, enhancing scheduling intelligence.

[0068] In one possible implementation of this application embodiment, the above-mentioned S204 can be specifically implemented by the following S501-S502, which will be described in detail below: S501. After sorting the orders to be assigned in descending order of urgency, assign them in turn to the candidate vehicles with the highest current load capacity score that still meet all soft and hard constraints after assigning the order.

[0069] The analysis process for the load factor score is as follows: Based on the total weight and volume of the cargo currently loaded on the vehicle, the weight utilization rate and volume utilization rate are obtained by dividing them by the maximum allowable load and the maximum cargo hold volume, respectively. These are then weighted and summed according to preset weights to obtain the full load score of the candidate vehicle, which represents the resource utilization efficiency of the candidate vehicle after accepting the order.

[0070] The process of analyzing order urgency is as follows: Given the time urgency and customer importance, for any order to be reassigned Its urgency Defined as: ; in, Customer level weighting; This is the current timestamp of the scheduling system; For orders The deadline for the time window specified by the customer; The shortest possible travel time from the current location of the target vehicle or candidate vehicle to the order's destination; It is a very small positive number, used to prevent the denominator from being zero and causing numerical anomalies.

[0071] It's important to note that the core purpose of this urgency formula is to scientifically and efficiently balance the two key business dimensions of "customer value" and "timeliness fulfillment risk" in dynamic order reallocation scenarios. Traditional scheduling methods often prioritize customers solely based on remaining time windows or static customer levels, making it difficult to handle complex trade-offs in the real world, such as "high-value customers about to expire" versus "low-value orders having ample availability but consuming vehicle resources." This formula addresses this by weighting customer levels... Using the numerator and the remaining buffer time between the expected delivery time and the time window deadline as the denominator, a non-linearly coupled urgency index was constructed: on the one hand, the more important the customer ( The larger the denominator (the larger the remaining delivery time), the higher the order priority. Conversely, the shorter the remaining delivery time (the smaller the denominator), the more exponentially the urgency is amplified. This is especially true when an order is nearing or has already exceeded its time limit, as the urgency level rises sharply. This ensures that the system can automatically focus on the high-risk, high-value tasks that require the most intervention. This design is not only mathematically simple and computationally inexpensive, suitable for millisecond-level real-time decision-making, but also achieves the intelligent scheduling goals of "prioritizing key tasks, controlling risks, and improving efficiency" in terms of business logic. This significantly improves the on-time delivery rate of high-priority customers and the overall capacity utilization efficiency.

[0072] The analysis process for customer level weighting is as follows: Get an order The historical fulfillment value score corresponds to the customer. The historical fulfillment value score is calculated based on the customer's order gross profit within a preset number of days in the past, and is weighted by time decay. Calculate the current scheduling sensitivity score for the customer. The current scheduling sensitivity score is the ratio of the order price to the minimum cost required to accept the order among all candidate vehicles. Obtain the customer's service stability coefficient, which is generated by normalizing its historical cancellation rate, complaint rate, and on-time delivery rate. After normalizing the historical performance value score, current scheduling sensitivity score, and service stability coefficient, the scores are weighted and summed according to preset weights, and then mapped to the (0,1) interval using the Sigmoid function to obtain the customer level weight. .

[0073] Among them, the historical performance value score is calculated using the exponential decay weighted average method: assuming the customer has in the past... Within a day The first order, the The gross profit per order is Then its historical performance value is divided for: ; in, ; As a decay factor, more recent orders have a higher weight, reflecting recent changes in customer value.

[0074] The current scheduling sensitivity score reflects the marginal benefit of the order under the current system state, that is, the revenue that can be brought by unit cost.

[0075] Service stability coefficient It is generated by linear combination and normalization of three indicators, for example:

[0076] The range of values ​​is: A higher value indicates more stable and reliable customer behavior.

[0077] It should be noted that customer level weighting Instead of static labels, it is a dynamic score that integrates historical performance, current scheduling environment and service quality. Its core design is to upgrade "customer value" from "manual rating" to "data-driven, context-aware" intelligent assessment. This mechanism can not only identify high-value customers, but also automatically avoid low-stability customers (such as those who frequently cancel), thereby improving overall resource utilization and business returns while ensuring service quality.

[0078] Based on the aforementioned technical solutions, in the order reallocation process after vehicle interruption, blindly allocating orders solely based on urgency can easily lead to fragmented capacity, with multiple vehicles operating at low loads, resulting in resource waste and increased costs. Conversely, pursuing full vehicle loads may delay high-priority orders and harm customer experience. Therefore, this solution adopts a collaborative allocation strategy of "urgency-driven + load-capacity-guided": orders to be allocated are sorted in descending order of urgency, incorporating both time window urgency and customer value, ensuring that high-risk, high-value tasks are prioritized; under the premise of satisfying all soft and hard constraints (such as load capacity, time window, driving time, etc.), orders are allocated to candidate vehicles with the highest current load-capacity score, i.e., vehicles that are already loaded but still have spare capacity. This strategy maximizes the utilization rate of single-vehicle resources while ensuring service priority, effectively avoiding "one vehicle, one order" or empty runs. Compared to single-objective allocation methods, this solution achieves Pareto optimization of service quality and operational efficiency, improving the on-time delivery rate of high-priority orders while reducing overall transportation costs and carbon emissions. It combines commercial rationality and engineering practicality, and is a key mechanism for intelligent scheduling systems to achieve refined and green operation.

[0079] S502. After completing the initial allocation, check if any candidate vehicle has a time window violation or load overload conflict; if so, then in the local subgraph formed by the conflicting candidate vehicle and other candidate vehicles within its one-hop or two-hop adjacency range, perform up to two rounds of order exchange within the local subgraph to eliminate the local conflict and generate a feasible allocation scheme.

[0080] Among them, a one-hop adjacent candidate vehicle is a candidate vehicle that is directly connected to the conflicting vehicle in the mission migration feasibility map; a two-hop adjacent candidate vehicle is a candidate vehicle that is not directly connected to the conflicting vehicle but has at least one common adjacent vehicle.

[0081] It should be noted that if a feasible reallocation plan cannot be generated, the remaining unallocated orders will be marked as high-disturbance orders, triggering the global lightweight rescheduling module. The global lightweight rescheduling module only optimizes the sub-path segments containing high-disturbance orders, rather than replanning the entire network.

[0082] Among them, a one-hop adjacent candidate vehicle is a candidate vehicle that is directly connected to the conflicting vehicle through a directed edge in the task migration feasibility graph, indicating that it was evaluated as capable of accepting the relevant order in the initial allocation phase; a two-hop adjacent candidate vehicle is a candidate vehicle that is not directly connected to the conflicting vehicle, but has at least one common adjacent vehicle (i.e., there is a connection with a path length of 2); a local subgraph is an induced subgraph composed of the conflicting vehicle, its one-hop adjacent vehicle, and its two-hop adjacent vehicle; "at most two rounds of order exchange" means that the backtracking search depth does not exceed 2: the first round attempts to transfer the conflicting order from the original vehicle to the one-hop adjacent vehicle. If it leads to a new conflict, the second round allows the adjacent vehicle to further transfer its own order to the two-hop adjacent vehicle, forming a two-level adjustment chain A→B→C. Chain operations exceeding two steps will not be executed.

[0083] In some implementations, conflict detection is performed in real time after each order insertion. This includes checking whether the total vehicle load exceeds the rated limit, whether the total volume is over-capacity, and whether the estimated delivery time of all orders (including newly inserted orders) on the path falls within their respective customer time windows. The construction of local subgraphs can be completed in constant time using graph traversal algorithms (such as BFS). The order exchange search adopts a recursive backtracking strategy, verifying the satisfaction of hard constraints at each step and pruning infeasible branches. If a feasible solution is found at depth 1, the exploration at depth 2 is terminated early to improve efficiency.

[0084] It should be noted that if, after exhausting all exchange combinations with a depth not exceeding 2 within a local subgraph, a feasible reallocation scheme satisfying all soft and hard constraints still cannot be generated, the conflict is determined to be highly perturbed. The system marks the remaining unassigned orders as highly perturbed orders and triggers the restricted range rescheduling module. The restricted range rescheduling module only performs joint optimization on local path segments containing highly perturbed orders and their associated vehicles (including the original target vehicle and all candidate vehicles that have participated in the allocation attempt). For example, it replans the sub-paths between these vehicles from their current position to the completion of the highly perturbed order, while the predetermined tasks and paths of the remaining vehicles remain unchanged. This avoids the high computational overhead and scheduling oscillations caused by full-network path replanning and ensures that the system maintains a second-level response capability under strong robustness.

[0085] Based on the above technical solutions, in the dynamic order reallocation process, although the initial allocation is based on compatibility scores and full-load strategies, conflicts such as time window violations or overload may still occur due to local coupling effects (e.g., multiple orders inserted into the same path). Directly backtracking and retrying is inefficient; triggering a full-network rescheduling results in high computational overhead and severe system jitter, making it difficult to meet real-time requirements. Therefore, this solution proposes a local conflict repair mechanism: once a conflict is detected, order exchange operations with a depth not exceeding two rounds are performed only within a local subgraph consisting of the conflicting vehicle and its one-hop and two-hop adjacent vehicles. This design precisely limits the repair scope; the one-hop vehicle is the original feasible candidate, and the two-hop vehicle provides indirect adjustment space, preserving problem relevance while avoiding ineffective expansion. By limiting the search depth to two rounds, backtracking attempts are ensured to be completed within milliseconds, balancing solution quality and response speed. Compared to global optimization, this method significantly reduces computational complexity; compared to abandoning conflicting orders, it effectively salvages high-value tasks. The system achieves robust fault tolerance capabilities with "small scope, fast response, and high success rate," representing a key technological breakthrough in ensuring the stability and feasibility of large-scale real-time scheduling systems.

[0086] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

Claims

1. A logistics transportation management method based on artificial intelligence, characterized in that, include: Obtain the current set of orders awaiting delivery and their real-time location information for the target vehicle; wherein, the target vehicle is the vehicle whose transportation task has been interrupted; Based on preset hard constraints, candidate vehicles that are geographically close to the target vehicle and meet the carrying capacity are selected, and a task migration feasibility map is constructed. The nodes of the task migration feasibility map represent the candidate vehicles and the orders to be delivered. The edge weights are determined based on the carrying cost of the orders carried by the candidate vehicles. The carrying cost is obtained by weighted summation of detour cost, timeliness breach cost, and service priority compensation. The task migration feasibility map is input into a pre-trained lightweight graph neural network model, which outputs the compatibility score of each candidate vehicle for each order to be delivered in the order set. Based on the compatibility score, the urgency of the orders to be delivered is assessed; based on the urgency of the orders to be delivered, the delivery task plan for each candidate vehicle is updated.

2. The logistics transportation management method based on artificial intelligence according to claim 1, characterized in that, The method for constructing the task migration feasibility map includes: Each order in the set of orders to be delivered to the target vehicle is taken as an order node, and each candidate vehicle in the set of candidate vehicles is taken as a vehicle node, together forming a node set of the graph; For each order node and each vehicle node, determine whether the preset soft constraint compatibility conditions are met; if yes, then establish a directed edge between the order node and the vehicle node; otherwise, do not establish a directed edge between them; the soft constraint compatibility conditions include: there is at least one path insertion position such that the time window of the order is not violated and the vehicle is not blacklisted by the customer corresponding to the order; if satisfied, then establish a directed edge between them. For each established directed edge, calculate its edge weight; Configure a node feature vector for each order node, which includes order weight, volume, start and end time of time window, and customer geographical location; configure a node feature vector for each vehicle node, which includes vehicle current location, remaining load capacity, remaining cargo hold volume, current cumulative driving time, and historical order fulfillment rate; Based on the node set, directed edges, edge weights, and node feature vectors, a feasibility map for task migration is generated.

3. The logistics transportation management method based on artificial intelligence according to claim 2, characterized in that, The method for obtaining the edge weights includes: Calculate the detour distance increment caused by inserting the order into the current delivery route of the candidate vehicle, and use the ratio of the detour distance increment to a preset baseline distance as the detour cost; The estimated delivery time of the order is calculated based on the inserted path. If the estimated delivery time exceeds the deadline of the time window specified by the customer, the ratio of the timeout amount to the width of the time window of the order is used as the time-out penalty; otherwise, the time-out penalty is set to zero. Obtain the normalized priority coefficient of the customer corresponding to the order. and will As a service priority compensation item; The detour cost, time-sensitive breach of contract cost, and service priority compensation item are weighted and summed to obtain the acceptance cost, and the acceptance cost is assigned as the weight of the edge.

4. The logistics transportation management method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the compatibility score includes: The graph neural network model is used to perform multi-round message passing on the order nodes and vehicle nodes in the task migration feasibility map to generate the embedding vectors of each order node and each vehicle node. Based on the embedding vector of any vehicle node, the embedding vector of the corresponding order node, and the features of the edge between them, the compatibility score of the vehicle for the order is calculated through a preset compatibility prediction module. The compatibility prediction module is a multilayer perceptron containing at least one fully connected layer, and its output, after being processed by an activation function, represents the probability of matching feasibility.

5. The logistics transportation management method based on artificial intelligence according to claim 1, characterized in that, The updated delivery task plan for each candidate vehicle includes: After sorting the orders to be assigned in descending order of urgency, they are assigned sequentially to the candidate vehicles with the highest current load capacity score that still meet all soft and hard constraints after the order is assigned; wherein, the load capacity score is used to characterize the resource utilization efficiency of the candidate vehicle after accepting the order; After the initial allocation is completed, it is checked whether any candidate vehicle has a time window violation or load overload conflict. If so, then in the local subgraph formed by the conflicting candidate vehicle and other candidate vehicles within its one-hop or two-hop adjacency range, at most two rounds of order exchange are performed within the local subgraph to eliminate the local conflict and generate a feasible allocation scheme. Here, the one-hop adjacency candidate vehicle is the candidate vehicle that is directly connected to the conflicting vehicle in the task migration feasibility graph; the two-hop adjacency candidate vehicle is the candidate vehicle that is not directly connected to the conflicting vehicle but has at least one common adjacency vehicle.

6. The logistics transportation management method based on artificial intelligence according to claim 5, characterized in that, The method for obtaining the load factor score includes: Based on the total weight and volume of the cargo currently loaded on the vehicle, divide them by the maximum allowable load capacity and the maximum cargo hold volume respectively to obtain the weight utilization rate and volume utilization rate. Then, sum them according to the preset weights to obtain the full load rating of the candidate vehicle.

7. The logistics transportation management method based on artificial intelligence according to claim 5, characterized in that, The assessment of the urgency of orders to be delivered includes: Given the time urgency and customer importance, for any order to be reassigned Its urgency Defined as: ； in, Customer level weighting; This is the current timestamp of the scheduling system; For orders The deadline for the time window specified by the customer; The shortest possible travel time from the current location of the target vehicle or candidate vehicle to the order's destination; It is a very small positive number, used to prevent the denominator from being zero and causing numerical anomalies.

8. The logistics transportation management method based on artificial intelligence according to claim 7, characterized in that, The method for determining the customer level weight includes: Get an order The historical performance value score corresponds to the customer. The historical performance value score is calculated based on the customer's order gross profit within a preset number of days in the past and is weighted by time decay. Calculate the current scheduling sensitivity score for the customer, where the current scheduling sensitivity score is the ratio of the order price to the minimum cost required to accept the order among all candidate vehicles; Obtain the customer's service stability coefficient, which is generated based on the normalization of its historical cancellation rate, complaint rate, and on-time delivery rate. After normalizing the historical performance value score, current scheduling sensitivity score, and service stability coefficient, the scores are weighted and summed according to preset weights, and then mapped to the (0,1) interval using the Sigmoid function to obtain the customer level weight. .

9. The logistics transportation management method based on artificial intelligence according to claim 1, characterized in that, The method for screening candidate vehicles includes: Obtain the real-time location of all other vehicles in the dispatch system that are not in a faulty state; Calculate the road network distance from each vehicle to the target vehicle, and retain vehicles whose road network distance does not exceed a preset neighborhood radius threshold as the initial screening vehicle set; For each vehicle in the initial screening vehicle set, the following hard constraints are checked sequentially: The remaining load capacity of the vehicles initially screened is not less than the total weight of the set of orders to be delivered; The remaining cargo hold volume of the initially screened vehicles shall not be less than the total volume of the set of orders to be delivered; After inserting any order from the set of orders to be delivered into the current delivery route of the initial screening vehicle, there is at least one insertion position such that the estimated delivery time of the order falls within the time window specified by its customer. The cumulative driving time of the vehicles screened in the initial screening after accepting new orders shall not exceed the legally mandated continuous driving time limit; The initial screening process ensures that the vehicle model and onboard equipment configuration match the transportation requirements of all orders in the set of orders to be delivered. When all hard constraints are met, the vehicles initially screened are considered candidate vehicles.

10. An artificial intelligence-based logistics transportation management system, operating based on the artificial intelligence-based logistics transportation management method according to any one of claims 1-9, characterized in that, This includes modules for acquiring, building, and allocating data; The acquisition module is used to acquire the current set of orders to be delivered and the real-time location information of the target vehicle; The construction module is used to filter out candidate vehicles that are geographically close to the target vehicle and meet the carrying capacity based on preset hard constraints, and to construct a task migration feasibility map. The allocation module is used to input the task migration feasibility map into a pre-trained lightweight graph neural network model and output the compatibility score of each candidate vehicle for each order to be delivered in the order set. Based on the compatibility score, the urgency of the orders to be delivered is assessed; Based on the urgency of the orders to be delivered, update the delivery task plan for each candidate vehicle.

Images