A vehicle dynamic scheduling method and system based on graph embedding grey wolf constraint

Abstract

The present application relates to the technical field of intelligent logistics operations optimization, and discloses a grey wolf constrained vehicle dynamic scheduling method based on graph embedding, comprising: receiving a vehicle scheduling request; reading structure constraint parameters related to the vehicle scheduling request from an embedding vector database, the structure constraint parameters representing the importance of node structure and the correlation between nodes in a transportation network; generating an initial population according to the structure constraint parameters and path search structure constraint rules determined by the structure constraint parameters, taking the initial population as a current population, and performing iterative convergence to output a Pareto optimal path solution set; detecting dynamic events, analyzing the events, and delimiting local subgraphs; reusing path search structure constraint rules and structure constraint parameters to generate a local Pareto optimal solution set, and selecting and replacing a sub-path to update an original path, and then continuing execution. The present application solves the fundamental defects of existing vehicle scheduling methods in terms of structural knowledge loss and slow dynamic response, and comprehensively improves the optimization calculation efficiency and dynamic response capability of vehicle scheduling schemes.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics operations optimization technology, and in particular to a vehicle dynamic scheduling method and system based on graph embedding gray wolf constraints. Background Technology

[0002] In multimodal transport and complex transportation network environments, the vehicle scheduling problem typically manifests as a multi-objective path optimization challenge under large-scale networks, strong constraints, and dynamic changes. In complex networks encompassing various transportation modes, node types, and dynamic transport, scheduling objectives usually include transportation cost, transportation time, reliability, and combinations thereof. The solution space exhibits overall structural dependence and strong semantic correlation; the formation of the optimal path depends not only on the cost weights of nodes and edges but also on network topology and semantic constraints. Existing techniques often employ metaheuristic algorithms to abstract the vehicle scheduling problem into a black-box function, ignoring the risk of random nodes in the transportation network disrupting critical path connectivity. These algorithms use structure-independent mathematical convergence mechanisms, relying on objective function evaluation, resulting in significant search space redundancy and low convergence efficiency. Even if the numerical objectives of the optimal path are closest, the structural over-reliance on a single vulnerable edge limits the stability of the solution set. Furthermore, the algorithms are prone to getting trapped in suboptimal local optima and incur high costs for replanning dynamic events.

[0003] Existing techniques typically use graph embedding results as state features or policy inputs. When a sudden dynamic event occurs in the network structure, local subgraphs of the affected nodes are extracted and re-input into a graph attention network for online embedding learning. However, forward propagation of local subgraphs requires complex multi-round message passing, and online iteration by calling the original network model is slow and computationally expensive, which is not conducive to the system's rapid response to dynamic events. Furthermore, dynamic events are infinitely diverse, and the mapping between events and states may be well distributed during local training, but the generalization ability drops sharply for dynamic events that occur for the first time in the actual vehicle scheduling process or for complex dynamic events. Summary of the Invention

[0004] The purpose of this invention is to provide a vehicle scheduling method that integrates the structural knowledge of multimodal transport graph neural networks into the gray wolf optimization search process, taking into account both multi-objective Pareto optimization and rapid response to dynamic events. The aim is to solve the technical problems of existing technologies that treat the vehicle scheduling problem as a black box, ignoring the structural stability of the transportation network; and the slow speed of online iteration of the original network model and the high cost of replanning when encountering sudden dynamic events.

[0005] In a first aspect, embodiments of the present invention provide a vehicle dynamic scheduling method based on graph embedding gray wolf constraints, comprising:

[0006] Receive vehicle dispatch requests that include origin, destination, cargo information, and time requirements;

[0007] Retrieves structural constraint parameters related to vehicle dispatch requests from the embedded vector database, which characterize the structural importance of nodes and the inter-node relationships in the transportation network.

[0008] Based on the structural constraint parameters and the determined path search structural constraint rules, a biased random walk is performed to generate an initial population composed of candidate paths, which serves as the current population. The following steps are then iterated until convergence:

[0009] Calculate the objective function value of each path in the current population on the preset scheduling target, and perform non-dominated sorting. Calculate the Pareto ranking of each path through non-dominated sorting and crowding degree, and calculate the cosine similarity between the average node embedding of each path and the predefined reference embedding as the structural potential score. Select α, β, and δ leader wolves based on the weighted sum of Pareto ranking and structural potential score.

[0010] Based on the preset probability corresponding to each leader wolf, cross segments are randomly selected from the paths of each leader wolf to replace the corresponding segments in the other paths. The mutation probability obtained by nonlinear convergence factor mapping is used to perform node mutation or transportation mode replacement on the replaced paths to generate a new generation of population as the current population for the next iteration.

[0011] After the population iteration converges, a set of Pareto optimal path solutions is output, and one of them is selected as the current path and executed.

[0012] Detect dynamic events and determine the local sub-paths they affect in the current path; analyze the scope of the event's influence and delineate a local subgraph centered on the local sub-paths; reuse the path search structure constraint rules and the structure constraint parameters to generate a local initial population within the local subgraph, generate a local Pareto optimal solution set through a preset finite number of iterations, and select alternative sub-paths from it to replace the local sub-paths, and continue execution with the updated scheduling path.

[0013] Secondly, embodiments of the present invention provide a vehicle dynamic scheduling system based on graph embedding gray wolf constraints, comprising:

[0014] The dispatch request receiving module is used to receive vehicle dispatch requests that include origin, destination, cargo information and time requirements;

[0015] The structural constraint module is used to read structural constraint parameters related to vehicle scheduling requests from the embedded vector database, which characterize the structural importance of nodes in the transportation network and the correlation between nodes.

[0016] The gray wolf search optimization module is used to perform biased random walks based on the structural constraint parameters and the determined path search structural constraint rules, generate an initial population composed of candidate paths as the current population, and iterate the following steps until convergence:

[0017] Calculate the objective function value of each path in the current population on the preset scheduling target, and perform non-dominated sorting. Calculate the Pareto ranking of each path through non-dominated sorting and crowding degree, and calculate the cosine similarity between the average node embedding of each path and the predefined reference embedding as the structural potential score. Select α, β, and δ leader wolves based on the weighted sum of Pareto ranking and structural potential score.

[0018] Based on the preset probability corresponding to each leader wolf, cross segments are randomly selected from the paths of each leader wolf to replace the corresponding segments in the other paths. The mutation probability obtained by nonlinear convergence factor mapping is used to perform node mutation or transportation mode replacement on the replaced paths to generate a new generation of population as the current population for the next iteration.

[0019] After the population iteration converges, a set of Pareto optimal path solutions is output, and one of them is selected as the current path and executed.

[0020] The local re-optimization module is used to detect dynamic events and determine the local sub-paths affected by them in the current path; analyze the scope of the event's influence and delineate a local subgraph centered on the local sub-paths; reuse the path search structure constraint rules and the structure constraint parameters to generate a local initial population in the local subgraph, generate a local Pareto optimal solution set through a preset finite number of iterations, and select alternative sub-paths from it to replace the local sub-paths, so that the updated scheduling path can continue to be executed.

[0021] Thirdly, embodiments of the present invention provide an electronic device, including:

[0022] One or more processors;

[0023] Memory, used to store one or more programs;

[0024] When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.

[0025] Fourthly, embodiments of the present invention provide a computer-readable storage medium that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform the method described in the first aspect.

[0026] The technical solution provided by this invention includes receiving a vehicle scheduling request that includes a starting point, destination, cargo information, and time requirements; reading structural constraint parameters related to the vehicle scheduling request and characterizing the importance and correlation of nodes in the transportation network from an embedded vector database; performing a biased random walk based on the structural constraint parameters and the path search structural constraint rules they determine, generating an initial population composed of candidate paths; iterating through the initial population, outputting a Pareto optimal path solution set, and selecting one of these as the current path for execution; when a dynamic event is detected, determining a local sub-path and analyzing its impact range, and delineating a local subgraph; reusing the path search structural constraint rules, performing a search within the local subgraph and iteratively generating a Pareto optimal solution set, selecting an alternative sub-path, and updating the scheduling path to continue execution.

[0027] Compared with existing technologies, this invention utilizes a separation mechanism between offline learning and online use. While ensuring in-depth utilization of structural knowledge, it reduces the computational latency of online scheduling to the level of database queries, thus balancing the dual requirements of knowledge depth and response time. Next, by embedding the learned graph neural network structural knowledge into the graph neural network, it deeply integrates it into key stages of multi-objective optimization algorithms, such as population initialization, leader wolf selection and evaluation, discrete position updates, and rolling time-domain dynamic re-optimization. Under the guidance of the leader wolf, the population evolves in a direction that simultaneously satisfies objective optimization and structural rationality, effectively avoiding the search direction deviation caused by existing black-box functions that only focus on the objective function value while ignoring path structure quality. This systematically solves the fundamental defects of existing vehicle scheduling methods in terms of missing structural knowledge, low search efficiency, subjective multi-objective trade-offs, and slow dynamic response, achieving a comprehensive improvement in the optimization computational efficiency and dynamic event response capability of vehicle scheduling schemes. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall workflow and data flow of the present invention.

[0029] Figure 2 This is a flowchart illustrating a vehicle dynamic scheduling method based on graph embedding gray wolf constraints according to the present invention.

[0030] Figure 3 This is a flowchart of the gray wolf search optimization algorithm of the present invention.

[0031] Figure 4 This is a timing diagram of the dynamic response mechanism for local re-optimization in this invention.

[0032] Figure 5 This is a schematic diagram of the heterogeneous graph attention network model of the present invention.

[0033] Figure 6This is a schematic diagram of the structure of a vehicle dynamic scheduling system based on graph embedding gray wolf constraints according to the present invention. Detailed Implementation

[0034] The present invention will be described in detail below through embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0035] In this invention, the terms "in one possible embodiment," "exemplary," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "in one possible embodiment," "exemplary," or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "in one possible embodiment," "exemplary," or "for example" is intended to present the relevant concepts in a specific manner.

[0036] In multimodal transport scenarios, facing complex transportation networks and ever-changing real-time environments, how to efficiently and intelligently perform dynamic vehicle scheduling is a key issue in improving supply chain operational efficiency and reducing logistics costs.

[0037] However, existing vehicle scheduling methods typically employ rule-based or experience-based static path planning, which fails to adequately consider the differentiated importance of different nodes and edges in the transportation network topology, resulting in low path search efficiency. Traditional multi-objective optimization algorithms often use completely random initialization of the population, lacking utilization of the transportation network structure, leading to slow convergence and getting trapped in local optima. When facing dynamic events during transportation, existing scheduling systems often require online retraining of local paths, resulting in slow response speeds, high computational resource consumption, and difficulty in meeting the timeliness requirements of real-time scheduling.

[0038] Therefore, there is an urgent need in this field for a vehicle scheduling scheme that can deeply integrate the structural knowledge of transportation networks into the optimization search process and take into account both multi-objective Pareto optimization and rapid response to dynamic events.

[0039] To this end, the present invention utilizes graph neural networks to perform offline embedding learning on transportation networks, extracting structural knowledge of nodes and edges as constraint parameters. When a scheduling request arrives, only a database read operation is needed to obtain the structural knowledge, without the need to run graph neural network inference in real time.

[0040] The vehicle dynamic scheduling method based on graph embedding gray wolf constraints provided by this invention is essentially about integrating structural knowledge into the population initialization, leader wolf selection, and discrete position update of the multi-objective gray wolf optimization algorithm to achieve structure-aware intelligent path search. To clarify the solution, we will first combine... Figure 1The overall workflow of this invention embodiment will be described in detail below. The workflow will be explained in detail in two stages.

[0041] Figure 1 This is a schematic diagram of the overall workflow and data flow of the present invention. It is mainly divided into two stages: the first stage is the offline learning and preparation stage, and the second stage is the online optimization, execution, and dynamic response stage.

[0042] Figure 1 The first phase of the project aims to extract and persistently store the structural knowledge of the transportation network, providing a data foundation for the current stage. The specific data flow is as follows:

[0043] First, network modeling and data preparation are performed, collecting historical dynamic and static data. The multimodal transport network is abstracted and constructed as a directed weighted graph, with transport hubs as nodes (e.g., ports, warehouses, distribution centers) and transport modes as directed edges (e.g., road, rail, waterway, air). The node feature vectors of each node and the edge feature vectors of each directed edge are extracted and initialized as basic graph data.

[0044] Then, the base graph data is input into the GNN model training and embedding generation module. This module uses a heterogeneous graph attention network to train the base graph data offline, iteratively updates the hidden state of each node through a multi-layer message passing mechanism, learns the topological structure information of the transportation network, the functional semantic information of nodes, and the directed edge transportation pattern relationship information, and obtains the results through forward propagation.

[0045] Finally, the node embedding matrix and edge embedding matrix obtained from the training, as well as the node structural importance parameters and inter-node structural correlation parameters calculated from the node embedding matrix and edge embedding matrix, are associated and stored in the GNN embedding database.

[0046] Figure 1 The goal of the second phase is to leverage the structural knowledge generated offline in the first phase to perform path optimization, plan execution, and dynamic event vectors in real-time scheduling scenarios. The specific data flow is as follows:

[0047] The first step is to accept scheduling requests and perform initial optimization. After receiving vehicle scheduling requests that include the origin, destination, cargo information, and time requirements, the system reads the structural constraint parameters related to the request from the GNN embedding database. Based on the structural constraint parameters, the system determines the path search structural constraint rules to generate an initial population. Then, it calls the GNN-MOGWO (GNN stands for Graph Neural Network; MOGWO stands for Multi-Objective Grey Wolf Optimizer) optimization engine to iteratively optimize the initial population and obtain a Pareto optimal path solution set.

[0048] The second step is scheme selection and plan issuance. From the set of Pareto optimal path schemes, a Pareto optimal path is selected according to the current scheduling preference. The scheduling preference can be set in advance to prioritize ensuring arrival time, followed by standard cost, but this is not limited here. The path sequence of this scheme can be: origin port → road transportation → regional warehousing center → rail transportation → urban distribution center → road transportation → destination delivery.

[0049] The third step involves rolling time-domain monitoring and dynamic event detection. For example, during route execution, the system continuously monitors the real-time status of the transportation network. When goods arrive at the regional warehousing center, the system detects a dynamic event: the next segment of the railway transportation line is temporarily interrupted due to equipment failure, with an estimated repair time of 6 hours. This dynamic event will cause the original route to be unable to execute on time.

[0050] The fourth step is local re-optimization and plan repair. Specifically, the local sub-path affected by the dynamic event in the current execution path is identified as the path segment from the regional warehousing center to the city distribution center, and the corresponding local scope of influence is determined. The system reuses existing path search structure constraint rules and parameters from the GNN embedded database and calls the GNN-MOGWO optimization engine again within this local scope. This local scope only includes the regional warehousing center, the city distribution center, and several intermediate nodes reachable between them.

[0051] The GNN-MOGWO optimization engine generates a local initial population within this local range, and iterates through a preset finite number of iterations, which is much smaller than the initial optimization iterations. Then, it generates a local Pareto optimal solution set and selects an alternative sub-path from it.

[0052] The fifth step is to update and continuously monitor the process. Specifically, the affected local sub-paths in the original execution path are replaced with alternative sub-paths to form the updated complete path. The updated path scheme is then issued for execution, and continuous rolling time-domain monitoring is performed.

[0053] If new dynamic events are detected during subsequent monitoring, the local re-optimization process in step four is repeated to form a closed-loop dynamic response mechanism of "monitoring - detection - re-optimization - update - monitoring".

[0054] Compared with the prior art, the technical solution provided by the present invention has at least the following technical effects:

[0055] 1. Significantly improved search efficiency. This invention uses graph neural networks to learn the topological structure of transportation networks offline, quantifying it into node structural importance parameters and inter-node structural correlation parameters. Based on this, path search structural constraint rules are generated, and the search space is pruned based on structural knowledge, achieving an improvement from indiscriminate traversal to targeted intelligent search.

[0056] 2. Significantly improved initial solution quality. This invention generates an initial population through path search structural constraint rules, ensuring that each path in the initial population tends to pass through key hubs with high structural importance, and performs structural filtering on transportation patterns. This provides a higher-quality search starting point for subsequent iterative optimization, effectively accelerating the overall convergence process of the algorithm.

[0057] 3. Leader Wolf Selection Considers Both Objective Performance and Structural Potential. This invention introduces the structural potential score obtained from GNN embedding learning into the leader wolf selection mechanism. By selecting the leader wolf through a weighted sum of Pareto ranking and structural potential score, the selected leader wolf not only performs well on Pareto objectives but also possesses higher potential at the structural level of the transportation network. This guides the population to evolve in a direction that simultaneously satisfies objective optimization and structural rationality, effectively avoiding the search direction bias caused by traditional methods that only focus on the objective function value and ignore the quality of path structure.

[0058] 4. Significantly improved response speed for dynamic events. This invention utilizes a local re-optimization mechanism to perform a limited number of iterations of re-optimization only within the subgraph corresponding to the local sub-path affected by dynamic events. This reduces the number of iterations from full convergence to a limited number, thereby improving the response speed of path adjustments for dynamic events while ensuring the quality of alternative solutions.

[0059] After a detailed description of the overall workflow of the present invention, the following will describe in detail a vehicle dynamic scheduling method based on graph embedding gray wolf constraints provided by an embodiment of the present invention.

[0060] like Figure 2 As shown in the figure, a vehicle dynamic scheduling method based on graph embedding gray wolf constraints provided by an embodiment of the present invention may include the following steps:

[0061] S210 receives vehicle dispatch requests that include origin, destination, cargo information, and time requirements.

[0062] Specifically, this step is the interface between the scheduling system and the upstream business system. The system receives vehicle scheduling requests from the transportation management system, order management system, or manual scheduling terminal. These requests include: origin (the transportation hub from which the goods depart); destination (the target transportation hub to which the goods need to be delivered); goods information, including attributes such as goods type, weight, volume, and whether they are cold chain goods or dangerous goods; and time requirements (the goods must arrive at their destination before a specified time). These four types of information collectively constitute the input boundary conditions for all subsequent calculation steps, defining the spatial scope, business constraints, and time constraints of the scheduling task.

[0063] S220: Read structural constraint parameters related to vehicle scheduling requests from the embedded vector database, which characterize the structural importance of nodes in the transportation network and the correlation between nodes.

[0064] Specifically, the system determines a set of candidate nodes reachable within a preset number of hops based on the origin and destination in the vehicle dispatch request and the directed weighted graph of basic multimodal transport. The directed weighted graph of basic multimodal transport is constructed during the offline learning phase, in which transport hubs are abstracted as nodes in the graph, transport modes are abstracted as directed edges in the graph, and the weight of the edges represents attributes such as transport cost, time, or reliability.

[0065] Next, two types of structural constraint parameters are read from a pre-built embedding vector database. This database is generated during the offline learning phase by training the directed weighted graph of the transport network using a heterogeneous graph attention network with multi-layer message passing. During training, each node updates its hidden state by repeatedly receiving and aggregating information from its neighbors. After iteration, each node's final embedding vector no longer represents only its own attributes but also encodes the node's complete topological structure information within its multi-hop neighborhood, the functional semantic information of its neighbors, and the transport pattern relationship information of the connecting edges.

[0066] The first type of structural constraint parameter is the node structural importance parameter. This parameter measures the criticality of a single node in the transportation network topology. During the training of graph attention networks, nodes located in key positions in the network topology, such as highly connected hubs or essential nodes on multiple transportation paths, receive more information aggregation from neighboring nodes in multi-layer message passing, resulting in more fully activated values ​​in each dimension of their final embedding vector, leading to a larger magnitude of the embedding vector. Therefore, the L2 norm of each node's final embedding vector is normalized to the maximum value across the entire network and used as the structural importance parameter for that node. After normalization, the parameter ranges from 0 to 1, with a larger value indicating higher structural importance of the node in the transportation network. For example, the structural importance parameter of key nodes such as large ports and core railway hubs is usually close to 1, while the parameter of small, end-point distribution points is usually lower.

[0067] The second type of structural constraint parameter is the inter-node structural correlation parameter. This parameter measures the degree of correlation between two nodes at the transportation network structure level. During the training of graph attention networks, nodes with similar functions or structural roles in the network topology are considered. For example, nodes that are both large ports and located in the core of the logistics corridor receive similar neighborhood information through message passing, and are thus mapped to similar directions in the embedding space.

[0068] Therefore, the cosine similarity between the final embedding vectors of two nodes is used as a parameter for the structural association between nodes. Cosine similarity measures the consistency of the orientation of two vectors in high-dimensional space, and its value ranges from -1 to +1. The closer the value is to +1, the more consistent the orientation of the two nodes in the embedding space, and the more suitable they are for connecting each other in the transportation path; the closer the value is to 0 or -1, the greater the difference in the structural roles of the two nodes, and the less suitable they are for direct connection in the path.

[0069] S230, based on the structural constraint parameters and the determined path search structural constraint rules, generates an initial population as the current population, and iterates the gray wolf optimization search until convergence, outputting a set of Pareto optimal path solutions, such as... Figure 3 As shown.

[0070] Specifically, this step is the core of the entire vehicle scheduling method, comprising two sub-processes: initial population generation and iterative optimization. The role of the path search structural constraint rules is to impose structural knowledge-based constraints on the range of candidate nodes and the jump relationships between nodes, thus limiting the path search space to a structurally reasonable range.

[0071] The biased walk process is as follows: Starting from the starting point, in each expansion step, among the candidate neighbor nodes filtered by the path search structural constraint rules and parameters, the probability of each candidate node being selected as the next hop is not equal. Instead, the structural importance parameter of each candidate node is transformed into a jump probability through a probability function, so that nodes with higher structural importance have a greater probability of being selected as the next hop, while nodes with lower structural importance retain a non-zero selection probability. After selecting the next hop node, a transportation mode is randomly selected with uniform probability from the set of feasible transportation modes between the current node and the next hop node. The selected node and transportation mode are then appended to the path sequence.

[0072] Repeat the above process until the endpoint is reached, forming a complete candidate path. Perform a biased random walk independently, utilizing the inherent randomness of probability sampling to generate several differentiated candidate paths, forming the initial population. This initial population then becomes the current population for iterative optimization. Figure 3 Population initialization is shown in the diagram.

[0073] Next, iteratively optimize as follows until convergence:

[0074] Step 1: Calculate the objective function value of each path on the preset scheduling target.

[0075] Specifically, the scheduling objectives are multiple conflicting optimization goals. For each candidate path in the current population, the objective function value for each scheduling objective is calculated separately.

[0076] Step 2: Perform non-dominated sorting, and calculate the Pareto ranking of each path using non-dominated sorting and congestion calculation.

[0077] Specifically, the non-dominated sorting process is as follows: First, traverse all candidate path pairs in the population to determine the set of paths dominated by each path and the number of paths dominating that path (i.e., the number of dominated paths). Then, all paths with a dominated number of zero are assigned to the first frontier level. These paths are not dominated by any other path in the entire population and constitute the Pareto optimal front of the current population. Next, from the set of paths dominated by the paths in the first frontier level, paths whose dominated number drops to zero are assigned to the second frontier level. This process is iterated step by step until all paths are assigned to their corresponding frontier levels. The smaller the frontier level number, the higher the Pareto quality of that path.

[0078] Crowding calculations are used to further differentiate the sparsity of distribution among paths within the same frontier level. Pareto ranking. The final determination adopts a two-level sorting rule: first, it is sorted in ascending order by frontier level number, with the smaller the level number, the higher the ranking; within the same frontier level, it is sorted in descending order by crowding degree, with the greater the crowding degree, the higher the ranking.

[0079] Step 3: Calculate the cosine similarity between the average node embedding of each path and the predefined reference embedding as the structural potential score.

[0080] Specifically, for each candidate path in the population, the average node embedding of that path is first calculated. This is the arithmetic mean of the final embedding vectors of all nodes traversed by the path across all dimensions, resulting in a vector representing the average structural position of the path in the embedding space. Figure 3 The similarity between GNNs in the dataset.

[0081] Here, the predefined reference embedding R is a vector representing the global structural features of an ideal hub in the transportation network, pre-calculated and stored after the embedding vector database is generated. Its generation method can be as follows: using the node structural importance parameters of each node in the entire network as weight coefficients, a weighted average is performed on the final embedding vectors of all nodes in the transportation network. Since the weights of nodes with higher structural importance are much greater than those of nodes with lower importance, the weighted average result is biased towards the center of the high-importance node group in the embedding space. Normalization is then performed after the weighted average, that is, the vector is scaled to a unit length to eliminate the influence of the vector magnitude.

[0082] Next, the cosine similarity between the average node embedding and the predefined reference embedding is calculated. The inner product of two vectors is divided by the product of their L2 norms. The inner product operation involves multiplying corresponding elements of the two vectors along each dimension and then summing the results, reflecting the consistency of the two vectors' directions. Dividing by the norm product eliminates the influence of vector length, retaining only directional information. The closer the structural potential score is to +1, the closer the group of nodes traversed by the path is to the structural characteristics of an ideal hub in the embedding space, indicating a higher structural advantage.

[0083] Step 4: Select based on the weighted sum of Pareto ranking and structural potential score. α, β, δ Leading wolves, such as Figure 3 The overall score shown is used to select the leader.

[0084] Specifically, the gray wolf optimization algorithm simulates the social hierarchy and hunting behavior of gray wolf packs in nature. A social hierarchy exists within gray wolf packs: α The wolf is the highest-ranking animal, representing the leader of the pack and responsible for deciding the direction of the search; β Wolves are the second tier, serving as support. α The wolf verifies the correctness of the search direction; δ Wolves are at the third level, providing additional search direction references; the remaining individuals are ω wolves, who follow the guidance of the three leader wolves in their search.

[0085] The reciprocal of the Pareto ranking of each candidate path in the population The ranking sequence is transformed into a monotonically decreasing scoring format; and weighting coefficients are used to control the relative contribution of Pareto ranking and structural potential score to the overall score. The specific formula is as follows: .

[0086] in, For the overall score, These are the weighting coefficients. The cosine similarity between the average node embedding and the predefined reference embedding. This is the reciprocal of the Pareto ranking. When the weighting coefficient approaches 1, the overall score is mainly determined by the Pareto ranking, with the structural potential score contributing less; when the weighting coefficient approaches 0, the overall score is mainly determined by the structural potential score.

[0087] Step 5: Randomly select intersecting segments from the paths of each leader wolf and replace the corresponding segments in the remaining paths. Figure 3 Position update operation in the middle.

[0088] Specifically, this operation simulates the encirclement behavior of ω wolves in a gray wolf society, where they gravitate towards the leader. For the ω wolves in the middle group other than the leader wolf (i.e., the other candidate paths), it sequentially interacts with... α Wolf 、β Wolf δ Wolves perform probabilistic crossover operations. A preset probability is used to determine whether to crossover with a leader wolf. If crossover is performed, a continuous subsequence is randomly selected from the current individual's path sequence. A subsequence containing the same start and end nodes is then searched within the corresponding leader wolf's path sequence. If a match is found, that subsequence from the current individual is replaced with the corresponding subsequence from the leader wolf, and the superior path segments contained in the leader wolf's subsequence are injected into the population. The crossover probabilities for the three leader wolves decrease in the following order: highest for α wolf, second highest for β wolf, and lowest for δ wolf.

[0089] Step 6: Using the mutation probability obtained by nonlinear convergence factor mapping, perform node mutation or transportation mode replacement on the replaced path.

[0090] Specifically, by simulating the random hunting and exploration behavior in a gray wolf society, perturbations are introduced into the existing paths to explore new solution spaces. The non-convergence factor is calculated to control the algorithm's transition from global exploration to local utilization. After mapping the non-linear convergence factor to mutation probabilities, node mutation or transportation mode replacement is performed.

[0091] When the path after iterative replacement meets the termination condition, output the Pareto optimal solution set; if the termination condition for convergence is not met, return to the step of calculating the objective function value.

[0092] S240: After the population iteration converges, output the Pareto optimal path solution set, and select one of them as the current path to be executed.

[0093] Specifically, after iterative convergence, all candidate paths at the first front level in the final population are extracted to form a Pareto optimal path set, in which no two paths have a dominance relationship. Each path represents an unimprovable optimal compromise between scheduling objectives such as transportation cost, transportation time, and reliability.

[0094] From the set of Pareto optimal solutions, a path is selected as the current execution path according to the current scheduling preference and executed. The scheduling preference can be to prioritize arrival time, reduce costs to a limited extent, or ensure reliability to a limited extent. The system selects the path with the best performance in the corresponding preference direction on the Pareto optimal front based on the scheduling preference.

[0095] S250 detects dynamic events and determines the local sub-paths affected by them in the current path.

[0096] Specifically, such as Figure 4 As shown, during the current transportation route, the system continuously monitors the real-time status of the transportation network and detects dynamic events by connecting with external data sources such as IoT sensors, GPS systems, and traffic information databases. Dynamic events include, but are not limited to, road closures, temporary port congestion, vehicle malfunctions, and route interruptions due to extreme weather. Upon detecting a dynamic event, the system assesses its impact on the current route and determines the affected local sub-paths. The starting node of each local sub-path is the nearest unaffected node in the current route that the goods have already passed through, and the ending node is the nearest reachable alternative rendezvous node after the affected area.

[0097] Next, the scope of the event's impact is analyzed, and a local subgraph is defined centered on the local subpath. Using the start and end nodes of the local subpath as boundaries, a subgraph corresponding to this local area is extracted from the directed weighted graph of the transportation network. This subgraph contains only the set of nodes reachable from the start to the end node and the set of directed edges connecting these nodes. The number of nodes in this subgraph is much smaller than the number of nodes in the global transportation network, thus limiting the computational scope of subsequent re-optimization to a smaller local region.

[0098] Then, the GNN-MOGWO engine is invoked for local optimization, reusing the path search structural constraint rules and structural constraint parameters to generate a local initial population within the local subgraph. The local Pareto optimal solution set is generated through rapid optimization iteration, and alternative sub-paths are selected from them to replace the local sub-paths, so that the updated scheduling path can continue to be executed.

[0099] Specifically, within the local subgraph, a local initial population is generated through biased random walks, starting from the beginning node of the local sub-path and ending at the end node. The same iterative optimization process as S230 is executed, but the number of iterations is limited to a preset finite number, much smaller than the total number of iterations in the initial optimization, to ensure response speed. After a finite number of iterations, a local Pareto optimal solution set is output. From the local Pareto optimal solution set, an alternative sub-path is selected according to a selection strategy consistent with the current scheduling preference; this is the optimized sub-path.

[0100] The affected path segments are replaced, a new plan is generated, and execution instructions are reissued using the updated scheduling path, while performance metrics are recorded in real time. Simultaneously, a rolling time-domain control mode is employed to continuously monitor dynamic event information.

[0101] The technical solution provided by this invention includes receiving a vehicle scheduling request that includes a starting point, destination, cargo information, and time requirements; reading structural constraint parameters related to the vehicle scheduling request and characterizing the importance and correlation of nodes in the transportation network from an embedded vector database; performing a biased random walk based on the structural constraint parameters and the path search structural constraint rules they determine, generating an initial population composed of candidate paths; iterating through the initial population, outputting a Pareto optimal path solution set, and selecting one of these as the current path for execution; when a dynamic event is detected, determining a local sub-path and analyzing its impact range, and delineating a local subgraph; reusing the path search structural constraint rules, performing a search within the local subgraph and iteratively generating a Pareto optimal solution set, selecting an alternative sub-path, and updating the scheduling path to continue execution.

[0102] Compared with existing technologies, this invention utilizes a separation mechanism between offline learning and online use. While ensuring in-depth utilization of structural knowledge, it reduces the computational latency of online scheduling to the level of database queries, thus balancing the dual requirements of knowledge depth and response time. Next, by embedding the learned graph neural network structural knowledge into the graph neural network, it deeply integrates it into key stages of multi-objective optimization algorithms, such as population initialization, leader wolf selection and evaluation, discrete position updates, and rolling time-domain dynamic re-optimization. Under the guidance of the leader wolf, the population evolves in a direction that simultaneously satisfies objective optimization and structural rationality, effectively avoiding the search direction deviation caused by existing black-box functions that only focus on the objective function value while ignoring path structure quality. This systematically solves the fundamental defects of existing vehicle scheduling methods in terms of missing structural knowledge, low search efficiency, subjective multi-objective trade-offs, and slow dynamic response, achieving a comprehensive improvement in the optimization computational efficiency and dynamic event response capability of vehicle scheduling schemes.

[0103] exist Figure 2 Based on the illustrated embodiment, as one implementation of this invention, the structural constraint parameters include node structural importance parameters and inter-node structural correlation parameters. Constructing the embedded vector database includes steps a1 to a5. The model structure and information transmission mechanism are as follows: Figure 5 As shown:

[0104] Step a1: Construct a basic multimodal transport directed weighted graph with transport hubs as nodes and transport modes as directed edges. ,like Figure 5 As shown, the node set VEach node in the graph represents a transportation hub, and the types of transportation hubs include, but are not limited to, ports, warehouses, distribution centers, railway freight stations, and airport cargo hubs; the set of directed edges... E Each edge in the diagram represents a transportation mode from one node to another, including road, rail, waterway, and air transport.

[0105] Step a2 involves extracting the node feature vector for each node in the directed weighted graph of basic multimodal transport; and extracting the edge feature vector for each directed edge.

[0106] Specifically, the node feature vector of each node is formed by concatenating static features representing the inherent attributes of the transportation hub and dynamic features representing the real-time operational status of the transportation hub. For example... Figure 5 As shown, the node features include: type, which is the category attribute of the transportation hub, using one-hot encoding to convert discrete category labels into numerical vectors; coordinates, represented by latitude and longitude; and capacity, which is the designed throughput capacity of the transportation hub.

[0107] The edge feature vector of each edge is formed by concatenating static features representing the basic attributes of the transportation mode and dynamic features representing the real-time attributes of the service. For example... Figure 5 As shown, the edge features include: transportation mode, such as highway, railway, waterway, and air, which also uses one-hot encoding to convert discrete transportation mode labels into numerical vectors; cost, i.e., the standard freight cost of the transportation mode; time, i.e., the standard transportation time of the transportation mode; and reliability, i.e., the historical probability of the transportation mode completing the transportation on time.

[0108] Step a3: Map the node feature vectors and edge feature vectors to the initial heterogeneous graph attention network and train it offline through a multi-layer message passing mechanism. In each layer of message passing, each node takes the hidden state of its neighboring nodes and the feature vectors of the connecting edges as input. The attention coefficient is obtained by calculating the feature similarity between the node and its neighboring nodes. After weighted aggregation and non-linear activation, the hidden state of the node is updated.

[0109] Each layer of message passing includes the following three steps:

[0110] First, calculate the raw attention score. , will node Transformation of hidden state, neighbors Transformation of hidden states and directed edges ,in, m For the transportation mode, the transformed feature vectors are concatenated and then subjected to linear mapping and nonlinear activation to obtain a scalar score. The mathematical expression is as follows:

[0111]

[0112] in, For nodes In the l The hidden state vector of the layer, For nodes In the l The hidden state vector of the layer, For directed edges The edge feature vectors, For the first l Layer node feature transformation matrix, For the first l The edge feature transformation matrix of the layer, For the first l The transpose of the attention parameter vector of the layer. This is a modified linear unit activation function with leakage.

[0113] Next, normalization is performed using the softmax function, and the mathematical expression is as follows:

[0114]

[0115] in, For nodes All incoming neighbor nodes gather, For natural index, For nodes To the incoming neighbor node The scalar score is obtained through linear mapping and nonlinear activation.

[0116] Second, such as Figure 5 The labeled weighted aggregated neighbor information, nodes The aggregated message is the attention-weighted sum of the hidden states of all its incoming edge neighbor nodes. The mathematical expression is as follows:

[0117]

[0118] in, The attention coefficient is obtained from the above calculation process. For neighboring nodes The hidden state after transformation matrix mapping. The higher the attention coefficient of a neighbor, the higher the weight of its information in the aggregated message and the greater its contribution; the lower the attention coefficient of a neighbor, the smaller its contribution.

[0119] Third, update the node's hidden state. The aggregated message, after being processed by a non-linear activation function, is used as the node's hidden state in the next layer, as shown in the following formula:

[0120]

[0121] Here, σ is a nonlinear activation function that enables the multilayer network to learn nonlinear combinations of node features. For nodes In the hidden state of the next layer.

[0122] Step a4: After training is complete, input the final embedding feature vectors of all nodes and directed edges into the trained heterogeneous graph attention network, and obtain the node embedding matrix H_V and the edge embedding matrix H_E through forward propagation.

[0123] Specifically, forward propagation is the process by which data flows from the input layer to the output layer along the network structure in a neural network. The edge embedding matrix integrates the structural semantic information of the two endpoints with the transport pattern attribute information of the edge itself into a multi-dimensional embedding space.

[0124] Step a5, for each node in the node embedding matrix The embedding vector is moduloed and normalized to obtain the node structure importance parameters of each node. The specific formula is as follows:

[0125]

[0126] in, For nodes The L2 norm of the final embedding vector, This represents the maximum value of the L2 norm of the final embedding vector for all nodes.

[0127] Next, the node structure correlation parameters are obtained by calculating the cosine similarity between the embedding vectors of each edge in the edge embedding matrix. The specific formula is as follows:

[0128]

[0129] Among them, among them, The inner product operation is performed on the two final embedding vectors. and These are the L2 norms of the two final embedding vectors, respectively.

[0130] Finally, the calculated node structural importance parameters and inter-node structural correlation parameters are associated and stored in the embedded vector database.

[0131] In this embodiment, during the training process of the heterogeneous graph attention network, nodes with similar functions or structural roles in the network topology, such as nodes that are both large ports or located at the core of a logistics corridor, receive similar neighborhood information patterns through multi-layer message passing, and are thus mapped to similar directions in the embedding space. This database, as the final output of the offline learning phase, is directly read and used in the online scheduling phase without requiring real-time execution of the heterogeneous graph attention network for inference computation.

[0132] exist Figure 2 Based on the illustrated embodiment, as one implementation of this invention, the path search structural constraint rules are determined according to the structural constraint parameters. These rules are obtained in any of the following ways:

[0133] The first method is to compare the node structure importance parameters of all neighboring nodes of the current node with a preset importance threshold, select neighboring nodes whose node structure importance parameters are higher than the importance threshold, and form a whitelist set corresponding to that node; in each step of the path search expansion, the next node can only be selected from the whitelist set corresponding to the current node.

[0134] Specifically, by setting an importance threshold, candidate nodes are hard-truncated and filtered, limiting the candidate range for each search step to nodes with structural importance labels. Assume that each node in the original graph has an average of... d The proportion of outgoing neighbor nodes that meet the importance threshold is: ( If ), then the search branch for each node starts from d Reduced to The importance threshold can be determined by balancing search efficiency and path reachability.

[0135] The second approach is to sort all neighboring nodes of the current node in descending order according to their importance parameters, and select the top K neighboring nodes to form a set of priority candidate nodes for the current node. In each step of the path search, node selection is performed only within the set of priority candidate nodes. Here, K is a preset positive integer.

[0136] Specifically, by retaining K optimal candidate neighbor nodes for each node, a uniform and controllable pruning of the search space is achieved. The value of K is set by balancing search efficiency and path diversity.

[0137] The third method is to preset a connection threshold. For any two nodes, if the structural correlation parameter between the nodes is greater than or equal to the connection threshold, it is determined that the node pair is allowed to connect directly; otherwise, it is determined that direct jump is prohibited.

[0138] Specifically, based on the determination results of all node pairs, a binary constraint matrix is ​​constructed between node pairs, where a matrix element of 1 indicates that the connection is allowed, and 0 indicates that the jump is prohibited. This binary constraint matrix is ​​a simulation, with the number of rows and columns equal to the total number of nodes in the network. During path search, the matrix element values ​​between the current node and the next hop candidate node are queried according to the binary constraint matrix to determine whether a direct jump is allowed, thus performing structural filtering.

[0139] This constraint is not based on simple filtering of superficial indicators such as physical distance or transportation cost, but on the structural relationships in the deep embedding space learned by the graph neural network. Node pairs with insufficient structural correlation in the embedding space may be prohibited from direct jumps even if they have direct connections on the physical network, due to their significantly different structural roles.

[0140] The fourth method: Incorporate structural correlation parameters between nodes. Input a preset non-negative monotonically decreasing penalty mapping function to obtain the jump penalty coefficient. The specific formula is shown below:

[0141]

[0142] in, For nodes With nodes The penalty coefficient for jumps between them To penalize the base coefficient, The attenuation coefficient is added to the objective function evaluation of the corresponding directed edge during path search. No connection is rigidly excluded; instead, differentiated cost penalties give implicit advantages to jump directions with high structural relevance in the optimization process.

[0143] The four methods in this embodiment of the invention can be used individually, or multiple methods can be selected and combined according to the needs of the actual scheduling scenario to apply structural constraints to different dimensions of the search space.

[0144] exist Figure 2 Based on the illustrated embodiment, as one implementation of this invention, a biased random walk is performed according to the path search structure constraint rules and structure constraint parameters to generate an initial population composed of candidate paths, including the following steps b1 to b2:

[0145] Step b1 encodes the candidate paths as a sequence of alternating nodes and transport modes.

[0146] Specifically, candidate paths X It can be represented as .in, Indicates a transportation node. Indicates the mode of transportation.

[0147] Step b2: In each step of the bias walk, the structural potential index of the generated sequence from the starting point to the current node is calculated based on the structural constraint parameters, and the selection of the next node and transportation mode is guided by the structural potential index.

[0148] Specifically, structural potential indicators are obtained through at least one of the following methods:

[0149] The first approach calculates the cumulative value of the node structural importance parameter of all nodes in the generated path sequence as a structural potential indicator. The larger the cumulative value, the higher the structural importance of the node in the transportation network topology.

[0150] The second approach involves calculating the average vector of all node embedding matrices in the generated path sequence as the sequence average embedding, and then calculating the cosine similarity between the sequence average embedding and the predefined ideal hub reference embedding to obtain the node embedding similarity aggregate value, which serves as a structural potential indicator.

[0151] Specifically, let the path be to the th j Step, the generated partial path sequence passes through the following nodes: The final embedding vector of each node is , For indexed sequences, the average embedding of the sequence is... The calculation formula is as follows:

[0152]

[0153] Among them, the predefined ideal hub reference embedding R It is a fixed vector pre-computed and stored after being generated in the embedding vector database, used to represent the global structural feature prototype of the "ideal hub" in the transportation network. (Node embedding similarity aggregation value) The calculation formula is as follows:

[0154]

[0155] The third approach divides all nodes into several preset levels based on the importance parameters of the node structure. The corresponding level sequence is "start → region → core → region → end". The level changes of edges in the generated path sequence are evaluated to determine whether they conform to a predefined level flow pattern. The proportion of edges conforming to the level flow pattern in the generated path sequence is calculated to obtain the node structure level matching degree, which serves as a structural potential indicator.

[0156] The node structure hierarchy matching degree quantifies the degree of matching between the path and the ideal hierarchy flow pattern. It can identify disordered paths in the early stage of path generation and guide the wandering direction back to a reasonable hierarchy flow pattern in a timely manner.

[0157] In this embodiment of the invention, three indicators are set to measure the quality of the path at the structural level from different perspectives. These structural potential indicators can be used individually or in combination.

[0158] exist Figure 2 Based on the illustrated embodiment, as one implementation of this invention, the objective function value of each path in the current population on the preset scheduling target is calculated, and non-dominated sorting is performed. The Pareto ranking of each path is obtained through non-dominated sorting and crowding calculation, including the following steps c1 to c5:

[0159] Step c1: Preset scheduling objectives, which are at least one combination of transportation cost, transportation time, and historical reliability.

[0160] Specifically, transportation cost is the sum of the transportation costs of each directed edge in the path segment by segment. Transportation time is the sum of the transportation time of each directed edge in the path segment by segment, plus the sum of the transit processing time of each intermediate node except the origin and destination segment by segment. The overall reliability of the path is the product of the historical reliability of each directed edge in the path. Since the reliability of each segment is a positive number not greater than 1, the product decreases as the number of path segments increases, reflecting the physical law that the longer the path or the lower the reliability of the routes it passes through, the lower the overall reliability.

[0161] Step c2, performing a non-dominated sort on the candidate paths in the current population, including:

[0162] For any two candidate paths in the initial population, if the objective function value of the first candidate path is not inferior to that of the second candidate path on all scheduling objectives and is superior to that of the second candidate path on at least one scheduling objective, then the first candidate path is determined to dominate the second candidate path; record the set of paths dominated by each candidate path and the number of paths dominated by the candidate path.

[0163] Candidate paths with a dominance of 0 are assigned to the first frontier level; candidate paths whose dominance is reduced to 0 after decreasing by 1 within the path set dominated by each candidate path in the first frontier level are assigned to the second frontier level. The same operation is performed on the paths in the second frontier level to identify the third level; the same logic is continued to iterate until all candidate paths in the population are assigned to the corresponding frontier level.

[0164] Step c3: Perform congestion calculation on candidate paths within the same frontier level.

[0165] Specifically, crowding is an indicator that measures the sparsity of candidate paths within the same frontier level in the objective function space. It includes the following steps:

[0166] First, for each scheduling objective, all candidate paths within that level are sorted in ascending order according to their objective function values.

[0167] Second, the crowding contribution of the candidate paths at both ends of the sorted sequence is set to infinity.

[0168] Third, for each candidate path in the middle of the sorting sequence, the absolute value of the difference between it and the objective function value of its two adjacent candidate paths on the scheduling objective is normalized and then added to the congestion of that candidate path, as shown in the following formula:

[0169]

[0170] in, For index labels, For the updated congestion level, The congestion level before the update. In order to achieve the goal The path to the right adjacent position in ascending order The objective function value, In order to achieve the goal The path of the left adjacent position in ascending order The objective function value, For targets within this frontier level The maximum value, For targets within this frontier level The minimum value.

[0171] Step c4: Traverse all scheduling targets to complete the accumulation and obtain the final congestion of each candidate path.

[0172] Step c5: The smaller the frontier level, the higher the Pareto ranking; within the same frontier level, the greater the congestion, the higher the Pareto ranking.

[0173] This invention unifies the optimization direction of all objectives to minimization, transforming reliability maximization into negative logarithmic reliability minimization. This transformation leverages the monotonically increasing property of the logarithmic function and its property of converting product-product relationships into summations. Pareto ranking ensures that population evolution progresses towards the Pareto optimal front, prioritizing the retention of better paths. Crowding ensures that the solutions on the Pareto front are evenly distributed and fully covered, preventing the front from degenerating into a few solutions that only cover a local region. The combination of these two criteria allows the optimization process to balance convergence towards the Pareto front with diversity evenly distributed on the Pareto front, representing the core dual objectives in multi-objective optimization algorithm design.

[0174] exist Figure 2 Based on the illustrated embodiment, as one implementation of the present invention, the mutation probability obtained by nonlinear convergence factor mapping is used to perform node mutation or transportation mode replacement on the replaced path.

[0175] Specifically, the nonlinear convergence factor for the current iteration is obtained by exponentially decreasing the ratio of the current iteration number to the total number of iterations, wherein the nonlinear convergence factor... From initial coefficients With decay rate control coefficient Joint control, the specific formula is as follows:

[0176]

[0177] in, This represents the current iteration number. This represents the total number of iterations.

[0178] The mutation probability of the current iteration number is mapped to a pre-defined monotonically decreasing mapping function. The specific formula is shown below:

[0179]

[0180] in, Let be the mutation probability in the t-th iteration. The maximum mutation probability is configured in advance. The mapping coefficient is used to control the nonlinear mapping relationship between the mutation probability and the convergence factor, and its value range is... The boundary value analysis of the mutation probability is as follows:

[0181] First, when the nonlinear convergence factor is greater than a preset threshold (which can be a value of 1), the value of the mutation probability is increased, and a larger proportion of nodes and directed edges in the population are mutated. The path structure changes significantly in each generation, driving candidate paths to explore new regions. This is beneficial for extensively exploring the search space and maintaining population diversity, preventing the algorithm from converging to a local region too early.

[0182] Second, when the nonlinear convergence factor is less than or equal to a preset threshold, which can be a value of 1, the value of the mutation probability is reduced. Only a small number of nodes and directed edges in the population are mutated. The path structure is only slightly adjusted in each generation. The search is focused on the refined region near the current Pareto front, driving the candidate path to conduct a small-range search near the optimal region. Through fine-tuning, the optimal solution is continuously approached.

[0183] This invention presents a discrete location update mechanism that integrates segment crossover and probabilistic mutation. It introduces random perturbations into the existing path and simulates the hunting and exploration behavior of a gray wolf through probability-driven node mutation and transportation mode replacement. This successfully extends the gray wolf optimization algorithm from a continuous space to a discrete combinatorial optimization space, enabling it to directly handle the discrete coding problem of multimodal transport path optimization.

[0184] This invention also provides a vehicle dynamic scheduling system 60 based on graph embedding gray wolf constraints, such as... Figure 6 As shown, it includes:

[0185] The dispatch request receiving module 610 is used to receive vehicle dispatch requests that include origin, destination, cargo information and time requirements;

[0186] The structural constraint module 620 is used to read structural constraint parameters related to vehicle scheduling requests from the embedded vector database, which characterize the structural importance of nodes in the transportation network and the correlation between nodes;

[0187] The gray wolf search optimization module 630 is used to perform a biased random walk based on the structural constraint parameters and the determined path search structural constraint rules, generate an initial population composed of candidate paths as the current population, and iterate the following steps until convergence:

[0188] Calculate the objective function value of each path in the current population on the preset scheduling target, and perform non-dominated sorting. Calculate the Pareto ranking of each path through non-dominated sorting and crowding degree, and calculate the cosine similarity between the average node embedding of each path and the predefined reference embedding as the structural potential score. Select α, β, and δ leader wolves based on the weighted sum of Pareto ranking and structural potential score.

[0189] Based on the preset probability corresponding to each leader wolf, cross segments are randomly selected from the paths of each leader wolf to replace the corresponding segments in the other paths. The mutation probability obtained by nonlinear convergence factor mapping is used to perform node mutation or transportation mode replacement on the replaced paths to generate a new generation of population as the current population for the next iteration.

[0190] After the population iteration converges, a set of Pareto optimal path solutions is output, and one of them is selected as the current path and executed.

[0191] The local re-optimization module 640 is used to detect dynamic events, determine the local sub-paths affected by the events in the current path, analyze the scope of the event's influence, delineate a local subgraph centered on the local sub-paths, reuse the path search structure constraint rules and the structure constraint parameters, generate a local initial population in the local subgraph, generate a local Pareto optimal solution set through a preset finite number of iterations, and select alternative sub-paths from the set to replace the local sub-paths, so that the updated scheduling path can continue to be executed.

[0192] This invention also provides an electronic device, comprising:

[0193] One or more processors;

[0194] Memory, used to store one or more programs;

[0195] When the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any of the above embodiments.

[0196] This invention also provides a computer-readable storage medium, characterized in that, when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is able to perform the method described in any of the above embodiments.

[0197] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A vehicle dynamic scheduling method based on graph embedding gray wolf constraints, characterized in that, include: Receive vehicle dispatch requests that include origin, destination, cargo information, and time requirements; Retrieves structural constraint parameters related to vehicle dispatch requests from the embedded vector database, which characterize the structural importance of nodes and the inter-node relationships in the transportation network. Based on the structural constraint parameters and the determined path search structural constraint rules, a biased random walk is performed to generate an initial population composed of candidate paths, which serves as the current population. The following steps are then iterated until convergence: Calculate the objective function value of each path in the current population on the preset scheduling target, and perform non-dominated sorting. Calculate the Pareto ranking of each path through non-dominated sorting and crowding degree, and calculate the cosine similarity between the average node embedding of each path and the predefined reference embedding as the structural potential score. Select α, β, and δ leader wolves based on the weighted sum of Pareto ranking and structural potential score. Based on the preset probability corresponding to each leader wolf, cross segments are randomly selected from the paths of each leader wolf to replace the corresponding segments in the other paths. The mutation probability obtained by nonlinear convergence factor mapping is used to perform node mutation or transportation mode replacement on the replaced paths to generate a new generation of population as the current population for the next iteration. After the population iteration converges, a set of Pareto optimal path solutions is output, and one of them is selected as the current path and executed. Detect dynamic events and determine the local sub-paths they affect in the current path; analyze the scope of the event's influence and delineate a local subgraph centered on the local sub-paths; reuse the path search structure constraint rules and the structure constraint parameters to generate a local initial population within the local subgraph, generate a local Pareto optimal solution set through a preset finite number of iterations, and select alternative sub-paths from it to replace the local sub-paths, and continue execution with the updated scheduling path.

2. The method according to claim 1, characterized in that, The structural constraint parameters include node structural importance parameters and inter-node structural correlation parameters. The embedded vector database is pre-generated by the following steps: Construct a basic multimodal transport directed weighted graph with transport hubs as nodes and transport modes as directed edges; For each node in the basic multimodal transport directed weighted graph, a node feature vector is extracted. The node feature vector includes: static node features representing the inherent attributes of the transport hub, and dynamic node features representing the real-time operating status of the transport hub. For each directed edge, an edge feature vector is extracted. The edge feature vector includes: edge static features representing the basic attributes of the transportation mode, and edge dynamic features representing the real-time state of the transportation mode. The node feature vector and the edge feature vector are input into the heterogeneous graph attention network, and offline training is performed through a multi-layer message passing mechanism. In each layer of message passing, each node takes the hidden state of its neighboring nodes and the feature vector of the connecting edges as input. The attention coefficient is obtained by calculating the feature similarity between the node and its neighboring nodes, and the hidden state of the node is updated after weighted aggregation and nonlinear activation. After training, the final embedding feature vectors of all nodes and directed edges are input into the trained heterogeneous graph attention network, and the node embedding matrix and edge embedding matrix are obtained through forward propagation. The node structural importance parameter of each node is obtained by taking the modulus of the embedding vector of each node in the node embedding matrix and normalizing it; the structural correlation parameter between nodes is obtained by calculating the cosine similarity between the embedding vectors of each edge in the edge embedding matrix; the calculated node structural importance parameter and the structural correlation parameter between nodes are stored in the embedding vector database.

3. The method according to claim 1, characterized in that, The path search structural constraint rules are determined based on the structural constraint parameters; the path search structural constraint rules are obtained in any of the following ways: The first method is to compare the node structure importance parameters of all neighboring nodes of the current node with a preset importance threshold, select neighboring nodes whose node structure importance parameters are higher than the importance threshold, and form a whitelist set corresponding to that node; in each step of the path search expansion, the next node can only be selected from the whitelist set corresponding to the current node. The second approach is to sort all neighboring nodes of the current node in descending order according to their importance parameters, and select the top K neighboring nodes to form a set of priority candidate nodes for the current node. In each step of the path search, node selection is performed only within the set of priority candidate nodes. Here, K is a preset positive integer. The third method involves setting a connection threshold. For any two nodes, if the structural correlation parameter between them is greater than or equal to the connection threshold, then the node pair is deemed to be allowed to connect directly; otherwise, direct jumps are prohibited. Based on the determination results of all node pairs, a binary constraint matrix is ​​constructed between the node pairs, where a matrix element of 1 indicates that a connection is allowed and a 0 indicates that a jump is prohibited. During the path search process, the binary constraint matrix is ​​used to query whether direct jumps between the current node and candidate nodes are allowed, thus performing structural filtering. The fourth method is to input the structural correlation parameters between nodes into a preset non-negative monotonically decreasing penalty mapping function to obtain the jump penalty coefficients and store them as a path jump coefficient table; during the path search process, priority is given to selecting the jump direction of nodes with high structural correlation.

4. The method according to claim 1, characterized in that, Based on the path search structure constraint rules and structure constraint parameters, a biased random walk is performed to generate an initial population composed of candidate paths, including: Candidate paths are encoded as sequences of alternating nodes and transportation modes; In each step of the biased random walk, the structural potential index of the generated path sequence from the starting point to the current node is calculated based on the structural constraint parameters, and the selection of the next node and transportation mode is guided by the structural potential index. The structural potential index is obtained through at least one of the following methods: The first method is to calculate the cumulative value of the node structure importance parameters of all nodes in the generated path sequence as an indicator of structural potential. The second approach involves calculating the average vector of all node embedding matrices in the generated path sequence as the sequence average embedding, and then calculating the cosine similarity between the sequence average embedding and the predefined ideal hub reference embedding to obtain the node embedding similarity aggregate value, which serves as a structural potential indicator. The third approach involves dividing all nodes into several preset levels based on the importance parameters of the node structure, evaluating whether the level changes of edges in the generated path sequence conform to the predefined level flow pattern, calculating the proportion of edges that conform to the level flow pattern in the generated path sequence, and obtaining the node structure level matching degree as a structural potential indicator.

5. The method according to claim 1, characterized in that, Calculate the objective function value of each path in the current population on the preset scheduling objective, and perform non-dominated sorting. Obtain the Pareto ranking of each path through non-dominated sorting and crowding calculation, including: A preset scheduling objective is defined as a combination of at least one of transportation cost, transportation time, and historical reliability. Perform non-dominated sorting on candidate paths in the current population, including: For any two candidate paths in the initial population, if the objective function value of the first candidate path is not inferior to that of the second candidate path on all scheduling objectives and is superior to that of the second candidate path on at least one scheduling objective, then the first candidate path is determined to dominate the second candidate path; record the set of paths dominated by each candidate path and the number of paths dominated by the candidate path. Candidate paths with a dominance of 0 are assigned to the first frontier level; candidate paths whose dominance is reduced to 0 after decreasing by 1 within the path set dominated by each candidate path in the first frontier level are assigned to the second frontier level, and this process is repeated step by step until all candidate paths in the population are assigned to the corresponding frontier level. The congestion calculation for candidate paths within the same frontier level includes the following steps: for each scheduling target, sort all candidate paths within that level in ascending order according to their objective function values; set the congestion contribution of candidate paths at both ends of the sorted sequence to infinity; for each candidate path in the middle of the sorted sequence, normalize the absolute value of the difference between its objective function value and that of its two adjacent candidate paths on that scheduling target, and then add it to the congestion of that candidate path. Iterate through all scheduling targets to complete the accumulation and obtain the final congestion of each candidate path; The smaller the frontier level, the higher the Pareto ranking; within the same frontier level, the greater the crowding, the higher the Pareto ranking.

6. The method according to claim 1, characterized in that, Using the mutation probability obtained from a nonlinear convergence factor mapping, node mutation or transportation mode replacement is performed on the replaced path, including: Based on the ratio of the current iteration number to the total number of iterations, the nonlinear convergence factor of the current iteration is obtained by exponentially decreasing the ratio. This nonlinear convergence factor is then mapped to the mutation probability of the current iteration number using a preset monotonically decreasing mapping function. The nonlinear convergence factor is jointly controlled by the initial coefficient and the decay rate control coefficient. Traverse each intermediate node in the candidate path after replacement, and determine whether to perform node mutation based on the mutation probability; for nodes determined to perform mutation, select replacement nodes according to the path search structure constraint rules. Traverse each directed edge in the candidate path after replacement, and determine whether to perform transportation mode replacement based on the mutation probability; for directed edges determined to be transportation mode replacement, select alternative directed edges for replacement according to the path search structure constraint rules.

7. The method according to claim 6, characterized in that, The nonlinear convergence factor is mapped to the mutation probability of the current iteration number through a preset monotonically decreasing mapping function, where: When the nonlinear convergence factor is greater than a preset threshold, the value of the mutation probability is increased to drive the candidate path to explore new regions; When the nonlinear convergence factor is less than or equal to a preset threshold, the value of the mutation probability is reduced to drive the candidate path to conduct a small-range search near the optimal region.

8. A vehicle dynamic scheduling system based on graph embedding gray wolf constraints, characterized in that, include: The dispatch request receiving module is used to receive vehicle dispatch requests that include origin, destination, cargo information and time requirements; The structural constraint module is used to read structural constraint parameters related to vehicle scheduling requests from the embedded vector database, which characterize the structural importance of nodes in the transportation network and the correlation between nodes. The gray wolf search optimization module is used to perform biased random walks based on the structural constraint parameters and the determined path search structural constraint rules, generate an initial population composed of candidate paths as the current population, and iterate the following steps until convergence: Calculate the objective function value of each path in the current population on the preset scheduling target, and perform non-dominated sorting. Calculate the Pareto ranking of each path through non-dominated sorting and crowding degree, and calculate the cosine similarity between the average node embedding of each path and the predefined reference embedding as the structural potential score. Select α, β, and δ leader wolves based on the weighted sum of Pareto ranking and structural potential score. Based on the preset probability corresponding to each leader wolf, cross segments are randomly selected from the paths of each leader wolf to replace the corresponding segments in the other paths. The mutation probability obtained by nonlinear convergence factor mapping is used to perform node mutation or transportation mode replacement on the replaced paths to generate a new generation of population as the current population for the next iteration. After the population iteration converges, a set of Pareto optimal path solutions is output, and one of them is selected as the current path and executed. The local re-optimization module is used to detect dynamic events and determine the local sub-paths affected by them in the current path; analyze the scope of the event's influence and delineate a local subgraph centered on the local sub-paths; reuse the path search structure constraint rules and the structure constraint parameters to generate a local initial population in the local subgraph, generate a local Pareto optimal solution set through a preset finite number of iterations, and select alternative sub-paths from it to replace the local sub-paths, so that the updated scheduling path can continue to be executed.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is enabled to perform the method as described in any one of claims 1-7.

Images