Multimodal transport cargo arrival time prediction method based on graph neural network

Abstract

The invention relates to the technical field of logistics management, and discloses a multimodal transport cargo arrival time prediction method based on a graph neural network, and the method comprises the steps: receiving an arrival time prediction request for a target cargo; planning a candidate transport path according to the arrival time prediction request, and dynamically extracting a sub-graph from the global multimodal transport network; endowing each node and each directed edge with a feature vector, endowing each sub-graph with a global feature vector, and inputting the sub-graphs after feature initialization into a graph neural network model; the method comprises the following steps of: generating an initial state vector for each node through a graph neural network model; iteratively updating the state vector of the node through a multi-layer message transmission mechanism, and simulating the cascade propagation effect of the transportation event on the path network; and based on the final state vector of the destination node after iteration updating, regression is carried out to obtain a probability distribution parameter of the arrival time. According to the method, cascade propagation of transportation delay in the network can be simulated, probabilistic prediction containing uncertainty measurement is output, and the prediction precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of logistics management, and in particular to a multi-modal freight arrival time prediction method based on a graph neural network. BACKGROUND

[0002] In the multi-modal freight transportation scenario, accurately predicting the arrival time (ETA) of freight from the starting point to the destination is crucial for improving supply chain transparency, optimizing inventory management, and enhancing customer satisfaction. However, existing techniques typically predict each transportation leg (such as highway, railway, and sea transportation) independently and then simply add them up. The flowchart is shown in Figure 1 Specifically, first, the highway leg prediction time T1, the sea transportation leg prediction time T2, and the railway leg prediction time T3 are independently predicted, and then T1, T2, and T3 are simply added up to obtain the final ETA prediction result.

[0003] However, this method of existing techniques completely ignores the inherent cascading effect in the multi-modal transportation network. For example, delays in the previous leg of the journey will inevitably cause a chain reaction in the operation plan of the subsequent nodes (such as port loading and unloading, railway connection), and the independent prediction model cannot capture the complexity of this dynamic propagation, resulting in a low accuracy of the ETA prediction result. SUMMARY

[0004] The purpose of the present application is to dynamically convert the multi-modal transportation network in the physical world into a computable graph structure, and use the unique message passing mechanism of the graph neural network to simulate the cascading propagation of transportation delays in the network, and finally output a probabilistic prediction containing uncertainty metrics.

[0005] In a first aspect, an embodiment of the present application provides a multi-modal freight arrival time prediction method based on a graph neural network, comprising: receiving an arrival time prediction request for a target freight, the arrival time prediction request containing the origin, the destination, and the freight characteristics; planning a candidate transportation path according to the arrival time prediction request, and dynamically extracting a subgraph from a pre-defined global multi-modal transportation network based on the candidate transportation path, the subgraph being composed of all nodes and directed edges connecting these nodes that the candidate transportation path passes through, the nodes representing logistics hubs, and the directed edges representing feasible transportation routes between nodes; assigning a node feature vector to each node in the subgraph, assigning an edge feature vector to each directed edge, and assigning a global feature vector to the subgraph, to obtain a subgraph after feature initialization; the node feature vector of a node comprises static features representing inherent attributes of a logistics hub corresponding to the node and dynamic features representing a real-time operation state of the logistics hub; the edge feature vector of a directed edge comprises static features representing inherent attributes of a transportation line corresponding to the directed edge and dynamic features representing a real-time traffic state of the transportation line; and the global feature vector comprises macro context information related to the prediction task; inputting the subgraph after feature initialization into a pre-trained graph neural network model; and performing the following operations by using the graph neural network model: generating an initial state vector for each node based on the node feature vector; updating the state vector of each node through a multi-layer message passing mechanism, wherein each layer performs information transmission by combining the edge feature vector and the state vectors of connected nodes, so that the state of each node can fuse the information of its multi-hop neighbor nodes to simulate the cascading propagation effect of transportation events on the path network; based on the final state vector of the destination node after iteration, regressing to obtain the probability distribution parameters of the arrival time.

[0006] Optionally, the step of updating the state vector of each node through the multi-layer message passing mechanism comprises the following steps performed in each layer: for each directed edge connecting a source node and a target node, generating a transmission message by using a message function, combining the edge feature vector of the directed edge, the current state vector of the source node, and the current state vector of the target node; for each target node, converging the transmission messages corresponding to all incoming edges of the target node by using an aggregation function to obtain an aggregated message of the target node; wherein all incoming edges are directed edges pointing to the target node; for each target node, updating the state vector of the target node by using an update function, combining the current state vector of the target node and the aggregated message.

[0007] Optionally, the step of regressing to obtain the probability distribution parameters of the arrival time based on the final state vector of the destination node after iteration comprises: performing a graph readout operation on the final state vector of the destination node; inputting the readout vector into a prediction head, wherein the prediction head is implemented by a fully connected neural network; outputting parameters for defining a probability distribution model by using the prediction head.

[0008] Optionally, the probability distribution model is Gaussian distribution, and the parameters output by the prediction head include an expected mean μ and a variance σ 2 of the arrival time.

[0009] Optionally, the construction process of the predefined global multimodal network includes: acquiring multi-source heterogeneous data, the multi-source heterogeneous data including logistics hub attribute data, transportation line data, and macro-environment data; processing the acquired multi-source heterogeneous data to eliminate data heterogeneity, wherein the processing methods include data cleaning, standardization, and alignment; generating a graph data structure G=(V,E,U) based on the processed data, wherein V is a node set, each node represents a logistics hub and has a corresponding node feature vector; E is a directed edge set, each directed edge represents a feasible transportation line between nodes and has a corresponding edge feature vector; and U is a global feature vector representing the macro state of the entire network.

[0010] Optionally, the static features representing the inherent attributes of the logistics hub include at least one of the following: hub type, design throughput capacity, geographic location coordinates, and fixed processing time length benchmark; and the dynamic features representing the real-time operation state of the logistics hub include at least one of the following: real-time queue length, current equipment availability, and weather forecast in a specific time window in the future. The static features of the inherent attributes of the transportation line include at least one of the following: fixed length of the line, default transportation mode, historical average speed, and standard transportation rate; and the dynamic features of the real-time traffic state of the transportation line include at least one of the following: current average driving speed, real-time traffic congestion index, and current period transportation discount. The global feature vector includes at least one of the following: international fuel price index, whether it is a national statutory holiday, and macroeconomic fluctuation index.

[0011] Optionally, the training process of the graph neural network model includes: acquiring historical transportation order data to construct a training set, wherein each training sample includes: a transportation path sequence, real arrival and departure times at each node on the path, and node feature vectors and edge feature vectors corresponding to the historical time; training the graph neural network model end-to-end using the training set, and achieving the training target by minimizing the negative log-likelihood loss function, wherein the training target is to minimize the difference between the probability distribution predicted by the model and the real arrival time distribution; updating the parameters of the graph neural network model using a backpropagation algorithm and an optimizer iteratively until the graph neural network model converges.

[0012] Optionally, the method further includes a graph neural network model updating step: adding real data of completed transportation orders as new samples to the training set; incrementally training the deployed graph neural network model using the new samples to realize online learning or periodic updating of the graph neural network model.

[0013] In a second aspect, an embodiment of the present application provides a multi-modal transport goods arrival time prediction system based on a graph neural network, comprising: a prediction request receiving module configured to receive an arrival time prediction request for a target good, the arrival time prediction request containing a departure location, a destination, and a good characteristic; a subgraph extraction module configured to plan a candidate transportation path according to the arrival time prediction request, and dynamically extract a subgraph from a pre-defined global multi-modal transport network based on the candidate transportation path, the subgraph being composed of all nodes and directed edges connecting the nodes passed through by the candidate transportation path, the nodes representing logistics hubs and the directed edges representing feasible transportation routes between the nodes; a subgraph feature initialization module configured to assign a node feature vector to each node in the subgraph, assign an edge feature vector to each directed edge, and assign a global feature vector to the subgraph, to obtain the subgraph after feature initialization; the node feature vector of a node including static features representing inherent attributes of a logistics hub corresponding to the node and dynamic features representing real-time operation status of the logistics hub; the edge feature vector of a directed edge including static features representing inherent attributes of a transportation route of the directed edge and dynamic features representing real-time traffic status of the transportation route; and the global feature vector including macro context information related to the prediction task; a graph neural network model prediction module configured to input the subgraph after feature initialization into a pre-trained graph neural network model; and perform the following operations by using the graph neural network model: generating an initial state vector for each node based on the node feature vector; updating the state vector of each node through a multi-layer message passing mechanism, wherein each layer performs information passing by combining the edge feature vector and the state vector of the connected node, so that the state of each node can fuse the information of its multi-hop neighbor nodes to simulate the cascading propagation effect of transportation events on the path network; based on the final state vector of the destination node after iteration, regressing to obtain probability distribution parameters of the arrival time.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, comprising: one or more processors; a memory configured 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 of the first aspect.

[0015] The technical scheme provided by the embodiment of the application comprises the following steps: receiving an arrival time prediction request for target goods; planning a candidate transportation path according to the arrival time prediction request, and dynamically extracting a subgraph from a predefined global multimodal transport network based on the candidate transportation path; assigning a node feature vector to each node in the subgraph, assigning an edge feature vector to each directed edge, and assigning a global feature vector to the subgraph to obtain a subgraph after feature initialization; inputting the subgraph after feature initialization into a pre-trained graph neural network model; performing the following steps by using the graph neural network model: generating an initial state vector for each node based on the node feature vector; iteratively updating the state vector of each node by using a multi-layer message passing mechanism, wherein each layer performs information transmission by combining the edge feature vector and the state vector of the connected node, so that the state of each node can fuse the information of the multi-hop neighbor nodes to simulate the cascading propagation effect of the transportation event on the path network; and based on the final state vector of the destination node after iteration, regressing to obtain the probability distribution parameters of the arrival time.

[0016] Compared with the prior art, the application abstracts the multimodal transport network into a dynamic graph structure and uses the message passing mechanism of the graph neural network, thereby systematically solving the fundamental defect that the traditional segmented prediction model cannot capture the cascading effect, and realizing a qualitative leap in prediction accuracy from a local and one-sided perspective to a systematic and global perspective. At the same time, the probabilistic distribution parameters output by the model provide a standard for measuring uncertainty for users, so that the prediction result is upgraded from a single reference number to a tool supporting risk assessment and emergency planning. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 For the prior art, each transportation segment is usually predicted independently, and then the arrival time is obtained by simply adding up the results; Figure 2 For the schematic diagram of the overall architecture and data flow of the overall technical scheme of the application; Figure 3 For the flowchart of the multimodal transport goods arrival time prediction method based on the graph neural network provided by the embodiment of the application; Figure 4 For the structural schematic diagram of the multimodal transport goods arrival time prediction system based on the graph neural network provided by the embodiment of the application; Figure 5 For the structural schematic diagram of the electronic device provided by the embodiment of the application. DETAILED DESCRIPTION

[0018] The application will be described in detail below through examples. The examples are implemented on the basis of the technical solutions of the application, and detailed implementation modes and specific operation processes are given, but the protection scope of the application is not limited to the examples below.

[0019] In the multi-modal transport scenario, accurately predicting the arrival time (ETA) of goods from the starting point to the destination is crucial for improving supply chain transparency, optimizing inventory management, and enhancing customer satisfaction. However, existing technologies usually predict each transportation leg (such as road, rail, and sea) independently and then simply add them up. The flowchart is shown in Figure 1 Specifically, first, the road segment prediction time T1, the sea segment prediction time T2, and the rail segment prediction time T3 are independently predicted, and then T1, T2, and T3 are simply added up to obtain the final ETA prediction result.

[0020] However, this method of the existing technology completely ignores the inherent cascading effect in the multi-modal transport network. For example, the delay of the previous leg will inevitably cause a chain reaction in the operation plan of the subsequent nodes (such as port loading and unloading, and rail connection), and the independent prediction model cannot capture the complexity of this dynamic propagation, resulting in low accuracy of the ETA prediction result.

[0021] Moreover, the above method of the existing technology has the following technical problems: The model lacks sufficient feature extraction capability. Specifically, traditional machine learning models have difficulty in effectively processing the topological structure of the transport network. They treat each hub and line as an isolated feature, and cannot deeply mine the complex spatial dependency and mutual influence between nodes; Moreover, the prediction result does not provide uncertainty measurement. Specifically, most systems only output a single determined ETA value, and cannot provide a probability distribution or confidence interval. In actual logistics, decision-makers need to know "under the probability of 95%, when will the goods arrive", rather than a single, usually inaccurate time point; In addition, data utilization is insufficient. Specifically, real-time dynamic data (such as the current degree of port congestion and weather conditions) is not deeply fused with network structure data for prediction.

[0022] Therefore, there is an urgent need in the field for an ETA prediction scheme that can systematically model the entire transport network, accurately quantify the cascading effect, and provide probabilistic prediction results.

[0023] The present application aims to solve the "blind spot" and "short board" existing in the current multimodal transport ETA prediction technology. The traditional method adopts the mode of segmented independent prediction and simple summation, and the fundamental defect is that it ignores the dynamic correlation and cascade effect between nodes in the transport network, cannot simulate the chain impact of the delay of the previous section on the operation of the subsequent section, and leads to a large deviation between the prediction result and the actual situation, and the reference value is limited.

[0024] Therefore, the present application proposes to abstract the entire multimodal transport network as a graph structure, and use the message passing mechanism unique to the graph neural network (GNN) to deeply learn and simulate the propagation path and influence strength of dynamic events such as delay between hubs and lines, so that the model has a "global view" and makes end-to-end comprehensive reasoning from the system level, rather than local isolated judgment.

[0025] In addition, the goal of the present application is not only to realize a more accurate point estimate, but also to output a probabilistic prediction distribution (such as a confidence interval). This provides the user with a crucial uncertainty measure, enabling risk assessment and emergency planning, thereby upgrading the ETA prediction from a reference number to a reliable decision support tool, and comprehensively improving the resilience, transparency and operational efficiency of the supply chain.

[0026] The multimodal transport ETA prediction system based on the graph neural network provided by the present application is characterized by converting the complex multimodal transport network into a graph structure and using the characteristics of GNN to model the complex spatio-temporal dependence relationship therein. In order to clearly describe the scheme, first of all, the combination of Figure 2 The overall technical scheme of the embodiment of the present application is described in detail. The following describes in detail from two aspects of system composition and implementation steps.

[0027] Figure 2 The overall architecture and data flow diagram of the overall technical scheme of the present application. Figure 2 The four core modules of the system described in the present application and the data flow relationship therebetween are clearly shown, which constitute a complete closed loop from data perception, processing, intelligent reasoning to self-evolution. The workflow and core functions are described as follows: From Figure 2 As can be seen, the starting point is external multi-source data. Specifically, the system can continuously access massive, heterogeneous real-time and historical data from the physical world. These data sources include: port, airport, station operation management system (API), Internet of Things sensor (IoT), GPS and AIS signals of ships / vehicles, third-party weather services, traffic monitoring platforms, and public economic indicators, etc.

[0028] From Figure 2As can be seen, the first core module of the system is a multi-source data real-time fusion module. The multi-source data real-time fusion module, as the data hub of the system, continuously extracts data from multi-source heterogeneous data sources through various interfaces and technologies. Its data processing flow is as follows: the module performs a continuous ETL (Extraction-Transformation-Loading) process. It cleans, standardizes (unifies units and timestamps) and aligns (maps data from different sources to corresponding nodes v_i or edges e_ij) the obtained raw data, and finally outputs the regular feature vectors for consumption by the GNN model.

[0029] The second core module of the system is a transportation network graph construction module. This module is responsible for abstracting the physical transportation network into a computer-processable graph data structure G= (V, E, U). Its input can be: receiving standardized data from upstream and user query requests from the application layer. The processing process is: according to the origin, destination and cargo characteristics in the query request, planning the path and dynamically extracting a relevant subgraph from a predefined global multimodal transport network.

[0030] The following will elaborate on each parameter in the graph data structure G= (V, E, U).

[0031] Among them, V represents the node set. Each node v_i∈V represents a logistics hub (such as a port, a railway station, an airport, a warehouse). Each node has a high-dimensional feature vector x_i, which includes: static attributes and dynamic states. Among them, static attributes can include: hub type, design throughput, geographic location coordinates and fixed processing time benchmark, etc.; dynamic state includes real-time queue length, current equipment availability and future weather forecast in a specific time window.

[0032] E represents the edge set. Each directed edge e_ij∈E represents a feasible transportation route from node v_i to v_j. Each edge also has a feature vector a_ij, which includes: static attributes and dynamic states, where static attributes can include: fixed length of the route, default transportation mode, historical average speed, standard transportation rate. Dynamic state can include: current average speed, real-time traffic congestion index and current period transportation discount.

[0033] U represents the global context, which is a vector representing the macro state of the entire network, which can include: international oil price index, whether it is a national statutory holiday, macroeconomic fluctuation index, etc. This vector is irrelevant to specific nodes and edges, but is related to all predictions.

[0034] The third core module of the system is the GNN prediction model core module. This core module embodies the algorithm core of the invention and is a multi-level spatio-temporal graph neural network.

[0035] The input of the GNN prediction model is a featureized subgraph structure, which contains nodes, edges and their rich features related to the current prediction task. Moreover, the GNN prediction model defines the message passing mechanism, message function, aggregation function and update function.

[0036] Among them, the message passing (neighbor aggregation) mechanism is the core operation of GNN. At each layer, each node will receive messages from its "neighbors".

[0037] Message function: define how to generate a message m_ij from the combination of adjacent node v_j and edge e_ij features.

[0038] Aggregation function: node v_i aggregates messages from all neighbors, usually using attention mechanism (GAT) or sum / mean pooling, to get a comprehensive neighborhood information M_i. This process simulates the cascading effect: the status update of a port node depends on the aggregation of delay messages from upstream ships and congestion messages from downstream land routes.

[0039] Update function: the node updates its own state representation h_i by combining its own last layer features and aggregated neighborhood information M_i. By stacking multiple GNN layers, the final representation of each node can capture network information beyond multiple hops, with a "global view".

[0040] Finally, the updated representation vector of the target node is input into a fully connected neural network. This fully connected neural network does not directly output an ETA value, but outputs the parameters of a probability distribution. For example, if ETA follows a Gaussian distribution, two values are output: mean μ (expected arrival time) and variance σ² (uncertainty level). This design makes the prediction result contain rich risk information.

[0041] From Figure 2 It can also be seen that the core module four of the system is the model training and updating module, which is responsible for the initial training and continuous optimization of the model.

[0042] The training data of the model training and updating module is: use massive historical transportation order data, each order contains its complete path and real arrival and departure timestamps at each node.

[0043] The loss function of the model training and updating module is: negative log likelihood (NLL) loss. The training goal is to make the probability distribution predicted by the model fit the distribution of the real ETA data to the maximum extent.

[0044] The update mechanism of the model training and updating module: the system supports online learning or regular retraining, using new transportation data as training samples, so that the model can adapt to the changing logistics environment and realize self-evolution.

[0045] Compared with the prior art, the technical scheme provided by the application at least has the following technical effects: 1. Precision is significantly improved. By explicitly modeling the transport network through a graph structure and accurately capturing the cascading effect through a message passing mechanism, the paradigm of previous isolated prediction is changed, and the prediction precision is much higher than that of traditional methods.

[0046] 2. The prediction result is more valuable for decision-making. Instead of single-point prediction, probability prediction is provided, which provides a standard for measuring prediction uncertainty for users and supports more robust risk assessment and supply chain decision-making.

[0047] 3. Strong generalization ability. Graph neural networks can handle unseen node and path combinations as long as they can be embedded into learned network representations, which makes the system have good adaptability to network changes and new routes.

[0048] 4. Global optimal perspective. End-to-end prediction is performed from the perspective of the entire network, rather than simply segmenting and superimposing, fully considering all possible influencing factors and their propagation paths.

[0049] After the overall technical scheme of the application is described in detail, a multi-modal freight arrival time prediction method based on a graph neural network provided by an embodiment of the application will be described in detail. The core of the method is to dynamically convert the multi-modal freight network in the physical world into a computable graph structure, and use the unique message passing mechanism of the graph neural network to simulate the cascading propagation of transport delays in the network, and finally output a probabilistic prediction containing uncertainty measurement.

[0050] As shown in Figure 3 A multi-modal freight arrival time prediction method based on a graph neural network provided by an embodiment of the application can include the following steps: S310, receiving an arrival time prediction request for a target freight.

[0051] The arrival time prediction request contains the origin, destination and freight characteristics.

[0052] This step is the interactive interface of the system and the user or the upstream business system (such as the transport management system TMS).

[0053] The request is not a simple start-end point query, but contains the minimum information set necessary for intelligent prediction.

[0054] The origin and destination define the spatial range of the prediction task.

[0055] Cargo characteristics are the key to achieve personalized prediction. Different cargos (e.g. general cargo, refrigerated cargo, dangerous goods, high-value cargo) have significant differences in handling time and priority on the same hub and line due to the differences in loading and unloading, storage, transportation mode and regulatory requirements. For example, refrigerated cargo may need additional plug-in and inspection time at the port, and dangerous goods transportation may have speed limits on certain routes. Therefore, cargo characteristics are important inputs for path planning and time prediction.

[0056] S320, planning a candidate transportation path according to the arrival time prediction request, and dynamically extracting a subgraph from a predefined global multimodal transport network based on the candidate transportation path.

[0057] The subgraph is composed of all nodes passed through by the candidate transportation path and directed edges connecting the nodes, wherein the nodes represent logistics hubs and the directed edges represent feasible transportation routes between the nodes.

[0058] Specifically, this step is the core bridge connecting the specific business request and the abstract graph model, realizing the conversion from the "physical path" to the "computational subgraph". The system plans one or several (candidate) transportation paths based on the cost, time efficiency or specific strategy according to the origin, destination and cargo characteristics. This reflects the flexibility of the system, which can evaluate multiple feasible schemes at the same time.

[0059] Moreover, unlike using a fixed and large global network graph, the embodiment "tailors" a computational subgraph for each prediction request. This greatly reduces the model calculation complexity and ensures that the calculation resources are concentrated in the most relevant network local part for the current task. The subgraph accurately contains all nodes (logistics hubs) passed through by the candidate path and directed edges (transportation routes) connecting the nodes. This definition ensures that the model can perform information reasoning along the real transportation path, rather than performing invalid calculations on irrelevant network parts.

[0060] S330, assigning a node feature vector to each node in the subgraph, assigning an edge feature vector to each directed edge, and assigning a global feature vector to the subgraph, to obtain the subgraph after feature initialization.

[0061] The node feature vector of a node includes static features representing inherent properties of the logistics hub corresponding to the node and dynamic features representing the real-time operation state of the logistics hub. The edge feature vector of a directed edge includes static features representing inherent properties of the transportation route of the directed edge and dynamic features representing the real-time traffic state of the transportation route. The global feature vector includes macro context information related to the prediction task.

[0062] Specifically, this step can convert the original, featureless network skeleton into an input rich in spatiotemporal attributes that can be understood by GNN.

[0063] For the node feature vector, it contains both static and dynamic features. The static features describe the inherent properties of the logistics hub, such as type (port / railway station), designed throughput capacity (maximum handling capacity), geographic location, and fixed handling duration benchmark. These are the "identity information" of the node, which determines its basic processing capacity. The dynamic features describe the real-time state of the logistics hub, such as queue length, equipment availability, and future weather forecast. This is the "real-time health status" of the node, which directly affects its current processing efficiency. The combination of static and dynamic features enables the model to understand both the long-term capacity and the instantaneous state of the node.

[0064] For the edge feature vector, it also includes both static and dynamic features. The static features describe the inherent properties of the transportation line, such as fixed length, default transportation mode (road / rail / sea), historical average speed, and standard transportation cost. These define the basic passing properties of the line. The dynamic features describe the real-time state of the transportation line, such as current average driving speed (from GPS), real-time traffic congestion index, and current transportation cost discount. These reflect the current passing efficiency and cost of the line. This design enables the model to perceive the real-time impact of weather, traffic incidents, and other factors on the passing capacity of the line.

[0065] For the global feature vector, this vector captures macro factors related to the entire network and the prediction task, but cannot be attributed to any specific node or edge. For example: international fuel prices, which affect the cost of all transportation modes and may indirectly affect the scheduling strategy of carriers. National statutory holidays change the nationwide operation mode and affect the overall logistics efficiency. The macroeconomic fluctuation index may affect the overall freight demand and thus the network congestion level. By introducing global features, the model is given the ability to understand the impact of macro-environmental shocks.

[0066] S340, input the initialized subgraph into the pre-trained graph neural network model. Through the graph neural network model, the following operations are performed: Based on the node feature vector, an initial state vector is generated for each node.

[0067] Through a multi-layer message passing mechanism, the state vector of the node is iteratively updated. Each layer combines the edge feature vector with the state vector of the connected nodes to pass information, so that the state of each node can integrate the information of its multi-hop neighbor nodes to simulate the cascading propagation effect of transportation events on the path network.

[0068] Based on the final state vector of the destination node after iteration, the probability distribution parameters of the arrival time are regressed.

[0069] Specifically, this step is the intelligent core of the entire method, which receives the initialized subgraph and outputs the final prediction result through a series of calculations. This step mainly includes the following three sub-steps: The first sub-step is node embedding representation. The model first maps the feature vector of each node to a high-dimensional, numerical initial state vector through an encoding layer (such as linear transformation or MLP). This process unifies heterogeneous original features into a hidden space that is convenient for the model to process.

[0070] The second sub-step is cascade effect simulation, which is one of the core innovations of the present application. It is realized through a multi-layer message passing mechanism. Each layer performs a cycle of "message generation → message aggregation → state update".

[0071] Message generation: For an edge from node A to node B, the model generates a "message" from A to B by combining the state of A, the state of B, and the features of the edge itself. This message can be understood as "A's current state and the line condition between them will have what impact on B".

[0072] Message aggregation: Node B will receive messages from all incoming edges pointing to it (i.e., all its upstream neighbors). The model combines these messages through an aggregation function (such as attention weighted sum) to obtain a comprehensive "neighbor influence package".

[0073] State update: Node B combines its current state with this "neighbor influence package" and updates its state through an update function (such as GRU). This new state now contains information from its direct neighbors.

[0074] After K layers of such message passing, the final state vector of each node encodes the information of all its neighbors within K hops (i.e., through K edges). This means that the destination's state has perceived the propagation of the initial delay from the origin, as well as the congestion conditions of all the hubs and lines in the path. This perfectly simulates the real-world "one delay, everywhere affected" cascade effect, which is the key to overcoming the limitations of traditional segment prediction.

[0075] Moreover, the model does not output a single, deterministic ETA value, but outputs a parameter of a probability distribution (for example, assuming that ETA follows a Gaussian distribution, it outputs the mean μ and the variance σ²).

[0076] Mean μ: provides the expected arrival time, i.e., the most likely ETA.

[0077] Variance σ²: quantifies the uncertainty of the prediction. Large variance indicates that this transportation is greatly affected by uncertainty factors, with a wide range of arrival times; small variance indicates that the prediction is confident, with a more reliable arrival time.

[0078] This makes the prediction result from a reference number, upgrade to a risk assessment tool. Decision makers can make emergency plans accordingly (such as "I have 95% confidence that the goods will arrive within a time window of μ±2σ"), greatly improving the resilience and decision-making level of supply chain management.

[0079] The technical scheme provided by the embodiment of the application comprises the following steps: receiving an arrival time prediction request for target goods; planning a candidate transportation path according to the arrival time prediction request, and dynamically extracting a subgraph from a predefined global multimodal transport network based on the candidate transportation path; assigning a node feature vector to each node in the subgraph, assigning an edge feature vector to each directed edge, and assigning a global feature vector to the subgraph to obtain a subgraph after feature initialization; inputting the subgraph after feature initialization into a pre-trained graph neural network model; and performing the following steps by using the graph neural network model: generating an initial state vector for each node based on the node feature vector; iteratively updating the state vector of each node by using a multi-layer message passing mechanism, wherein each layer performs information transmission by combining the edge feature vector and the state vector of the connected node, so that the state of each node can fuse the information of its multi-hop neighbor nodes to simulate the cascading propagation effect of the transportation event on the path network; and based on the final state vector of the destination node after iteration, regressing to obtain the probability distribution parameters of the arrival time.

[0080] Compared with the prior art, the application abstracts the multimodal transport network into a dynamic graph structure and uses the message passing mechanism of the graph neural network to systematically solve the fundamental defect that the traditional segmented prediction model cannot capture the cascading effect, realizing a qualitative leap in prediction accuracy from a local and one-sided to a systematic and global. At the same time, the probabilistic distribution parameters output by the model provide a standard for users to measure uncertainty, so that the prediction result is upgraded from a single reference number to a tool supporting risk assessment and emergency planning.

[0081] In Figure 3 Based on the embodiment shown in the figure, as an implementation manner of the embodiment of the application, the state vector of each node is iteratively updated by using a multi-layer message passing mechanism, which comprises the following steps a1 to a3 performed in each layer: Step a1, for each directed edge connecting a source node and a target node, a message function is used to combine the edge feature vector of the directed edge, the current state vector of the source node and the current state vector of the target node to generate a transmission message.

[0082] Specifically, this step aims to quantitatively model the potential "influence" between any two directly connected nodes (through a directed edge) in the network. The input of the message function is a complete three-tuple information, which includes: 1. Current state vector of source node: represents the current operational status of the upstream hub (e.g., port A) (e.g., whether it is congested).

[0083] 2. Edge feature vector of directed edge: represents the real-time status of the transportation link connecting two hubs (e.g., the highway from A to B) (e.g., current driving speed, congestion index).

[0084] 3. Current state vector of target node: represents the current receiving and processing capacity of the downstream hub (e.g., warehouse B).

[0085] The message function M(·) (usually a learnable neural network, such as a small MLP) performs a comprehensive calculation on the three vectors to generate a passing message m_ij. This message is a high-dimensional vector that essentially encodes the numerical value of "the spatiotemporal impact of the transportation segment from node i to node j via edge e_ij at the current time." For example, it may encode the complex information "due to the congestion of port A and the highway limit, the expected arrival time at warehouse B will be delayed by 2 hours."

[0086] Step a2, for each target node, aggregate the passing messages corresponding to all its incoming edges to obtain the aggregated message of the target node. All incoming edges are all directed edges pointing to the target node.

[0087] Specifically, in a real network, a hub is usually simultaneously affected by multiple upstream hubs. This step aims to simulate how a node integrates all complex information from its upstream that is concurrent, possibly contradictory or reinforcing.

[0088] The input of this step is the message set {m_1j, m_2j,..., m_kj} generated by all incoming edges (i.e., all directed edges pointing to it) of the target node v_j. The aggregation function AGG(·) is used to combine these messages into a single, comprehensive aggregated message M_j. A simple sum or average pooling is one way of aggregation. However, a better solution is to introduce an attention mechanism (Graph Attention Network, GAT). At this time, the aggregation function will automatically assign different weights to messages m_ij from different neighbors. The size of the weight represents the importance of the influence of the neighbor node v_i on the current target node v_j. For example, a major upstream port that is experiencing severe delays will have a higher attention weight than a small transit station with slight congestion. This weighted aggregation allows the model to focus on the most critical sources of influence, greatly improving the accuracy of the simulation.

[0089] Step a3: For each target node, update its state vector by combining its current state vector with the aggregated message using the update function.

[0090] Specifically, this step simulates how a node adjusts its state after receiving comprehensive external influences. A node's state vector is a dynamic representation of its "identity" and "situation" in the entire network, and updating this state is to make the node "aware" of its new situation in the network.

[0091] The input to this step is the target node v_j's current state vector h_j (its self-awareness before the update); and the aggregated message M_j obtained from step a2 (the combined influence from the external environment). The update function U(·) (usually a recurrent neural network unit such as GRU, or a linear transformation plus a nonlinear activation function) receives these two inputs and computes the node's new state vector h_j'. The new state vector h_j' is no longer its isolated state, but rather a fusion of its own historical state and the real-time influences transmitted from all upstream neighbors through the transport lines.

[0092] The three steps a1 to a3 described above constitute one layer of message passing. By stacking K layers, the final state vector of each node can capture information within its K-hop neighborhood. This means that the state of a final destination can perceive the initial state and dynamic changes of all links in the entire transportation path (origin, all transit hubs, and all transportation routes). This mechanism fundamentally solves the cascading effect problem that traditional models cannot handle, achieving accurate dynamic, global, and end-to-end modeling of complex transportation networks. It is the algorithmic cornerstone that produces the outstanding technical effects of this invention.

[0093] exist Figure 3 Based on the illustrated embodiment, as one implementation of this invention, the probability distribution parameters of arrival time are obtained by regression based on the final state vector of the destination node after iterative updates. This may include the following steps, namely steps b1 to b3: Step b1: Perform graph readout operation on the final state vector of the destination node.

[0094] After multiple layers of message passing, the state vectors of all nodes in the entire subgraph have been updated. However, for this prediction task, our real focus is on the destination node. The purpose of graph readout is to selectively extract the information most relevant to the prediction target from the entire complex graph structure, while filtering out other redundant information.

[0095] The specific operation process of this step is: after obtaining the final state vector of all nodes in the entire subgraph, the graph reading function is applied to the graph. This operation directly selects the final state vector of the destination node as the output.

[0096] Step b2, input the read vector into the prediction head, which is realized by a fully connected neural network.

[0097] The vector obtained from the graph reading operation is still a high-dimensional, abstract feature representation, which needs to be mapped to a specific prediction target space. The prediction head is realized by a fully connected neural network. It can be one or more layers of perceptron, which has strong nonlinear fitting ability. This fully connected network learns how to decode the input condensed path state vector into specific numerical parameters that can define a probability distribution. It is an important part of model training, and its parameters are optimized together during training.

[0098] Step b3, output the parameters for defining the probability distribution model through the prediction head.

[0099] This is the essence of the invention that distinguishes it from traditional point prediction models. Instead of outputting a dogmatic, single numerical value, the model outputs the parameters of a probability distribution, thereby completely describing the likelihood of the prediction result. Among them, the number of output layer neurons of the prediction head directly corresponds to the number of parameters of the selected probability distribution model. The activation values of these neurons are the parameters of the distribution model.

[0100] As an optional implementation, the probability distribution model is Gaussian distribution, and the parameters output by the prediction head include the expected mean μ and variance σ² of the arrival time.

[0101] Specifically, Gaussian distribution has mature mathematical properties and efficient calculation, and its two parameters have very intuitive practical significance.

[0102] The first parameter is the expected mean μ: output by one neuron of the prediction head. It represents the most likely arrival time predicted by the model, i.e. the center position of the probability distribution. This is the value that traditional point prediction models try to give.

[0103] The second parameter is the variance σ²: output by another neuron of the prediction head (to ensure positive values, usually exponential transformation is performed on the neuron output). It quantifies the uncertainty or confidence of this prediction.

[0104] Through the present embodiment, the user obtains not a time point, but a time window (such as μ ± 2σ) and the corresponding occurrence probability (such as 95%). This fundamentally changes the way decision makers use prediction results. The size of the variance σ² directly reflects the fluctuation risk of the transportation task. A large σ² value indicates that the path is complex, the real-time state is unstable, or there are uncontrollable factors (such as severe weather), prompting the decision maker to prepare an emergency plan. A small σ² value indicates that the prediction is highly accurate and the journey is expected to be stable. Furthermore, it can enable logistics and supply chain management to shift from passive response to delay to active risk management. For example, a high-uncertainty order can be set with an earlier safety stock deadline or allocated more transportation resource buffers.

[0105] In Figure 3 Based on the embodiment shown, as an implementation manner of the embodiment of the present application, the construction process of the pre-defined global multimodal transport network can include the following steps, steps c1 to c3: Step c1, acquiring multi-source heterogeneous data.

[0106] The multi-source heterogeneous data includes logistics hub attribute data, transportation line data, and macro-environment data.

[0107] Specifically, the logistics hub attribute data is used to describe the inherent ability and real-time state of the node (V). The source of the logistics hub attribute data can be: the operation management system (API) of the port / station / airport, the enterprise resource planning system, the Internet of Things sensor, etc. The logistics hub attribute data can include static data such as the type, design throughput capacity, and geographic location of the hub, as well as dynamic data such as real-time queue length, equipment availability, and future weather forecast.

[0108] The transportation line data is used to describe the inherent attributes and real-time state of the edge (E). The source of the transportation line data can be: GPS and AIS data of ships / vehicles, traffic monitoring platforms (such as Gaode / Google Map API), dispatching systems of transportation enterprises, and transportation price platforms, etc. The transportation line data can include static data such as the fixed length, default transportation mode, and historical average speed of the line, as well as dynamic data such as the current average driving speed, real-time traffic congestion index, and current transportation discount.

[0109] The macro-environment data is used to describe global factors that affect the entire network and cannot be attributed to any node or edge. The source of the macro-environment data can be: public economic data API, government announcements, news data streams, etc. The macro-environment data can include international fuel price index, whether it is a national statutory holiday, macroeconomic fluctuation index, etc.

[0110] Step c2, processing the acquired multi-source heterogeneous data to eliminate data heterogeneity.

[0111] The processing manner includes data cleaning, standardization and alignment.

[0112] Specifically, this step is the key to whether the data can be effectively utilized. It processes the raw data obtained in step c1 into data that can be consumed by the model through a standardized ETL process.

[0113] Data cleaning is used to handle missing values, obvious outliers (such as negative driving speed) and noisy data in the raw data. To ensure data quality, avoid "garbage in, garbage out", and ensure the robustness of the model.

[0114] Data standardization is used to unify the dimensions and units of data from different sources (such as converting miles to kilometers and Fahrenheit to Celsius), and to apply uniform and accurate timestamps to all data. To eliminate the magnitude difference and timing confusion between data, and provide consistency assurance for subsequent feature alignment and model calculation.

[0115] Data alignment: This is the most core step. It accurately maps the cleaned and standardized data to the corresponding nodes, edges or global context in the global multimodal transport network. For example, mapping a piece of AIS trajectory data to the sea transport edge "Shanghai Port-Ningbo Port"; mapping the availability of a port crane to the node "Shanghai Port". To establish a one-to-one correspondence between physical world entities and graph elements (nodes, edges) in the digital twin network, and give the graph structure a real physical meaning and real-time state.

[0116] Step c3, based on the processed data, generate graph data structure G=(V,E,U).

[0117] Where V is the node set, each node represents a logistics hub and has a corresponding node feature vector. E is a directed edge set, each directed edge represents a feasible transport route between nodes and has a corresponding edge feature vector. U is a global feature vector representing the macro state of the entire network.

[0118] Specifically, this step instantiates the processed data into a specific, computer-processable computational graph object G=(V, E, U), completing the final delivery from data to model.

[0119] Where the node set is based on logistics hub data, instantiating each logistics hub as a node in the graph, and assembling the processed corresponding attribute data (static and dynamic) into the node feature vector of the node.

[0120] The set of directed edges is based on the transport line data, instantiating each feasible transport line as a directed edge in the graph, and assembling the corresponding processed line data (static and dynamic) into the edge feature vector of the edge. The directionality of the edge is crucial, as it accurately models the unidirectional flow characteristics of logistics.

[0121] The global feature vector is based on macro-environmental data, generating a global feature vector U. This vector is associated with the entire network and provides uniform macro-background information for all prediction tasks.

[0122] This embodiment systematically constructs a "predefined global multimodal transport network" through the standardized process of c1 (data acquisition) -> c2 (data processing) -> c3 (structure generation). This network is not a static "map" containing only topological relationships, but a deeply processed, feature-rich, and dynamically updated data model that synchronizes with the physical world. It is not only the input of the GNN model, but also the digital cornerstone of the entire invention for precise insight and deduction of the complex real world.

[0123] Based on the embodiment shown in Figure 1 As an implementation manner of the embodiment of the present application, the training process of the graph neural network model can include the following steps, steps d1 to d3: Step d1, obtaining historical transportation order data to construct a training set.

[0124] Each training sample contains: a transport path sequence, real arrival and departure times at each node on the path, and node feature vectors and edge feature vectors corresponding to the historical time.

[0125] This step aims to construct a data set that can fully reflect the dynamics of real-world transportation. The core is that each training sample is not an isolated number, but a complete and context-rich transportation process.

[0126] The transport path sequence provides the topological structure of the transportation task, defining the trajectory of the goods moving on the network. This is the basis for constructing dynamic subgraphs.

[0127] The real arrival and departure times at each node on the path are the supervision signals for training. They provide the real time consumption of goods at each key hub, which is the goal of model learning. More importantly, they imply the real information of node processing efficiency and line transport time.

[0128] The node and edge feature vectors corresponding to historical moments provide the context for high-precision modeling. They record the state of each hub and line at that specific moment (e.g., congestion level, weather, equipment status). This allows the model to learn what transportation outcomes are possible under specific network conditions, rather than seeking a fixed, average transportation time.

[0129] This three-in-one data construction approach, which combines "path structure + real timestamp + spatiotemporal context", provides the model with the necessary, comprehensive learning materials for learning cascade effects and complex dependencies.

[0130] Step d2 involves end-to-end training of the graph neural network model using the training set, achieving the training objective by minimizing the negative log-likelihood loss function.

[0131] The training objective is to minimize the difference between the probability distribution predicted by the model and the actual arrival time distribution.

[0132] Specifically, unlike traditional regression models that use mean squared error to pursue the accuracy of a single value, the training objective of this invention is to make the entire probability distribution predicted by the model as consistent as possible with the distribution of the real data.

[0133] Negative log-likelihood is a classic metric for measuring the goodness of fit of a probabilistic model. Its core idea is that for a sample with a true arrival time of t_true, if the probability distribution P(t) predicted by the model has a high probability density at t_true, then the "likelihood" of the prediction is large, and its negative log-likelihood loss is small; conversely, the loss is large.

[0134] Step d3: Use the backpropagation algorithm and optimizer to iteratively update the parameters of the graph neural network model until the graph neural network model converges.

[0135] Specifically, end-to-end training is a key feature of this implementation. From the node / edge feature encoding layer to the multi-layer message passing mechanism, and finally to the prediction head, the parameters of all modules participate in training as a whole. This means that how information is aggregated and how node states are updated in the message passing mechanism will be automatically adjusted and optimized based on the accuracy (and uncertainty) of the final ETA prediction.

[0136] Backpropagation is the engine of end-to-end training. It calculates the "contribution" (gradient) of each parameter to the loss from the output layer back to the input layer based on the negative log-likelihood loss value.

[0137] Optimizers (such as Adam) intelligently update model parameters using gradients calculated through backpropagation. Optimizers (such as Adam) can adaptively adjust the learning rate, enabling the training process to converge to a better solution more quickly and stably.

[0138] Through this end-to-end optimization approach, the model not only learns how to interpret the characteristics of nodes and edges, but more importantly, it autonomously discovers how information should propagate and aggregate within the transportation network to most effectively predict final arrival times. In other words, it automatically learns how to model cascading effects without requiring manual definition of complex propagation rules.

[0139] Based on the above embodiments, as an implementation of the present invention, the method further includes a graph neural network model update step, namely step e1 and step e2: Step e1: Add the real data of completed shipping orders as new samples to the training set.

[0140] Specifically, the system treats each completed shipment as a new and valuable piece of experimental data. Once the goods successfully arrive at their destination, the system automatically captures data such as the actual arrival and departure times of the order along the entire route.

[0141] Using the initial training data (step d1), the system retrospectively retrieves historical feature data (which can be queried from the historical database) of each node and edge on the corresponding path during the execution of the order. This data is then used to construct a new training sample.

[0142] Step e2 involves incrementally training the deployed graph neural network model using the new samples to enable online learning or periodic updates of the graph neural network model.

[0143] This step is the specific process by which the graph neural network model "learns." It uses new data to update the knowledge of the deployed model in an efficient way.

[0144] In practical applications, graph neural network models can be learned online or updated regularly.

[0145] For online learning, whenever a small number of new samples arrive (such as the completion of a single or small batch of orders), the system immediately, or within a very short delay, uses this new data to make a small update to the model parameters. In this way, the graph neural network model can adapt to changes in the environment almost in real time. For example, when a port suddenly experiences a drop in operational efficiency due to unforeseen circumstances, the system can quickly learn this new change after processing a few affected orders and reflect it in subsequent predictions. It is important to note that to prevent new knowledge from overwriting old knowledge, techniques such as replay buffers can be used to mitigate this.

[0146] For periodic updates, the system collects all new samples within a preset period (e.g., daily, weekly) and then performs a centralized incremental training on the model in a slightly larger batch. This method is more stable and easier to manage and monitor. It can integrate trend changes over a period of time, making the training process more efficient and resource scheduling more convenient. It is suitable for business scenarios where data accumulates rapidly and real-time requirements are not at the second level.

[0147] This implementation introduces a crucial "self-evolution" mechanism into the system. It ensures that the predictive system deployed in the production environment does not degrade in performance over time, but instead learns new lessons from continuous operation, becoming increasingly intelligent with use. This allows it to maintain the accuracy and reliability of its predictions when facing ever-changing logistics networks, emerging transportation modes, or unexpected global events.

[0148] This invention also provides a multimodal transport cargo arrival time prediction system 40 based on graph neural networks, such as... Figure 4 As shown, it includes: The prediction request receiving module 410 is used to receive an arrival time prediction request for the target goods, the arrival time prediction request including the place of origin, destination and goods characteristics; The subgraph extraction module 420 is used to plan candidate transportation routes according to the arrival time prediction request, and dynamically extract a subgraph from the predefined global multimodal transport network based on the candidate transportation routes. The subgraph consists of all nodes passed through by the candidate transportation routes and directed edges connecting these nodes. Nodes represent logistics hubs, and directed edges represent feasible transportation routes between nodes. The subgraph feature initialization module 430 is used to assign a node feature vector to each node in the subgraph, assign an edge feature vector to each directed edge, and assign a global feature vector to the subgraph to obtain the feature-initialized subgraph. The node feature vector of a node includes: static features representing the inherent attributes of the logistics hub corresponding to the node and dynamic features representing the real-time operating status of the logistics hub. The edge feature vector of a directed edge includes static features representing the inherent attributes of the transportation route of the directed edge and dynamic features representing the real-time traffic status of the transportation route. The global feature vector contains macroscopic contextual information related to the prediction task. The graph neural network model prediction module 440 is used to input the feature-initialized subgraph into the pre-trained graph neural network model; and to perform the following operation through the graph neural network model: generate an initial state vector for each node based on the node feature vector; The state vector of a node is iteratively updated through a multi-layer message passing mechanism. Each layer transmits information by combining the edge feature vector with the state vector of the connected node, so that the state of each node can be integrated with the information of its multi-hop neighbor nodes to simulate the cascading propagation effect of transportation events on the path network. Based on the final state vector of the destination node after iterative updates, the probability distribution parameters of the arrival time are obtained by regression.

[0149] This invention provides an electronic device 500, such as... Figure 5 As shown, it includes: One or more processors 501; Memory 502 is 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 described in the first aspect.

Claims

1. A method for predicting the arrival time of multimodal transport goods based on graph neural networks, characterized in that, include: Receive an arrival time prediction request for a target cargo, the arrival time prediction request including the origin, destination and cargo characteristics; Based on the arrival time prediction request, candidate transportation routes are planned, and a subgraph is dynamically extracted from the predefined global multimodal transport network based on the candidate transportation routes. The subgraph consists of all nodes traversed by the candidate transportation routes and directed edges connecting these nodes. Nodes represent logistics hubs, and directed edges represent feasible transportation routes between nodes. Each node in the subgraph is assigned a node feature vector, each directed edge is assigned an edge feature vector, and the subgraph is assigned a global feature vector to obtain the feature-initialized subgraph. The node feature vector of a node includes: static features representing the inherent attributes of the logistics hub corresponding to the node and dynamic features representing the real-time operating status of the logistics hub. The edge feature vector of a directed edge includes static features representing the inherent attributes of the transportation route of the directed edge and dynamic features representing the real-time traffic status of the transportation route. The global feature vector contains macroscopic contextual information related to the prediction task. The feature-initialized subgraph is input into a pre-trained graph neural network model. The graph neural network model performs the following operations: an initial state vector is generated for each node based on the node feature vector; the state vector of the node is iteratively updated through a multi-layer message passing mechanism, where each layer transmits information by combining the edge feature vector with the state vector of the connected node, so that the state of each node can incorporate the information of its multi-hop neighbor nodes to simulate the cascading propagation effect of transportation events on the path network; based on the final state vector of the destination node after iterative updates, the probability distribution parameters of the arrival time are obtained by regression.

2. The method according to claim 1, characterized in that, The iterative update of the node's state vector through a multi-layer message passing mechanism includes the following steps performed at each layer: For each directed edge connecting the source node and the target node, a message is generated by combining the edge feature vector of the directed edge, the current state vector of the source node, and the current state vector of the target node through the message function. For each target node, the aggregated message of the target node is obtained by aggregating the messages corresponding to all its incoming edges using an aggregation function; where all incoming edges are all directed edges pointing to the target node. For each target node, the state vector of the target node is updated by an update function, which combines its current state vector with the aggregated message.

3. The method according to claim 1, characterized in that, The probability distribution parameters for arrival time obtained by regressing the final state vector of the destination node after iterative updates include: Perform a graph readout operation on the final state vector of the destination node; The read vector is input into a prediction head, which is implemented by a fully connected neural network. The prediction head outputs parameters used to define the probability distribution model.

4. The method according to claim 3, characterized in that, The probability distribution model is a Gaussian distribution, and the parameters output by the prediction head include the expected mean μ and variance σ² of the arrival time.

5. The method according to claim 1, characterized in that, The construction process of the predefined global multimodal transport network includes: Acquire multi-source heterogeneous data, including logistics hub attribute data, transportation route data, and macro-environmental data; The acquired multi-source heterogeneous data is processed to eliminate data heterogeneity, wherein the processing methods include data cleaning, standardization and alignment; Based on the processed data, a graph data structure G=(V,E,U) is generated, where V is a set of nodes, each node represents a logistics hub and has a corresponding node feature vector; E is a set of directed edges, each directed edge represents a feasible transportation route between nodes and has a corresponding edge feature vector; and U is a global feature vector, representing the macroscopic state of the entire network.

6. The method according to claim 1, characterized in that, The static features characterizing the inherent attributes of a logistics hub include at least one of the following: hub type, designed throughput capacity, geographical location coordinates, and fixed processing time benchmark; the dynamic features characterizing the real-time operational status of a logistics hub include at least one of the following: real-time queue length, current equipment availability, and weather forecast for a specific future time window. The static characteristics of the inherent attributes of the transportation route include at least one of the following: fixed length of the route, default transportation mode, historical average speed, and standard freight rate; the dynamic characteristics of the real-time traffic status of the transportation route include at least one of the following: current average driving speed, real-time traffic congestion index, and freight discount for the current time period. The global feature vector includes at least one of the following: international fuel price index, whether it is a national statutory holiday, and macroeconomic fluctuation index.

7. The method according to claim 1, characterized in that, The training process of the graph neural network model includes: Historical transportation order data is obtained to construct a training set, where each training sample contains: a transportation path sequence, the actual arrival and departure times at each node of the path, and the node feature vector and edge feature vector at the corresponding historical moment; The graph neural network model is trained end-to-end using the training set, and the training objective is achieved by minimizing the negative log-likelihood loss function. The training objective is to minimize the difference between the probability distribution predicted by the model and the actual arrival time distribution. The parameters of the graph neural network model are iteratively updated using the backpropagation algorithm and optimizer until the graph neural network model converges.

8. The method according to claim 7, characterized in that, The method also includes a graph neural network model update step: Add real data of completed shipping orders as new samples to the training set; Incremental training of deployed graph neural network models can be performed using new samples to enable online learning or periodic updates of graph neural network models.

9. A multimodal transport cargo arrival time prediction system based on graph neural networks, characterized in that, include: The prediction request receiving module is used to receive an arrival time prediction request for the target goods, wherein the arrival time prediction request includes the origin, destination and goods characteristics; The subgraph extraction module is used to plan candidate transportation routes according to the arrival time prediction request, and dynamically extract a subgraph from the predefined global multimodal transport network based on the candidate transportation routes. The subgraph consists of all nodes traversed by the candidate transportation routes and directed edges connecting these nodes. Nodes represent logistics hubs, and directed edges represent feasible transportation routes between nodes. The subgraph feature initialization module is used to assign a node feature vector to each node in the subgraph, an edge feature vector to each directed edge, and a global feature vector to the subgraph, thus obtaining the feature-initialized subgraph. A node's feature vector includes: static features representing the inherent attributes of the logistics hub corresponding to the node and dynamic features representing the real-time operational status of the logistics hub. An edge feature vector of a directed edge includes static features representing the inherent attributes of the transportation route of the directed edge and dynamic features representing the real-time traffic status of the transportation route. The global feature vector contains macroscopic contextual information related to the prediction task. The graph neural network model prediction module is used to input the feature-initialized subgraph into the pre-trained graph neural network model; and to perform the following operations through the graph neural network model: generate an initial state vector for each node based on the node feature vector; The state vector of a node is iteratively updated through a multi-layer message passing mechanism. Each layer transmits information by combining the edge feature vector with the state vector of the connected node, so that the state of each node can be integrated with the information of its multi-hop neighbor nodes to simulate the cascading propagation effect of transportation events on the path network. Based on the final state vector of the destination node after iterative updates, the probability distribution parameters of the arrival time are obtained by regression.

10. 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 1-8.

Images