Multi-modal transport path planning method and device based on space-time probability hypergraph

Abstract

The application belongs to the field of intelligent logistics, and discloses a multimodal transport path planning method and device based on a space-time probability hypergraph, which comprises the following steps: acquiring space-time data of a target multimodal transport network and constructing a space-time probability hypergraph; predicting the delay probability and cancellation probability of each hyperedge and calculating the reliability weight vector of each hyperedge; based on the space-time probability hypergraph, screening path prefixes meeting preset hard constraint conditions to generate a candidate path prefix set; performing robust bundle search according to the candidate path prefix set and the reliability weight vector of each hyperedge to obtain a candidate path set; evaluating each candidate path in the candidate path set to obtain a target vector of each candidate path; performing Pareto hierarchical screening according to the target vector of the candidate path to output a transport path planning result of the target multimodal transport network; and the application can output multiple sets of main and backup transport paths considering time efficiency, cost, carbon emission and transport reliability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent logistics technology, and relates to multimodal transport intelligent logistics route planning technology, and in particular to a multimodal transport route planning method and device based on spatiotemporal probability hypergraph. Background Technology

[0002] The global cross-border supply chain has long faced a complex and ever-changing external environment. As the core mode of cross-regional freight transportation, multimodal transport links various transportation carriers such as roads, railways, sea, and air, passing through a large number of transit nodes such as ports, border crossings, and inland terminals. The entire transportation process is constantly affected by various uncertainties. At present, the scale of cross-border container transportation business continues to expand, and customers are simultaneously putting forward rigid requirements for transportation timeliness, comprehensive costs, carbon emission indicators, and transportation performance stability. Multimodal transport route planning has become the core technical support for supply chain scheduling.

[0003] Currently, traditional multimodal transport route planning technologies generally adopt static graph network modeling schemes, with nodes representing terminals and hyperedges representing single transport segments. They rely solely on fixed distances, fixed durations, and fixed freight rates as path weights, and the mainstream solution algorithms are shortest path algorithms and K-shortest path algorithms. Among them, traditional shortest path or K-shortest path algorithms that rely on static weights often assume that the transport network is stable and the constraints are simple. However, in practical applications, multiple sources of disturbances, such as port congestion, extreme weather, temporary cancellations of trains / flights, adjustments to port clearance windows, fuel and exchange rate fluctuations, and temporary regional travel restrictions, continuously change the actual operating status of the transport link. Traditional static route planning schemes are unable to provide robust and interpretable solutions when node states change rapidly. They suffer from problems such as inability to adapt to dynamic supply chains, lack of risk prediction, insufficient resilience of single paths, and difficulty in coordinating solutions to multiple constraints, leading to temporary route changes, cost overruns, or time-delays.

[0004] Specifically, firstly, traditional static route planning schemes fail to probabilistically model the two core uncertainties of transport segment delays and cancellations, making it impossible to quantify the overall transport risk of the route in advance. They can only passively adjust routes after transport failures occur, easily leading to uncontrolled costs and cargo delays. Secondly, existing schemes lack a time extension layer modeling mechanism, failing to express the temporal constraints of node transfer waiting and cross-transportation mode connections under different time slices. The resulting routes may have infeasible solutions such as insufficient transfer time or missed customs windows. Furthermore, existing methods separate hard constraint verification from route generation, failing to filter invalid branches that violate time windows, transport capacity, policy compliance, and carbon emission thresholds during the route prefix extension stage, resulting in massive computational loads and slow response times. Simultaneously, existing schemes employ a fixed-bundle-width general bundle search strategy, using the same extension scale for high-risk and low-risk paths, either missing high-quality, robust paths or invalidating redundant branches. In addition, existing schemes generally only output a single optimal path, failing to provide a multi-objective balanced primary and backup path combination. Faced with node failures and route cancellations, there are no reliable alternatives, resulting in insufficient supply chain resilience. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a multimodal transport route planning method and apparatus based on spatiotemporal probabilistic hypergraphs, in order to solve the technical problems that traditional static route algorithms cannot adapt to dynamic supply chains, lack risk prediction, have insufficient resilience of single routes, and are difficult to solve collaboratively with multiple constraints.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a multimodal transport route planning method based on a spatiotemporal probabilistic hypergraph, comprising: Acquire spatiotemporal data of the target multimodal transport network; wherein, the spatiotemporal data includes static structural data and dynamic environmental data of the multimodal transport network; Based on the spatiotemporal data of the target multimodal transport network, a spatiotemporal probabilistic hypergraph is constructed; Using a temporal graph neural network, the delay probability and cancellation probability of each hyperedge in the spatiotemporal probability hypergraph are predicted, and the reliability weight vector of each hyperedge is calculated. Based on the spatiotemporal probabilistic hypergraph and using a multi-constraint label setting algorithm, path prefixes that meet the preset hard constraints are selected to generate a candidate path prefix set. Based on the candidate path prefix set and the reliability weight vector of each hyperedge, a robust bundle search is performed to obtain the candidate path set; By using a multi-objective evaluation function, each candidate path in the candidate path set is evaluated to obtain the objective vector of each candidate path; Based on the target vector of each candidate path, Pareto hierarchical screening is performed on each candidate path in the candidate path set, outputting the Pareto optimal path set, and determining the main path and at least one alternative path from the Pareto optimal path set as the transportation path planning result of the target multimodal transport network.

[0007] Furthermore, the spatiotemporal probabilistic hypergraph includes a node set, a hyperedge set, and a time extension layer; each hyperedge in the hyperedge set represents an independent transportation sub-path and is attached with a set of probability attributes, including the hyperedge's on-time probability, delay probability, and cancellation probability; the time extension layer is used to replicate each node along the time axis into multiple time-slice nodes and configure waiting connection edges for each time-slice node.

[0008] Furthermore, the reliability weight vector for each hyperedge includes time reliability weight, cost reliability weight, and carbon emission reliability weight.

[0009] Furthermore, based on the spatiotemporal probabilistic hypergraph and utilizing a multi-constraint label setting algorithm, the process of filtering path prefixes that satisfy preset hard constraints and generating a candidate path prefix set is as follows: A label set is set for each spatiotemporal node in the spatiotemporal probabilistic hypergraph. Each label in the label set contains the cumulative time, cumulative cost, cumulative carbon emissions, occupied capacity, and policy area codes traversed from the source node to the corresponding spatiotemporal node. When expanding from the first spatiotemporal node to the second spatiotemporal node along the first hyperedge, constraint verification is performed according to preset hard constraints. The new labels generated by the expansion that passes the constraint verification are added to the label set of the second spatiotemporal node, and Pareto dominance pruning is performed on the label set of the second spatiotemporal node. All spatiotemporal nodes in the spatiotemporal probabilistic hypergraph are traversed, and a candidate path prefix set is output. The preset hard constraints include time window constraints, transportation capacity constraints, policy compliance constraints, and total carbon emission threshold constraints.

[0010] Furthermore, based on the candidate path prefix set and the reliability weight vector of each hyperedge, a robust bundle search is performed to obtain the candidate path set, as follows: In each layer of the robust bundle search, a comprehensive risk score is calculated for each candidate path prefix in the current layer; wherein the comprehensive risk score is calculated as a weighted sum of the delay probabilities and cancellation probabilities of each hyperedge contained in the candidate path prefix; Obtain the risk baseline value for the current layer; where the risk baseline value is the average of the comprehensive risk scores of all candidate path prefixes in the current layer; For candidate path prefixes whose comprehensive risk score is higher than the risk benchmark value multiplied by a predetermined multiple, the extension beamwidth of the corresponding candidate path prefix will be increased according to the preset initial beamwidth value. For candidate path prefixes whose comprehensive risk score is not higher than the risk benchmark value multiplied by a predetermined multiple, the extended beamwidth is kept at the preset initial beamwidth value.

[0011] Furthermore, the target vector for each candidate route includes expected transport time, expected transport cost, time risk value, cost risk value, total carbon emissions, reliability score, and route redundancy.

[0012] Secondly, the present invention provides a multimodal transport route planning system based on a spatiotemporal probabilistic hypergraph, used to implement the multimodal transport route planning method based on a spatiotemporal probabilistic hypergraph described in the first aspect, including: The spatiotemporal data acquisition module is used to acquire spatiotemporal data of the target multimodal transport network; wherein, the spatiotemporal data includes static structural data and dynamic environmental data of the multimodal transport network; The graph construction module is used to construct a spatiotemporal probabilistic hypergraph based on the spatiotemporal data of the target multimodal transport network. The weight calculation module is used to predict the delay probability and cancellation probability of each hyperedge in the spatiotemporal probability hypergraph using a temporal graph neural network, and calculate the reliability weight vector of each hyperedge. The prefix set generation module is used to filter path prefixes that meet preset hard constraints based on the spatiotemporal probabilistic hypergraph and using a multi-constraint label setting algorithm to generate a candidate path prefix set. The candidate path search module is used to perform robust bundle search based on the candidate path prefix set and the reliability weight vector of each hyperedge to obtain the candidate path set; The candidate path evaluation module is used to evaluate each candidate path in the candidate path set using a multi-objective evaluation function, and obtain the objective vector of each candidate path. The candidate route selection module is used to perform Pareto hierarchical selection on each candidate route in the candidate route set according to the target vector of each candidate route, output the Pareto optimal route set, and determine the main route and at least one alternative route from the Pareto optimal route set as the transportation route planning result of the target multimodal transport network.

[0013] Thirdly, the present invention provides an electronic device, comprising: A processor is used to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, performs the multimodal transport route planning method based on spatiotemporal probabilistic hypergraph as described in the first aspect.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multimodal transport route planning method based on spatiotemporal probabilistic hypergraph described in the first aspect.

[0015] Fifthly, the present invention provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the multimodal transport route planning method based on spatiotemporal probabilistic hypergraph described in the first aspect.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a multimodal transport route planning method based on a spatiotemporal probabilistic hypergraph. By constructing a spatiotemporal probabilistic hypergraph that integrates static structural data and dynamic environmental data, and combining it with a temporal graph neural network to predict the delay and cancellation probabilities of each transport hyperedge and generate a reliability weight vector, it achieves pre-quantitative prediction of transport link uncertainty risks, effectively avoiding problems such as runaway transport costs and cargo delays and defaults. Simultaneously, it utilizes a spatiotemporal expansion modeling mechanism to accurately describe the temporal constraints of node transfers and transport connections under different time slices, significantly improving the practical feasibility of route planning. Through a multi-constraint label setting algorithm, it filters invalid branches in terms of time windows, transport capacity, compliance, and carbon emissions during the route prefix expansion stage, significantly reducing the computational load of the algorithm and improving route response efficiency. Finally, it leverages robustness adapted to differences in link risks. The bundle search strategy avoids missing high-quality and robust paths and reduces redundant and ineffective calculations. Furthermore, by employing multi-objective evaluation and Pareto hierarchical filtering, it outputs the optimal primary and backup path combinations, achieving synergistic optimization across multiple constraints including timeliness, cost, risk, and carbon emissions. This significantly improves the robustness, flexibility, and overall supply chain resilience of cross-border multimodal transport routes. The invention explicitly models transport uncertainty and temporal sequence relationships using a spatiotemporal probabilistic hypergraph, quantifies transport risks using a temporal sequence graph neural network, and combines multi-constraint filtering, adaptive robust bundle search, and multi-objective Pareto optimization. This enables the output of multiple primary and backup transport routes that balance timeliness, cost, carbon emissions, and transport reliability in multi-source uncertainties such as port congestion, sudden weather changes, and policy shifts. It can be widely applied to scheduling scenarios for cross-border container multimodal transport, including sea-rail, air-rail, and land bridge transport.

[0017] The multimodal transport route planning system, electronic device, computer-readable storage medium, and computer program product based on spatiotemporal probabilistic hypergraphs provided by this invention possess all the advantages of the aforementioned multimodal transport route planning methods based on spatiotemporal probabilistic hypergraphs. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of the multimodal transport route planning method based on spatiotemporal probabilistic hypergraph provided by the present invention; Figure 2 The flowchart illustrates the principle of the multimodal transport route planning method based on spatiotemporal probabilistic hypergraph provided in Example 1. Figure 3 This is a structural block diagram of the multimodal transport route planning system based on spatiotemporal probabilistic hypergraph provided in Example 2; Figure 4 This is a structural block diagram of the electronic device provided in Example 4. Detailed Implementation

[0020] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0021] As attached Figure 1 As shown, this invention provides a multimodal transport route planning method based on a spatiotemporal probabilistic hypergraph, comprising the following steps: Step 100: Obtain the spatiotemporal data of the target multimodal transport network; wherein, the spatiotemporal data includes the static structural data and dynamic environmental data of the multimodal transport network.

[0022] Step 200: Construct a spatiotemporal probabilistic hypergraph based on the spatiotemporal data of the target multimodal transport network.

[0023] Step 300: Using a temporal graph neural network, predict the delay probability and cancellation probability of each hyperedge in the spatiotemporal probability hypergraph, and calculate the reliability weight vector of each hyperedge.

[0024] Step 400: Based on the spatiotemporal probabilistic hypergraph and using a multi-constraint label setting algorithm, filter out path prefixes that meet the preset hard constraint conditions and generate a candidate path prefix set.

[0025] Step 500: Based on the candidate path prefix set and the reliability weight vector of each hyperedge, perform robust bundle search to obtain the candidate path set.

[0026] Step 600: Use a multi-objective evaluation function to evaluate each candidate path in the candidate path set and obtain the objective vector of each candidate path.

[0027] Step 700: According to the target vector of each candidate path, perform Pareto hierarchical screening on each candidate path in the candidate path set, output the Pareto optimal path set, and determine the main path and at least one alternative path from the Pareto optimal path set as the transportation path planning result of the target multimodal transport network.

[0028] In the above implementation, by acquiring spatiotemporal data of the target multimodal transport network and constructing a spatiotemporal probabilistic hypergraph, the terminals, transport segments, and their temporal relationships in the multimodal transport network are integrated into a hyperedge structure. A temporal graph neural network is used to dynamically predict the delay and cancellation probabilities of the hyperedges affected by disturbances such as port congestion, weather disasters, and capacity fluctuations, thereby quantifying transport risks into a reliability weight vector. Based on this, a multi-constraint label setting algorithm is used to simultaneously verify the hard constraints of time windows, capacity, and carbon emissions during the path prefix expansion stage, pruning invalid branches in advance and reducing computational redundancy. Robust bundle search dynamically adjusts the expansion bundle width based on the reliability weights, optimizing... First, high-reliability paths are retained to avoid the omission of high-quality and robust paths by traditional fixed-bundle-width strategies. Pareto hierarchical screening is performed based on a multi-objective evaluation function to output a primary and backup path combination that balances timeliness, cost, risk, and carbon emission. Compared with the traditional static graph network modeling method that only relies on fixed distance, duration, and freight rate as weights, this invention embeds the time-varying state and probabilistic risk of the transportation link into the graph structure, completing risk prediction and multi-constraint collaborative verification before path generation. This makes the planning results dynamic, adaptable to the environment, controllable in risk, and resilient to the supply chain, effectively solving the problems of passively responding to failures, solving multiple constraints separately, and insufficient anti-interference capability of single paths in traditional methods.

[0029] The following specific embodiments further explain the multimodal transport route planning method based on spatiotemporal probabilistic hypergraphs provided by the present invention: Example 1 As attached Figure 2 As shown, this embodiment 1 provides a multimodal transport route planning method based on spatiotemporal probabilistic hypergraphs, including the following steps: Step 1: Obtain the spatiotemporal data of the target multimodal transport network; wherein, the spatiotemporal data includes the static structural data and dynamic environmental data of the multimodal transport network.

[0030] Interpretive, static structured data refers to the long-term, fixed, and unchanging basic attributes of the road network within the target multimodal transport network. Specifically, static structured data includes the location coordinates of all transport nodes within the target multimodal transport network, transport node type identifiers, physical connections between transport nodes, preset planned operating time windows for each transport segment, airport flight schedules, number of port berths, yard capacity, basic carbon emission factors for each transport segment, shortest road path length, basic customs clearance time at ports, and basic freight rate benchmarks for various transport vehicles. Optionally, static structured data can be sourced from logistics company road network databases, port terminal static archives, customs rules databases, and basic route archives of railway / airline operators. The data update cycle is monthly or quarterly, with no real-time time-series fluctuations.

[0031] Dynamic environmental data refers to real-time, time-series changes in disturbances and operational parameters. Specifically, dynamic environmental data includes the real-time number of vessels in port at each port node, the average port loading and unloading efficiency, the full-cycle record of historical delay events in each transportation segment, medium- and long-term weather forecast data, real-time fuel market prices, real-time cross-border settlement exchange rates, regional temporary policy restrictions, real-time airspace capacity, real-time port tidal data, typhoon path forecasts, real-time carbon trading market prices, and remaining carbon allowances for enterprises. Optionally, dynamic environmental data can be sourced from port real-time dispatch systems, meteorological service platforms, commodity trading platforms, customs real-time policy push interfaces, railway train real-time dispatch terminals, and air traffic control systems. This data has minute-level time-series update characteristics and can continuously change the risk and cost indicators of the transportation chain.

[0032] Optionally, after acquiring the raw spatiotemporal data, a unified preprocessing operation is also included. This preprocessing operation includes periodic feature encoding, spatial geographic feature embedding, confidence weighting of outlier data, interpolation and completion of missing time series data, unified conversion of cross-currency freight rates, and normalization of carbon emission units. After preprocessing, all spatiotemporal data are uniformly written into the time series graph database for storage, ensuring that the data is formatted uniformly, time-series aligned, and free from missing data and numerical deviations when called in subsequent steps.

[0033] For example, in the preprocessing operation, for historical delay event records, each record is bound with a corresponding meteorological condition label, port congestion level label, and transport carrier type label, which serve as input features for subsequent training and inference of the time series neural network; for sudden abnormal event signals such as extreme weather, port strikes, and line failures, the abnormality type, abnormality duration, affected nodes, and transport segments are separately marked in the preprocessing operation for subsequent triggering of conservative planning strategies.

[0034] Optionally, when starting a brand new route planning, the origin-end point OD pair corresponding to the route planning request is extracted first, and the OD-level cache library is searched to determine whether there is any unexpired historical cache data. If the cache is valid, the process of re-collecting part of the full data can be skipped, and only the indicators that have changed in the dynamic environment data are refreshed, which greatly reduces the overhead of duplicate data collection. Only when the cache does not exist or expires will the full static and dynamic spatiotemporal data be collected.

[0035] Step 2: Construct a spatiotemporal probabilistic hypergraph based on the spatiotemporal data of the target multimodal transport network. This involves constructing a spatiotemporal probabilistic hypergraph to complete the probabilistic and temporal graph structure modeling of the multimodal transport network. The hypergraph encapsulates multimodal transport combination sub-paths through hyperedges, describes transit waiting and temporal connection constraints through a time extension layer, and adds complete uncertainty probability attributes to each transport link.

[0036] In one possible implementation, the spatiotemporal probabilistic hypergraph includes a set of nodes, a set of hyperedges, and a temporal extension layer; wherein, the mathematical expression of the spatiotemporal probabilistic hypergraph is as follows:

[0037] in, It is a spatiotemporal probability hypergraph; A set of nodes; It is a set of superedges; A collection of node attributes; For the set of attributes of the superedge; It is a set of discrete time slices.

[0038] Explanatoryly, the node set corresponds to all physical stations, ports, airports, and border crossings, and the hyperedge set corresponds to any independent transportation sub-path containing one or more transportation modes. The time extension layer is used to assign a temporal dimension to the physical nodes. Each hyperedge in the hyperedge set represents an independent transportation sub-path and is attached with a set of probability attributes, including the hyperedge's on-time probability, delay probability, and cancellation probability. The time extension layer is used to replicate each node along the time axis into multiple time-slice nodes and configure waiting connection edges for each time-slice node.

[0039] For example, each hyperedge in the hyperedge set independently represents a complete and independent transportation sub-path. The sub-path can be composed of any two or more transportation carriers such as road + rail, rail + sea, and air + road. Each hyperedge is bound to a set of probability attributes, which fully includes three types of risk indicators for the transportation sub-path: on-time probability, delay probability, and cancellation probability. It also includes business attributes such as basic transportation time, basic transportation cost, basic carbon emissions, transportation capacity limit, policy area code of the route, and detour factor.

[0040] The implementation logic of the time extension layer is to copy each original physical node along the time axis corresponding to the planned time window to generate a large number of discrete time slice spatiotemporal nodes, so as to distinguish the transit status of the same physical node under different arrival and departure times. When constructing the time extension layer, the total planned time window and the preset minimum discrete time unit are read first. The minimum time unit is used to divide the complete planned time window into several independent time slices of equal duration. Each physical node will generate a unique corresponding spatiotemporal node in each time slice.

[0041] For example, two types of connection edges are built synchronously within the time extension layer. The first type is a waiting connection edge between adjacent time slice spatiotemporal nodes of the same physical node. The waiting connection edge is accompanied by waiting time cost, waiting energy consumption and carbon emissions, and risk accumulation coefficient during the waiting period. It is used to simulate the business scenario of goods arriving at the station in advance and waiting for the next transport shift. The second type is a transit connection edge between spatiotemporal nodes of different physical nodes. It is only created when the difference between the end time of the preceding super-edge and the start time of the following super-edge is greater than the station's minimum transit time threshold. The transit connection edge is accompanied by transit operation time and transit carbon emission factor. It is used to simulate the business logic of goods loading and unloading across transport carriers and port customs clearance transit.

[0042] Specifically, the construction process of the spatiotemporal probabilistic hypergraph is as follows: Step 21: Construct the basic node set. Read all physical stations, ports, airports, and border crossings from the static spatiotemporal data, assign a unique global node identifier to each physical node, and summarize them to form the original physical node set.

[0043] Step 22: Construct a set of hyperedges. Traverse all composable multi-mode transport sub-paths, generating a hyperedge for each continuous and independent intermodal transport link, and binding each hyperedge to a complete set of probability attributes; where the mathematical definition of the three indicators within the probability attribute set is: On-time probability... This represents the probability that the transport sub-path will arrive at its destination on time within the planned time window, and the probability of delay. This represents the probability that the arrival time of the transport will exceed the planned time threshold, and the probability of cancellation. This represents the probability that the transport segment or route is completely canceled and unable to carry goods. All three values ​​range from 0 to 1 and satisfy the following conditions: .

[0044] Step 23: Construct a time extension layer. Read the start and end times of the global planning time window and the preset minimum discrete time unit, and evenly divide the complete planning time window into... Each physical node generates a unique spatiotemporal node within each time slice, and the spatiotemporal node identifier is composed of the physical node ID and the time slice sequence number. Subsequently, two types of connecting edges are built: waiting connecting edges are created between adjacent spatiotemporal nodes of the same physical node; among them, waiting connecting edges are used to simulate the scenario of goods arriving at the port in advance and waiting for the shift; and transit connecting edges are created only between spatiotemporal nodes of different physical nodes when the transit time reaches the target, which is used to simulate the loading and unloading transit scenario across transportation modes.

[0045] It should be noted that traditional graph models only have single-layer nodes and single-edge structures, lacking the ability to encapsulate multimodal transport with hyperedges and to perform time-series modeling with extended time layers. They cannot distinguish between differentiated transfer scenarios arriving at the same terminal at different times, nor can they encapsulate multi-carrier combined transport links such as road-to-rail or sea-to-road transfers in one go. In this embodiment 1, the three types of information—road network topology, transport risk probability, and time-series transfer constraints—are decoupled and modeled to integrate multimodal transport, time-series waiting, transfer connections, and transport risk probabilities into a unified graph structure, providing a basis for subsequent reliability prediction and path constraint filtering. It provides a standardized modeling carrier with clear logic and facilitates feature reading by time-series graph neural networks, significantly reducing the difficulty of model feature extraction. Among them, the node set is the basic topological unit of the spatiotemporal probabilistic hypergraph, and all spatiotemporal nodes carry the cumulative constraint labels in the path expansion process; the hyperedge set is the carrier of transportation business logic, which fully quantifies the performance uncertainty of each intermodal transport link through the probabilistic attribute set; the time extension layer is the carrier of time-series constraints, which fully describes the time-series rules of transit waiting and cross-carrier connection through discrete time slices and two types of connecting edges, so as to fully express all business constraints of multimodal transport.

[0046] In this embodiment 1, by constructing the spatiotemporal probabilistic hypergraph as a three-layered component consisting of a node set, a hyperedge set with probability attributes, and a time extension layer, independent optimization and adjustment can be made for three types of business logic: road network topology, transportation risk, and time-series transfer. Hyperedges can be flexibly added or deleted to correspond to transportation sub-paths, the time slice granularity can be freely adjusted according to business accuracy requirements, and the node set can be added to stations and ports as needed. The overall modeling framework has strong business scalability and fully and explicitly models the three core uncertain probabilities of the transportation link, providing a standardized probability carrier for subsequent time-series graph neural network risk prediction and reliability weight calculation. This enables transportation risk to run through the entire route planning process, achieving full-link risk perception from modeling to solution.

[0047] For example, in the Qingdao-Moscow sea-rail intermodal transport scenario, in this embodiment 1, the entire combined link from Qingdao Port to Vladivostok Port by sea and then to Moscow by rail can be modeled as an independent hyperedge. The hyperedge is bound to the probability of sea freight delay and the probability of cancellation due to typhoon weather. At the same time, all physical nodes of Qingdao Port, Vladivostok Port and Moscow Railway Station are replicated in time slices according to the minimum time unit of 12 hours. Waiting connection edges and railway transfer connection edges are built between each time slice node to fully express the time sequence process of ships waiting for loading and unloading after arriving at the port and transferring to railway at the port.

[0048] Step 3: Using a spatiotemporal graph neural network, predict the delay probability and cancellation probability of each hyperedge in the spatiotemporal probability hypergraph, and calculate the reliability weight vector of each hyperedge.

[0049] Interpretive methods are used to predict the delay and cancellation probabilities of each hyperedge in the spatiotemporal probabilistic hypergraph, thereby quantifying the uncertainty risk of the transportation link. Based on full spatiotemporal data, the correlation between road network dynamic disturbances and transportation delays and cancellations is learned. Then, based on the predicted probability indicators, dynamic reliability weights that take into account three dimensions: timeliness, cost, and carbon emissions are generated. This allows risk indicators to be pre-integrated into path weight calculation, enabling path solving to naturally avoid high-risk sections.

[0050] In this embodiment 1, the temporal graph neural network used is a graph neural network model with a Bayesian probability output layer. The model is trained offline in advance based on historical full-cycle transportation data. During the training phase, the input features include historical delay event records for each hyperedge, weather forecast data, real-time number of ships in port, tidal data, congestion index, and dynamic environmental indicators of airspace visibility. The model training objective is to output the probability labels of delay events and cancellation events occurring on each hyperedge within the future planning time window. After training, the model weights are fixed. During the online inference phase, real-time dynamic spatiotemporal data is directly input to complete the probability prediction update.

[0051] For example, the online prediction update process is as follows: all the collected and preprocessed real-time dynamic environmental data are input into the temporal graph neural network. The model completes the temporal aggregation of node and hyperedge features and the transmission of neighborhood information based on the graph topology of the spatiotemporal probability hypergraph. Finally, it outputs the delay probability and cancellation probability of each hyperedge in the current planning cycle, and simultaneously outputs the baseline time weight, baseline cost weight, and baseline carbon emission weight.

[0052] In one possible implementation, the reliability weight vector for each hyperedge includes a time reliability weight, a cost reliability weight, and a carbon emission reliability weight. The time reliability weight is a time-dimensional risk amplification coefficient, used to quantify the amplified impact of transportation delays on overall transportation time; the higher the delay probability value, the greater the time reliability weight, which significantly increases the time loss weight of that hyperedge segment during path solving. The cost reliability weight is an economic cost risk amplification coefficient, simultaneously superimposing two types of economic losses: port warehousing costs due to delays and rebooking and transshipment costs due to cancellations, while also relating to the dual variables of delay probability and cancellation probability. The carbon emission reliability weight is a carbon emission risk amplification coefficient, mainly calculated based on the transportation segment detour factor. The detour factor is the ratio of the actual transportation path length to the shortest direct path length; the longer the detour distance, the higher the additional carbon emissions, and the carbon emission reliability weight is amplified accordingly.

[0053] Explanatoryly, each hyperedge corresponds to a unique set of reliability weight vectors. The vectors contain time reliability weights, cost reliability weights, and carbon emission reliability weights, which correspond to transportation timeliness risk, transportation cost risk, and carbon emission detour risk, respectively. The weight values ​​of time reliability weights, cost reliability weights, and carbon emission reliability weights will dynamically change with the probability of delay and the probability of cancellation, realizing the adjustment logic that the higher the risk, the greater the corresponding dimension weight.

[0054] Specifically, the reliability weight vector is expressed as follows:

[0055] in, This is the reliability weight vector; For time reliability weights; Cost reliability weight; As a weight for carbon emission reliability; The calculation process for time reliability weights is as follows:

[0056] in, As the baseline time weight; The time risk sensitivity coefficient; This represents the delay probability.

[0057] The cost-reliability weights are as follows:

[0058] in, Baseline cost weight; The risk coefficient for delay costs; To determine the cost and risk factor for cancellation; This represents the probability of cancellation.

[0059] The carbon emission reliability weights are as follows:

[0060] in, As a benchmark carbon emission weight; The carbon emission bypass sensitivity coefficient; The detour factor is determined based on the ratio of the actual path length to the shortest path length of the transport segment corresponding to the superedge.

[0061] In this embodiment 1, by setting a reliability weight vector that includes three dimensions: time, cost, and carbon emissions, the differentiated impacts of two types of transportation uncertainty events, delay and cancellation, on timeliness, economy, and carbon emissions can be quantified separately, avoiding the defect that a single weight cannot distinguish multi-dimensional losses. The reliability weight vector serves as the input parameter for robust bundle search and multi-objective evaluation function, running through the entire process of route generation and evaluation. This ensures that the calculation process of each candidate route fully incorporates the losses of the three types of transportation risks, and the final output of the primary and backup routes can simultaneously balance timeliness, cost, carbon emissions, and performance reliability, meeting the diversified operation and management needs of modern cross-border multimodal transport supply chains.

[0062] It should be noted that in the calculation of the reliability weight vector, time risk sensitivity coefficient, delay cost risk coefficient, cancellation cost risk coefficient, and carbon emission detour sensitivity coefficient are introduced to adaptively adjust the amplification of the reliability weight. Among them, the time risk sensitivity coefficient, delay cost risk coefficient, cancellation cost risk coefficient, and carbon emission detour sensitivity coefficient are optimized by using real transportation records within the historical transportation playback period. The optimization objective is to minimize the error between the predicted indicators and the actual transportation indicators, thereby ensuring that the weight adjustment range conforms to the actual business loss pattern.

[0063] Explanatoryly, when real-time abnormal event signals such as port strikes, extreme typhoons, and line failures are detected, a conservative strategy mode is triggered. The predicted delay probability and cancellation probability of all super-edges affected by the anomaly are increased by a preset increment, and the three types of reliability weights are amplified simultaneously. This forces the route solving stage to prioritize robust and low-disturbance transportation links, avoiding large-scale performance risks caused by extreme emergencies. This achieves a technical shift from post-event remediation to pre-event quantitative prediction of transportation risks. Traditional route planning schemes do not consider the probability of delay and cancellation events at all, assuming that all transportation segments can be fulfilled on time. Once a disturbance occurs, all route schemes become invalid.

[0064] In this embodiment 1, a time-series graph neural network is used to predict the probability of risk and transform the risk into path weights. This reduces the probability of high-risk paths being selected in subsequent search stages from the source, significantly reducing temporary route changes and breach of contract compensation issues during actual transportation. For example, in the land bridge transportation scenario where loading and unloading efficiency at the Khorgos Port is reduced and strong wind warnings are issued along the route, the time-series graph neural network will predict that the probability of delays exceeding the edge of all railways associated with the port will increase to 0.31 and the probability of cancellation will increase to 0.06. Based on this probability value, the time reliability weight and cost reliability weight are amplified, so that the system will automatically prioritize the selection of alternative routes passing through the Almaty Port in the subsequent path search stage, avoiding the original route of the high-risk Khorgos Port.

[0065] Step 4: Based on the spatiotemporal probabilistic hypergraph and using a multi-constraint label setting algorithm, path prefixes that satisfy preset hard constraints are selected to generate a candidate path prefix set. By selecting path prefixes that meet the preset hard constraints, all invalid path branches that violate business hard control rules are filtered out in the early stages of path expansion, significantly reducing the computational space of subsequent robust bundle search, improving overall solution efficiency, and ensuring that all final output paths naturally meet the four types of rigid control requirements: time, capacity, policy, and carbon emissions, with no unexecutable violating schemes.

[0066] The interpretable, multi-constraint label setting algorithm specifically binds a label set to each spatiotemporal node within the spatiotemporal probabilistic hypergraph. For each incomplete path prefix extending from the source node to the current spatiotemporal node, an independent label is generated and stored in the label set of that node. The label fully records all cumulative constraint indicators of the path prefix up to the current node, which are used for constraint verification during each hyperedge expansion.

[0067] Each tag contains five cumulative indicators: cumulative transportation time from the global source node to the current spatiotemporal node, cumulative transportation cost, cumulative carbon emissions, the cargo transportation capacity occupied by the route, and the set of all policy area codes that the route has passed through. The five indicators fully cover all the data required for the four types of pre-set hard constraint verification, without the need to retrieve external data to complete the verification.

[0068] The complete process of performing path prefix expansion and constraint verification is as follows: When extending the path prefix from the first spatiotemporal node to the second spatiotemporal node along any first hyperedge, all four types of preset hard constraints are checked sequentially. The four types of hard constraints are time window constraints, transportation capacity constraints, policy compliance constraints, and total carbon emission threshold constraints. The path extension is only allowed to take effect when all four types of checks pass.

[0069] The detailed judgment logic for the four types of hard constraint verification is as follows: The time window constraint check synchronously determines whether the departure time of the superedge matches the allowed departure time window of the starting spatiotemporal node and whether the arrival time matches the allowed arrival time window of the target spatiotemporal node. If any one of them is not satisfied, the current extension is discarded. The transportation capacity constraint check determines whether the current cargo volume of the path plus the transportation volume allocated to the superedge exceeds the maximum capacity limit of the superedge. If it exceeds the limit, it is discarded. The policy compliance constraint check superimposes the current superedge policy attribute with the code of the policy area traversed by the path to match the prohibition rules. If the prohibition is triggered, it is discarded. The total carbon emission threshold check determines whether the cumulative carbon emission of the path plus the current carbon emission value of the superedge exceeds the enterprise's preset total carbon emission threshold. If it exceeds the threshold, it is discarded.

[0070] After all single expansion verifications pass, new labels are generated and added to the label set of the second spatiotemporal node based on the cumulative indicators of the five categories of the original labels plus the incremental values ​​corresponding to the current superedge. After the new labels are added to the database, Pareto dominance pruning is immediately performed on the entire label set of the node to delete inferior path prefix labels that are completely dominated by other labels, further compressing the size of the label set and avoiding an explosive growth in the number of labels.

[0071] The Pareto dominance pruning logic includes: if the cumulative time, cumulative cost, cumulative carbon emissions, and occupied capacity of label A are all less than or equal to those of label B, and at least one of these indicators is strictly smaller, then label B is completely dominated by label A. The path prefix corresponding to label B has no optimization value and is directly deleted from the set. Only non-dominated labels are retained to participate in subsequent path expansion.

[0072] After traversing all spatiotemporal nodes within the spatiotemporal probabilistic hypergraph to complete the entire process of path prefix expansion, constraint verification, and label pruning, the complete path prefixes corresponding to the valid non-dominated labels remaining in all nodes are summarized and uniformly output as a set of candidate path prefixes for use in the next stage of robust bundle search.

[0073] Explanatoryly, traditional path search algorithms place constraint verification after the path is fully generated, resulting in a large number of invalid complete paths that violate hard rules, wasting a lot of computing power. In this embodiment 1, constraint filtering and pruning are completed in the early expansion stage before the path prefix is ​​fully generated, eliminating invalid branches from the source, significantly reducing the computational load of subsequent search and multi-objective evaluation, and improving the system response speed in high-concurrency path planning request scenarios. For example, in the scenario of cross-border transportation of high-value export goods, the enterprise presets a total carbon emission threshold of 100t. The cumulative carbon emission of a certain path prefix has reached 98t. If the carbon emission of the superedge to be expanded is 5t, the carbon emission constraint verification will fail, and the path expansion branch will be directly discarded and will not enter the subsequent complete path generation process, saving computing resources.

[0074] In one possible implementation, the process of generating a candidate path prefix set by filtering path prefixes that satisfy preset hard constraints based on a spatiotemporal probabilistic hypergraph and using a multi-constraint label setting algorithm is as follows: A label set is set for each spatiotemporal node in the spatiotemporal probabilistic hypergraph. Each label in the label set contains the cumulative time, cumulative cost, cumulative carbon emissions, occupied capacity, and policy area codes traversed from the source node to the corresponding spatiotemporal node. When expanding from the first spatiotemporal node to the second spatiotemporal node along the first hyperedge, constraint verification is performed according to preset hard constraints. The new labels generated by the expansion that passes the constraint verification are added to the label set of the second spatiotemporal node, and Pareto dominance pruning is performed on the label set of the second spatiotemporal node. All spatiotemporal nodes in the spatiotemporal probabilistic hypergraph are traversed, and a candidate path prefix set is output. The preset hard constraints include time window constraints, transportation capacity constraints, policy compliance constraints, and total carbon emission threshold constraints.

[0075] Specifically, the implementation process of the Multi-Constraint Labeling Algorithm (MC-LSA) is as follows: Step 41: Initialize the label set of all spatiotemporal nodes. The global source node initially contains only one blank label. The cumulative time, cumulative cost, cumulative carbon emissions, and occupied capacity in the blank label are all set to 0, and the set of policy area codes that have been passed through is empty. The initial label set of all other spatiotemporal nodes is empty.

[0076] Step 42: Traverse all spatiotemporal nodes of the spatiotemporal probability hypergraph, extract each label to be expanded in the node in turn, and read the five types of cumulative indicators inside the label.

[0077] Step 43: Traverse all downstream associated hyperedges of the current spatiotemporal node, and sequentially perform path extension operations from the current first spatiotemporal node along the first hyperedge to the second spatiotemporal node, while simultaneously performing four types of hard constraint checks, specifically including: The time window constraint verification is as follows: verify whether the departure time of the superedge falls within the allowed departure time window of the first spatiotemporal node and whether the expected arrival time of the superedge falls within the allowed arrival time window of the second spatiotemporal node. If either condition is not met, the current expansion is terminated.

[0078] The transportation capacity constraint verification is as follows: the occupied cargo capacity within the label is added to the current cargo volume carried by the superedge to determine whether it exceeds the maximum transportation capacity limit of the superedge. If it exceeds the limit, the current expansion is terminated.

[0079] Policy compliance constraint verification specifically involves merging the policy region code currently bound to the superedge with the set of policy region codes already passed through within the label, matching it against the system's stored prohibition rule base, and terminating this extension if any prohibition rule is triggered.

[0080] The carbon emission total threshold constraint verification is as follows: the cumulative carbon emission of the tag is added to the current super-edge basic carbon emission to determine whether it exceeds the enterprise's preset global carbon emission threshold. If it exceeds the threshold, the current expansion is terminated.

[0081] Step 44: After all four types of constraints have passed the verification, a new label is generated by superimposing the five cumulative indicators of the original label with the increment corresponding to the current superedge, and the new label is written into the label set of the second spatiotemporal node.

[0082] Step 45: Perform Pareto dominance pruning on the updated label set of the second spatiotemporal node, traverse all labels in the set, and delete inferior labels that are completely dominated by other labels. Step 46: Repeat steps 42-45 to complete the path prefix expansion operation for all spatiotemporal nodes and all hyperedges.

[0083] Step 47: Summarize the complete path prefixes corresponding to all non-dominated labels remaining in all spatiotemporal nodes, and output them as a set of candidate path prefixes.

[0084] Explanatory terms: Cumulative time within the label refers to the total transportation time from the origin to the current node of the path prefix; Cumulative cost refers to the total basic transportation cost of the entire journey; Cumulative carbon emissions refer to the total carbon emissions of the entire journey; Occupied capacity refers to the total weight / volume of goods carried by the path; The set of codes for policy areas traversed records the unique codes of all provinces, cities, border ports, and controlled areas through which the path passes; The four types of hard constraints correspond to the four rigid business rules of enterprises: logistics scheduling timeliness control, cargo space capacity control, cross-border policy compliance control, and dual carbon emission reduction control. Any path plan must meet all four types of constraints simultaneously to be implemented; Pareto dominance pruning is used to delete inferior path prefixes with no optimization value, control the size of the label set data, and avoid computing power explosion during path expansion.

[0085] It should be noted that traditional path search algorithms place constraint verification after the complete path generation is completed, which will generate a large number of invalid complete paths that violate hard control rules and consume a lot of computing resources. In this embodiment 1, constraint verification and pruning are performed in the early stage before the path prefix has been fully expanded, eliminating all illegal branches from the source, greatly reducing the computation space of subsequent robust bundle search, and improving the system response speed in high-concurrency path planning scenarios.

[0086] Specifically, by binding a set of labels containing five types of cumulative constraint indicators to each spatiotemporal node, four types of enterprise hard control rules are simultaneously verified during each hyperedge expansion of the path prefix. Only all compliant path prefixes are retained to participate in subsequent complete path expansion. At the same time, inferior prefixes with no optimization value are eliminated through Pareto dominance pruning. This ensures that all paths in the final output naturally meet the requirements of time, capacity, policy, and carbon emission control, and there are no non-compliant solutions that cannot be implemented. It also effectively reduces the number of path branches, balances the completeness of the solution and the consumption of computing power, and provides a concise and compliant candidate path prefix input for downstream robust bundle search. From the algorithm level, this improves the overall operational efficiency and feasibility of the path planning system.

[0087] Step 5: Based on the candidate path prefix set and the reliability weight vector of each hyperedge, perform robust bundle search to obtain the candidate path set. Specifically, based on the filtered effective path prefixes, an adaptive bundle width adjustment robust bundle search algorithm is used to complete the full path expansion, balancing path completeness and computational efficiency. For high-risk path prefixes, the bundle width is automatically increased to fully explore various feasible alternative detour schemes and improve the resilience of the supply chain in responding to failures.

[0088] Explained, robust bundle search is a hierarchical iterative path expansion algorithm. The whole process is executed layer by layer according to the jump level from the source node to the target node. Each level corresponds to one hyperedge expansion operation. Each level independently completes the entire process of comprehensive risk score calculation, risk benchmark value determination, bundle width adaptive adjustment, and path prefix expansion screening.

[0089] The first step in each layer of expansion is to calculate the comprehensive risk score for each candidate path prefix within the current layer. The comprehensive risk score is obtained by weighted summation of all hyperedge delay probabilities and cancellation probabilities contained in the path prefix. The higher the comprehensive risk score, the greater the overall transportation performance risk of the path prefix and the higher the probability of delays or cancellations.

[0090] After all the comprehensive risk scores of all path prefixes at the current level have been calculated, the arithmetic mean of all scores is taken as the risk benchmark value for that level. The risk benchmark value is used to distinguish between high-risk path prefixes and regular risk path prefixes. The threshold for this distinction is the risk benchmark value multiplied by a preset multiple, which is a fixed parameter that can be configured by the system.

[0091] The adaptive beamwidth adjustment logic is divided into two branches. In the first branch, if the comprehensive risk score of the path prefix is ​​higher than the risk benchmark value multiplied by a preset multiple, it is determined to be a high-risk path prefix, and the exclusive extended beamwidth of the path prefix is ​​increased based on the preset initial beamwidth value. In the second branch, if the comprehensive risk score of the path prefix is ​​not higher than the risk benchmark value multiplied by a preset multiple, it is determined to be a normal robust path prefix, and the extended beamwidth is kept equal to the preset initial beamwidth value.

[0092] The calculation of bundle width expansion introduces a risk ratio coefficient, which is the ratio of the current path prefix comprehensive risk score to the hierarchical risk benchmark value. The bundle width expansion value is positively correlated with the risk ratio coefficient. The higher the path risk, the greater the bundle width expansion, which can expand more alternative downstream transportation sections and fully explore detour alternatives.

[0093] At each level, the path prefix is ​​extended downstream according to the adjusted bundle width. Only the new path prefix with the best overall performance within the bundle width limit is retained to enter the next level of iteration. This process is repeated layer by layer until the path prefix reaches the global target node. All complete paths that reach the target node are summarized to form a candidate path set, which is then used in subsequent multi-objective evaluation steps.

[0094] In this embodiment 1, the adaptive beamwidth mechanism expands the search range for high-risk paths while simultaneously controlling the computational overhead of low-risk paths, balancing solution completeness with computational power consumption. The output candidate path set contains a sufficient number of differentiated detour schemes, improving the fault redundancy capability of the final path set. For example, a typhoon warning leads to a significant increase in the probability of shipping overruns at Qingdao Port being cancelled. The comprehensive risk score of the prefix for this type of path far exceeds the hierarchical risk benchmark value. In this case, the beamwidth of the prefix for this path is expanded, and two alternative branches, Busan transit by sea and land via Suifenhe Port, are simultaneously expanded and included in the candidate path set, ensuring that there are multiple executable alternative complete paths.

[0095] In one possible implementation, the process of obtaining the candidate path set by performing robust bundle search based on the candidate path prefix set and the reliability weight vector of each hyperedge is as follows: In each layer of the robust bundle search, a comprehensive risk score is calculated for each candidate path prefix in the current layer. The comprehensive risk score is calculated based on the weighted sum of the delay probabilities and cancellation probabilities of each hyperedge contained in the candidate path prefix. A risk benchmark value for the current layer is obtained, where the risk benchmark value is the average of the comprehensive risk scores of all candidate path prefixes in the current layer. For candidate path prefixes whose comprehensive risk score is higher than the risk benchmark value multiplied by a predetermined multiple, the bundle width of the corresponding candidate path prefix is ​​increased according to a pre-set initial bundle width value. For candidate path prefixes whose comprehensive risk score is not higher than the risk benchmark value multiplied by a predetermined multiple, the bundle width is kept at the pre-set initial bundle width value.

[0096] Specifically, the implementation process of robust beam search hierarchical adaptive beam width adjustment is as follows: Step 51: Initialize the global beamwidth initial value B0, the risk assessment predetermined multiple K, and the risk enhancement coefficient λ. All three types of parameters are configurable hyperparameters.

[0097] Step 52: Perform path expansion iteratively in layers, with each layer corresponding to a path jumping to a hyperedge once; the following process is executed independently for each layer operation.

[0098] Step 521: Read all candidate path prefixes at the current level. For a single path prefix, traverse all its contained hyperedges, retrieve the delay probability and cancellation probability of each hyperedge, and calculate the comprehensive risk score. The calculation process for the comprehensive risk score is as follows:

[0099] in, For comprehensive risk scoring; For delay weighting coefficients; To remove the weighting coefficient.

[0100] Step 522: Summarize the comprehensive risk scores of all path prefixes at the current level and calculate the arithmetic mean as the risk benchmark value for the current level.

[0101] Step 523: Calculate the high-risk threshold by traversing each path prefix at the current level; the process of calculating the high-risk threshold is as follows:

[0102] in, This is the threshold for determining high risk. This is the risk benchmark value for the current level.

[0103] If the overall risk score of a path prefix is ​​greater than the high-risk threshold, it is determined to be a high-risk path prefix, and the risk ratio is calculated. The calculation process for the risk ratio is as follows:

[0104] in, This represents the risk ratio.

[0105] At this point, the path prefix extension bundle width is as follows:

[0106] in, Extend the bundle width for the path prefix.

[0107] If the overall risk score of the path prefix is ​​less than or equal to the high-risk determination threshold, it is determined to be a normal robust path prefix, and the extended bundle width is fixed at the initial value of the global bundle width.

[0108] Step 524: Based on the bundle width adjusted for each path prefix, perform path extension downstream over the superedge, retaining only the new path prefix with the best overall performance within the bundle width limit to enter the next iteration.

[0109] Step 53: Iterate layer by layer until the path prefix reaches the global target node, summarize all complete paths that reach the target node, and generate a candidate path set.

[0110] Interpretive, the higher the comprehensive risk score, the higher the overall delay and cancellation risk of the transportation segments included in the route prefix, and the worse the supply chain performance stability; the risk benchmark value is used to distinguish between high and low risk route prefixes within the tier; the risk judgment predetermined multiple K is the risk classification threshold, used to define high-risk routes that need to be widened; the global bundle width initial value is the basic expansion scale of the regular robust route, the risk enhancement coefficient λ controls the expansion range of the high-risk route bundle width, and the risk ratio characterizes the degree of deviation of the risk of a single route prefix from the average risk of the tier.

[0111] It should be noted that traditional beam search algorithms use a fixed and uniform beam width for all path prefixes, which makes it impossible to expand sufficient alternative detour schemes for high-risk paths. Once the main line is out of service, there are no feasible backup paths. In this embodiment 1, the beam width is automatically expanded for path prefixes with high comprehensive risk scores, which can fully explore various differentiated detour alternative schemes. At the same time, the basic beam width is maintained for low-risk paths to avoid unnecessary computational power consumption.

[0112] Specifically, by calculating the comprehensive risk score of path prefixes and the hierarchical risk benchmark value in a hierarchical manner, high-risk and conventional robust path prefixes are distinguished and the expansion bundle width is adaptively adjusted. For high-risk transportation links, the search range is proactively expanded to generate a sufficient number of differentiated alternative transportation branches, thereby improving the fault redundancy capability of the final candidate path set. When sudden events such as port congestion, extreme weather, and line failures occur, there are multiple switchable detour paths. At the same time, the basic bundle width is maintained for low-risk stable paths to control the overall search computation volume, balance the completeness of path solution with the efficiency of computing power consumption, and adapt to multimodal transport business scenarios with high concurrency and real-time scheduling.

[0113] Step 6: Utilize a multi-objective evaluation function to evaluate each candidate path in the candidate path set, obtaining the objective vector for each candidate path. Specifically, a standardized multi-objective vector is generated by performing a multi-dimensional quantitative evaluation on all complete candidate paths output by the robust bundle search. This vector simultaneously covers expected indicators, tail risk indicators, sustainability indicators, performance stability indicators, and fault redundancy indicators, providing a unified quantitative evaluation basis for subsequent Pareto hierarchical multi-objective optimization.

[0114] In one possible implementation, the objective vector for each candidate route includes expected transport time, expected transport cost, time risk value, cost risk value, total carbon emissions, reliability score, and route redundancy. Specifically, the objective vector for each candidate route is represented as follows:

[0115] in, The target vector for each candidate path; The expected transport time is obtained by summing the basic transport times of all superedges of the path and the time reliability weights, which represents the average total transport time of the path under normal disturbance scenarios. The expected transportation cost is obtained by summing the base freight rates of all super-edges of the path and the cost reliability weights, which represents the average total transportation cost of the path under normal disturbance scenarios. To determine the time risk value, a massive sample of transport time is generated by performing Monte Carlo random sampling on the probability distribution of all hyperedge delays and cancellations along the path. The percentile value corresponding to the preset confidence level η is then used to obtain the worst-case transport time of the path under extremely severe disturbance scenarios. The cost risk value is obtained by generating a massive sample of transportation costs based on Monte Carlo sampling and taking the percentile value corresponding to the confidence level η, which represents the worst total transportation cost of the path under extremely severe disturbance scenarios. The total carbon emissions are calculated by directly summing the basic carbon emissions of all super-edges along the path, representing the total carbon emissions of the entire freight transport process. The reliability score is obtained by multiplying the probabilities of on-time delivery of all hyperedges along the path, representing the overall probability of on-time delivery throughout the entire path. The path redundancy is obtained by dividing the number of replaceable nodes within the path by the total number of nodes in the path. Replaceable nodes are defined as backup terminals or ports that have completely identical functions and can be switched for customs clearance / loading and unloading within a preset spatial distance threshold. The higher the value, the stronger the path fault switching capability.

[0116] Interpretive metrics include expected transit time and expected transit cost as average path performance indicators, used to evaluate business performance under normal and stable scenarios; time risk value and cost risk value as tail extreme risk indicators, used to quantify the maximum loss under sudden adverse scenarios such as extreme weather and port congestion, and to manage tail risks in the supply chain; total carbon emissions are used to meet the needs of enterprises for low-carbon emission reduction and carbon quota management; reliability score intuitively represents the probability of on-time delivery of the path; and path redundancy quantifies the path's alternative fault tolerance capability in the face of node failures and line outages.

[0117] For example, expected transportation time and expected transportation cost are calculated based on the base duration and base cost of all super-edges of the path, plus reliability weights, representing the average transportation indicators of the path under normal disturbance scenarios; time risk value and cost risk value are quantitative indicators of extreme risks at the tail of the path, solved using the Monte Carlo sampling method, which involves a large number of random samples of the probability distribution of delays and cancellations of all super-edges of the path, and extracting the percentiles of the pre-set confidence level as risk values ​​to measure the worst-case timeliness and worst-case cost of the path under extremely adverse scenarios; total carbon emissions are the sum of the base carbon emissions of all super-edges of the path; reliability score is obtained by multiplying the on-time probabilities of all super-edges of the path, representing the comprehensive probability of on-time performance throughout the entire path; path redundancy is the ratio of the number of replaceable nodes within the path to the total number of nodes. Replaceable nodes are defined as backup nodes with completely identical functions that can be switched for customs clearance / loading and unloading within a preset spatial distance range. The higher the index value, the stronger the path's fault switching capability.

[0118] It should be noted that traditional route planning only outputs a single time or cost evaluation indicator, which cannot simultaneously take into account the diversified business management needs such as extreme risks, carbon emission reduction, performance reliability, and fault redundancy. In this 1st embodiment, the target vector simultaneously covers five major business evaluation dimensions: average indicators, extreme risks, sustainability indicators, performance stability, and fault tolerance, which fully matches the all-dimensional operation and management objectives of modern cross-border multimodal transport enterprises.

[0119] By constructing a target vector containing seven differentiated evaluation indicators, the comprehensive business value of each candidate path can be fully quantified from multiple dimensions such as timeliness, cost, extreme tail risk, carbon emissions, performance reliability, and fault redundancy. The seven indicators naturally have mutual trade-offs and conflicts, and no single path can achieve the optimal performance of all seven indicators at the same time. Therefore, it provides a standardized quantitative evaluation basis for subsequent Pareto hierarchical multi-objective equilibrium optimization. The final Pareto optimal primary and backup paths can adapt to various differentiated business preferences such as timeliness priority, cost priority, low carbon priority, and robust risk resistance, and meet diversified scheduling decision-making needs.

[0120] Step 7: According to the target vector of each candidate path, perform Pareto hierarchical screening on each candidate path in the candidate path set, output the Pareto optimal path set, and determine the main path and at least one alternative path from the Pareto optimal path set as the transportation path planning result of the target multimodal transport network.

[0121] Explanatoryly, in this embodiment 1, multi-objective Pareto optimization is performed based on multi-dimensional objective vectors to eliminate inferior solutions that are completely dominated by other paths, thereby obtaining a globally balanced Pareto optimal path set. The main path and multiple differentiated alternative paths are output simultaneously, providing multiple feasible transportation solutions for scheduling operations and enhancing the resilience of the supply chain.

[0122] The core logic of Pareto hierarchical screening is the Pareto dominance criterion, as follows: If all seven-dimensional target vector indicators of path A are better than or equal to those of path B, and at least one dimension indicator is strictly better than that of path B, then path B is completely dominated by path A, path B has no business optimization value, and is directly removed from the candidate set. The layered operation will peel away all dominated paths layer by layer, and finally retain all non-dominated paths to form a Pareto optimal path set. No two paths in the set have a mutual dominance relationship, and they correspond to differentiated business preferences such as timeliness priority, cost priority, low carbon priority, and robust risk priority.

[0123] For example, after the Pareto optimal path set is generated, a comprehensive balanced solution is selected from the set as the main path according to the business preset priority rules, and at least two differentiated solutions are selected as alternative paths. The main path is adapted to normal and stable transportation scenarios, taking into account four core indicators: timeliness, cost, carbon emissions, and risk. The alternative paths focus on three sub-categories of needs: extreme risk resistance, low-carbon emission reduction, and low-cost management. When a node or transportation segment along the main path fails, the scheduling system can switch to the corresponding alternative path with one click.

[0124] Optionally, after completing the Pareto screening, two auxiliary optimization processes are executed. The first process is main path sensitivity analysis and node failure scenario simulation, which traverses the key nodes and key hyperedges of the main path, simulates two types of failure scenarios: node failure and tightening of transportation segment time windows, re-solves for feasible alternatives, generates a main path resilience assessment report, and adjusts the output ranking of candidate paths according to the resilience level. The second process is full set path compliance verification, which inputs all Pareto optimal paths into the compliance verification engine, matches them with embargo lists, customs windows, road restrictions, and green channel rules, eliminates non-compliant paths, and adds green channel labels and a complete compliance explanation report.

[0125] It should be noted that in this embodiment 1, the final output of the main path and alternative paths is uniformly pushed to the user dispatch terminal or the automated guided vehicle execution equipment, and is directly used for multimodal transport dispatch system path decision-making and cargo navigation execution; at the same time, a path breakdown report is output simultaneously, including the transport time, cost, carbon emission details, sources of delay / cancellation risk, sensitive nodes and alternative detour schemes for each segment, providing complete and interpretable support; traditional path planning only outputs a single optimal path without differentiated alternative schemes, and once the main line is blocked, the dispatch system has no available schemes; in this embodiment 1, the Pareto optimal main and alternative path combinations are output, covering multiple business preferences and fault response scenarios, which greatly improves the stability and fault tolerance of the cross-border multimodal transport supply chain; for example, the Pareto optimal set for Chongqing to Duisburg land bridge transport includes three types of paths: the time-priority alternative via Almaty port, the robust option of retaining the original Khorgos route with the Hami alternative detour scheme, and the low-carbon option of adding railway sections. The robust option is selected as the main path, and the time-priority and low-carbon options are output as two alternative paths and delivered to the dispatch terminal simultaneously.

[0126] It is worth noting that, in terms of robustness and performance design, this embodiment 1 adopts a local recalculation strategy. Specifically, when a local event occurs, such as a single port closure or border crossing restrictions, only the candidate and beam search are recalculated for the affected subgraph to avoid network-wide jitter. A decay weight is applied to data freshness to ensure that predictions and weights are based on the latest information. For high-concurrency requests, OD-level caching and result reuse are enabled, and expiration times are set for candidates and weights on high-frequency lines to ensure accuracy. Reliability prediction and beam search are executed in parallel in a GPU / CPU heterogeneous environment to compress response time. Anomaly detection triggers conservative parameters, such as increasing the weight of delay probability or cancellation probability or reducing beam width, to prevent extreme inputs from leading to overly aggressive paths.

[0127] Method effectiveness verification instructions: Verification Scenario 1: Taking the Chongqing-Horgos-Duisburg land corridor as the target multimodal transport network, the multimodal transport route planning method based on the spatiotemporal probabilistic hypergraph described above is used for route planning. The results show that: The Khorgos loading and unloading efficiency decreased by 20% and a strong wind warning was issued. The prediction module increased the probability of delay to 0.31 and the probability of cancellation to 0.06. The beam search output three types of solutions: timeliness priority, robustness priority, and low-carbon priority. Among them, timeliness priority specifically means changing the route to Almaty, with a timeliness increase of 8%, cost increase of 4.9%, and CO2 increase of 3%; robustness priority specifically means keeping the original route and configuring Hami as a backup detour, with a timeliness increase of 5%, cost increase of 2.5%, and a VaR_95 timeliness reduction of 18 hours; low-carbon priority specifically means adding railway sections, with CO2 reduction of 12%, timeliness increase of 6%, and cost increase of 6.4%. The solution explanation report found that the delay risk mainly comes from the decline in border inspection and loading and unloading efficiency.

[0128] Verification Scenario 2: Taking the Qingdao-Vladivostok-Moscow sea-rail intermodal transport network as the target multimodal transport network, the multimodal transport route planning method based on the spatiotemporal probabilistic hypergraph described above is used for route planning. The results show that: The typhoon has compressed the berthing window, increasing the probability of cancellation to 0.12. The berthing width has been automatically increased, and the main route has been changed to transit through Busan (timeliness +9%, cost +5%, CO2 +4%). The alternative is: delayed departure + Khabarovsk Railway acceleration plan, with a backup land port detour.

[0129] Verification Scenario 3: A multimodal transport network targeting high-value goods combining air and rail was established. The multimodal transport route planning method based on spatiotemporal probabilistic hypergraphs described above was used for route planning. The results showed that: The insurance requires a timeliness confidence level of ≥0.9. Under VaR constraints, edges with high cancellation probability are removed, and dual backup paths are output to ensure that the total timeliness does not exceed the threshold under a 90% confidence condition.

[0130] The multimodal transport route planning method described in Embodiment 1 overcomes the limitations of unilateral deterministic graphs by constructing a spatiotemporal probabilistic hypergraph. This hypergraph model explicitly carries delay / cancellation distributions and multi-segment intermodal constraints. Secondly, reliability-driven dynamic weights enable proactive risk perception in route decisions, avoiding reactive route changes. Furthermore, robust multi-objective bundle search, while satisfying hard constraints, outputs primary and backup routes in a Pareto hierarchy and uses VaR to constrain tail risks, reducing rework rates. In addition, the integrated interpretable and compliant output directly provides decomposition and sensitivity analysis, improving customer and audit acceptance. The engineering-ready performance and resilience design ensures stable response even in high-concurrency, rapidly changing scenarios.

[0131] In this embodiment 1, the container multimodal transport route algorithm based on spatiotemporal probabilistic hypergraphs and robust multi-objective programming achieves robust, interpretable, and compliant route decisions in complex dynamic environments through the synergy of uncertainty modeling, reliability prediction, and robust bundle search. This provides an implementable technical solution for multiple modes such as rail + road, sea-rail intermodal transport, and air-rail intermodal transport. Optionally, the solution space coverage can be expanded by introducing hyperedge-level simulation annealing or genetic algorithms to complement bundle search; carbon trading / quotas can be incorporated into constraints and incentives to form joint optimization of carbon costs and routes; the reliability model can be shared among multiple carriers by using federated learning to protect data privacy while improving prediction accuracy; route explanations and contingency plans can be automatically generated by combining generative models to shorten communication time; and the global computational pressure can be further reduced by introducing hierarchical partitioning solutions on multi-regional networks.

[0132] Example 2 This embodiment 2 provides a multimodal transport route planning system based on a spatiotemporal probabilistic hypergraph, used to implement the multimodal transport route planning method based on a spatiotemporal probabilistic hypergraph described in embodiment 1 above, as shown in the attached figure. Figure 3 As shown, the multimodal transport route planning system based on spatiotemporal probabilistic hypergraph includes a spatiotemporal data acquisition module, a graph construction module, a weight calculation module, a prefix set generation module, a candidate route search module, a candidate route evaluation module, and a candidate route screening module.

[0133] The spatiotemporal data acquisition module is used to acquire spatiotemporal data of the target multimodal transport network; the spatiotemporal data includes static structural data and dynamic environmental data of the multimodal transport network.

[0134] The graph construction module is used to construct a spatiotemporal probabilistic hypergraph based on the spatiotemporal data of the target multimodal transport network.

[0135] The weight calculation module is used to predict the delay probability and cancellation probability of each hyperedge in the spatiotemporal probability hypergraph using a temporal graph neural network, and calculate the reliability weight vector of each hyperedge.

[0136] The prefix set generation module is used to filter path prefixes that meet preset hard constraints based on the spatiotemporal probabilistic hypergraph and using a multi-constraint label setting algorithm to generate a candidate path prefix set.

[0137] The candidate path search module is used to perform robust bundle search based on the candidate path prefix set and the reliability weight vector of each hyperedge to obtain the candidate path set.

[0138] The candidate path evaluation module is used to evaluate each candidate path in the candidate path set using a multi-objective evaluation function, and obtain the objective vector of each candidate path.

[0139] The candidate route selection module is used to perform Pareto hierarchical selection on each candidate route in the candidate route set according to the target vector of each candidate route, output the Pareto optimal route set, and determine the main route and at least one alternative route from the Pareto optimal route set as the transportation route planning result of the target multimodal transport network.

[0140] It should be noted that all relevant content of each step involved in the multimodal transport route planning method based on spatiotemporal probabilistic hypergraph described in Embodiment 1 can be referenced to the functional description of the corresponding functional module of the multimodal transport route planning system based on spatiotemporal probabilistic hypergraph in Embodiment 2, and will not be repeated here.

[0141] The module division in this embodiment 2 is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods. In addition, each functional module in this embodiment 2 can be integrated into a processor, exist as a separate physical entity, or have two or more modules integrated into one module. The integrated module can be implemented in hardware or as a software functional module.

[0142] Example 3 As attached Figure 4 As shown, this embodiment 3 provides an electronic device, including: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the multimodal transport route planning method based on spatiotemporal probability hypergraph; or, the processor for executing the computer program to implement the functions of each module in the above-mentioned multimodal transport route planning system based on spatiotemporal probability hypergraph.

[0143] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a preset function, the instruction segments describing the execution process of the computer program in the electronic device.

[0144] The electronic device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above are examples of electronic devices and do not constitute a limitation on the electronic device. It may include more components than described above, or combine certain components, or different components. For example, the electronic device may also include a communication interface, input / output devices, network access devices, and a bus.

[0145] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (OPGs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor, or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the electronic device via various communication interfaces and lines.

[0146] The memory can be used to store the computer program and / or module. The processor implements various functions of the electronic device by running or executing the computer program and / or module stored in the memory and by calling the data stored in the memory.

[0147] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function (such as sound playback, image playback, etc.). The data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards, secure digital cards, flash memory cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0148] Example 4 This embodiment 4 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the multimodal transport route planning method based on a spatiotemporal probabilistic hypergraph.

[0149] If the modules / units of the multimodal transport route planning system based on spatiotemporal probability hypergraph are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0150] Based on this understanding, the present invention can implement all or part of the processes in the above-mentioned multimodal transport route planning method based on spatiotemporal probabilistic hypergraphs. This can also be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above-mentioned multimodal transport route planning method based on spatiotemporal probabilistic hypergraphs. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or preset intermediate forms, etc.

[0151] The computer-readable storage medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0152] Example 5 This embodiment 5 provides a computer product, which includes a computer program stored in a computer-readable storage medium. The processor of the electronic device reads the computer program from the computer-readable storage medium and executes the computer program, so that the electronic device can execute the multimodal transport route planning method based on spatiotemporal probabilistic hypergraph described in embodiment 1, which will not be described again here.

[0153] It should be noted that those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above methods.

[0154] The multimodal transport route planning method based on spatiotemporal probabilistic hypergraph described in this invention obtains complete spatiotemporal data of the target multimodal transport network as the modeling basis. This method can fully cover the static road network structure and real-time dynamic disturbance information, providing comprehensive and real-time data support for subsequent probability prediction and route solving. It avoids the model distortion problem caused by traditional static graphs relying solely on fixed basic data. By constructing a spatiotemporal probabilistic hypergraph containing a set of nodes, a set of hyperedges with probabilistic attributes, and a time extension layer, it can uniformly abstract multimodal transport sub-paths into hyperedges. At the same time, it uses time slice nodes and waiting connection edges to fully characterize the transfer waiting and cross-modal time sequence connection logic, intuitively quantifying the three types of uncertainty probabilities of punctuality, delay, and cancellation for each transport link. This solves the technical defects of traditional graph models that cannot simultaneously express multimodal transport and time sequence constraints and lack risk probability modeling.

[0155] Secondly, by using a temporal graph neural network to predict the delay and cancellation probabilities of each hyperedge based on full spatiotemporal data, and generating a reliability weight vector containing three dimensions: time, cost, and carbon emissions, transportation risks can be quantified and integrated into the path weight calculation. This enables proactive perception of path risks, allowing the path solving process to prioritize avoiding transportation segments with high disturbance and high default risk, thus preventing performance losses caused by passive route rerouting after transportation failures. Furthermore, by using a multi-constraint label setting algorithm to simultaneously verify four types of hard constraints—time window, transportation capacity, policy compliance, and total carbon emissions—during the path prefix expansion stage, and by using Pareto dominance pruning to eliminate invalid path branches, the computational space of subsequent robust bundle search is significantly reduced, improving the overall path solving response speed. At the same time, it ensures that all output paths meet the hard control rules of the business and that there are no unenforceable or non-compliant solutions.

[0156] Furthermore, through a robust bundle search mechanism that adaptively adjusts the bundle width, the bundle width is expanded for path prefixes with high comprehensive risk scores while maintaining the basic bundle width for low-risk paths. This not only fully explores alternative detour paths in high-risk scenarios to ensure supply chain resilience but also controls computational redundancy for low-risk paths, balancing solution completeness and computational efficiency. A multi-objective evaluation function generates a multi-dimensional objective vector that includes expected transportation time, expected transportation cost, time risk value, cost risk value, total carbon emissions, reliability score, and path redundancy. This simultaneously covers optimization objectives such as timeliness, cost, carbon emissions, tail risk, fulfillment stability, and fault tolerance, overcoming the limitation of traditional single-objective path optimization that can only weigh a single indicator. Pareto hierarchical screening yields a globally optimal equilibrium Pareto optimal path set, from which the main path and at least one alternative path are simultaneously output. This provides multiple differentiated and robust transportation solutions for multimodal transport scheduling. When a main path node or transportation segment fails, alternative paths can be quickly switched, significantly improving the resilience of the cross-border supply chain.

[0157] The method described in this invention can be deployed on various hardware platforms such as multimodal transport scheduling platforms, port terminal management terminals, and cloud logistics computing servers. It is compatible with various multimodal transport business scenarios such as sea-rail intermodal transport, air-rail intermodal transport, land bridge cross-border transport, and parallel customs clearance at multiple ports, providing a complete and feasible technical solution for intelligent route scheduling in the complex and dynamic environment of the global supply chain.

[0158] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

Claims

1. A multimodal transport route planning method based on spatiotemporal probabilistic hypergraph, characterized in that, include: Acquire spatiotemporal data of the target multimodal transport network; wherein, the spatiotemporal data includes static structural data and dynamic environmental data of the multimodal transport network; Based on the spatiotemporal data of the target multimodal transport network, a spatiotemporal probabilistic hypergraph is constructed; Using a temporal graph neural network, the delay probability and cancellation probability of each hyperedge in the spatiotemporal probability hypergraph are predicted, and the reliability weight vector of each hyperedge is calculated. Based on the spatiotemporal probabilistic hypergraph and using a multi-constraint label setting algorithm, path prefixes that meet the preset hard constraints are selected to generate a candidate path prefix set. Based on the candidate path prefix set and the reliability weight vector of each hyperedge, a robust bundle search is performed to obtain the candidate path set; By using a multi-objective evaluation function, each candidate path in the candidate path set is evaluated to obtain the objective vector of each candidate path; Based on the target vector of each candidate path, Pareto hierarchical screening is performed on each candidate path in the candidate path set, outputting the Pareto optimal path set, and determining the main path and at least one alternative path from the Pareto optimal path set as the transportation path planning result of the target multimodal transport network.

2. The multimodal transport route planning method based on spatiotemporal probabilistic hypergraph according to claim 1, characterized in that, The spatiotemporal probabilistic hypergraph consists of a node set, a hyperedge set, and a time extension layer. Each hyperedge in the hyperedge set represents an independent transportation sub-path and is associated with a set of probability attributes, including the hyperedge's on-time probability, delay probability, and cancellation probability. The time extension layer is used to replicate each node along the time axis into multiple time-slice nodes and configure waiting connection edges for each time-slice node.

3. The multimodal transport route planning method based on spatiotemporal probabilistic hypergraph according to claim 1, characterized in that, The reliability weight vector for each hyperedge includes time reliability weight, cost reliability weight, and carbon emission reliability weight.

4. The multimodal transport route planning method based on spatiotemporal probabilistic hypergraph according to claim 1, characterized in that, The process of generating a candidate path prefix set by using a spatiotemporal probabilistic hypergraph and a multi-constraint label setting algorithm to filter path prefixes that meet preset hard constraints is as follows: A label set is set for each spatiotemporal node in the spatiotemporal probabilistic hypergraph. Each label in the label set contains the cumulative time, cumulative cost, cumulative carbon emissions, occupied capacity, and policy area codes traversed from the source node to the corresponding spatiotemporal node. When expanding from the first spatiotemporal node to the second spatiotemporal node along the first hyperedge, constraint verification is performed according to preset hard constraints. The new labels generated by the expansion that passes the constraint verification are added to the label set of the second spatiotemporal node, and Pareto dominance pruning is performed on the label set of the second spatiotemporal node. All spatiotemporal nodes in the spatiotemporal probabilistic hypergraph are traversed, and a candidate path prefix set is output. The preset hard constraints include time window constraints, transportation capacity constraints, policy compliance constraints, and total carbon emission threshold constraints.

5. The multimodal transport route planning method based on spatiotemporal probabilistic hypergraph according to claim 1, characterized in that, The process of obtaining the candidate path set by performing robust bundle search based on the candidate path prefix set and the reliability weight vector of each hyperedge is as follows: In each layer of the robust bundle search, a comprehensive risk score is calculated for each candidate path prefix in the current layer; wherein the comprehensive risk score is calculated as a weighted sum of the delay probabilities and cancellation probabilities of each hyperedge contained in the candidate path prefix; Obtain the risk baseline value for the current layer; where the risk baseline value is the average of the comprehensive risk scores of all candidate path prefixes in the current layer; For candidate path prefixes whose comprehensive risk score is higher than the risk benchmark value multiplied by a predetermined multiple, the extension beamwidth of the corresponding candidate path prefix will be increased according to the preset initial beamwidth value. For candidate path prefixes whose comprehensive risk score is not higher than the risk benchmark value multiplied by a predetermined multiple, the extended beamwidth is kept at the preset initial beamwidth value.

6. The multimodal transport route planning method based on spatiotemporal probabilistic hypergraph according to claim 1, characterized in that, The objective vector for each candidate route includes expected transport time, expected transport cost, time risk value, cost risk value, total carbon emissions, reliability score, and route redundancy.

7. A multimodal transport route planning system based on spatiotemporal probabilistic hypergraph, characterized in that, The method for implementing the multimodal transport route planning method based on spatiotemporal probabilistic hypergraph as described in any one of claims 1-6 includes: The spatiotemporal data acquisition module is used to acquire spatiotemporal data of the target multimodal transport network; wherein, the spatiotemporal data includes static structural data and dynamic environmental data of the multimodal transport network; The graph construction module is used to construct a spatiotemporal probabilistic hypergraph based on the spatiotemporal data of the target multimodal transport network. The weight calculation module is used to predict the delay probability and cancellation probability of each hyperedge in the spatiotemporal probability hypergraph using a temporal graph neural network, and calculate the reliability weight vector of each hyperedge. The prefix set generation module is used to filter path prefixes that meet preset hard constraints based on the spatiotemporal probabilistic hypergraph and using a multi-constraint label setting algorithm to generate a candidate path prefix set. The candidate path search module is used to perform robust bundle search based on the candidate path prefix set and the reliability weight vector of each hyperedge to obtain the candidate path set; The candidate path evaluation module is used to evaluate each candidate path in the candidate path set using a multi-objective evaluation function, and obtain the objective vector of each candidate path. The candidate route selection module is used to perform Pareto hierarchical selection on each candidate route in the candidate route set according to the target vector of each candidate route, output the Pareto optimal route set, and determine the main route and at least one alternative route from the Pareto optimal route set as the transportation route planning result of the target multimodal transport network.

8. An electronic device, characterized in that, include: A processor is used to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, performs the multimodal transport route planning method based on a spatiotemporal probabilistic hypergraph as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the multimodal transport route planning method based on spatiotemporal probabilistic hypergraph as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the multimodal transport route planning method based on a spatiotemporal probabilistic hypergraph as described in any one of claims 1-6.

Images