Cargo transportation scheduling method and system based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent logistics scheduling, and discloses a cargo transportation scheduling method and system based on artificial intelligence. The method comprises the following steps: performing standardized formatting preprocessing on original transportation task data to obtain standard task description data; task space-time coding is executed, and task space-time coding data containing time sensitivity weights and path conflict marks are generated; inputting the coded data into a trained deep neural network scheduling optimization model, and outputting a group of preliminary scheduling schemes; inputting the initial scheme into a multi-scheme game decision module, and selecting an optimal scheduling scheme based on a preset game rule and a real-time road condition information flow; and performing dynamic environment deduction on the optimal scheme, and generating an execution deduction path with a time label and a resource label. According to the method, constraint perception is enhanced through space-time coding, dynamic preferential selection is realized through game decision, and the accuracy and adaptability of cargo transportation scheduling are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics scheduling technology, specifically to a cargo transportation scheduling method and system based on artificial intelligence. Background Technology

[0002] In the field of freight transportation scheduling, traditional methods often rely on fixed rules or linear programming algorithms for route planning and time scheduling. These methods are usually based on static data and ignore the dynamic changes in transportation tasks. When dealing with complex scheduling problems, existing technologies often consider cargo attributes, path nodes, and time constraints separately, which can easily lead to conflicts or delays in the actual implementation of scheduling schemes.

[0003] Existing technical solutions have shortcomings, primarily in their inability to effectively quantify the time urgency of transportation tasks and the lack of pre-emptive identification of route resource competition. Conventional encoding methods merely convert spatiotemporal information into numerical vectors without injecting scheduling semantics, making it difficult for subsequent optimization models to perceive key constraints. Furthermore, scheduling decisions often employ single-objective optimization, neglecting the game-theoretic relationships between multiple options in dynamic environments, with insufficient integration of real-time traffic information, resulting in poor adaptability of the solutions. While existing deep learning methods can handle large amounts of data, their input features often lack explicit conflict markers and time-sensitivity indicators, leading to unstable performance of the model's output solutions in complex scenarios.

[0004] This invention addresses how to explicitly express time sensitivity and path conflict during the encoding phase to enhance the model's constraint awareness. It also addresses how to select the optimal scheduling scheme from multiple preliminary schemes using a game theory mechanism, incorporating real-time dynamic information, thereby improving decision robustness. Solving these problems requires novel data encoding methods and decision-making frameworks that go beyond the simple processing of spatiotemporal information by conventional techniques. Summary of the Invention

[0005] The purpose of this invention is to provide a cargo transportation scheduling method and system based on artificial intelligence to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a cargo transportation scheduling method based on artificial intelligence, the method comprising:

[0007] The raw transportation task data, which includes cargo attributes, route nodes, and scheduled time, is standardized and formatted preprocessed to obtain standard task description data.

[0008] Perform task spatiotemporal coding on the standard task description data to generate task spatiotemporal coded data that includes time sensitivity weights and path conflict markers;

[0009] The task spatiotemporal encoded data is input into a trained deep neural network scheduling optimization model, and the deep neural network scheduling optimization model outputs a set of preliminary scheduling schemes corresponding to the original transportation task data.

[0010] The set of preliminary scheduling schemes is input into a multi-scheme game decision module, and an optimal scheduling scheme is selected from the set of preliminary scheduling schemes based on the preset game rules and real-time traffic information flow.

[0011] The optimal scheduling scheme is dynamically simulated to generate an execution simulation path with time and resource tags.

[0012] Preferably, performing task spatiotemporal coding on the standard task description data includes the following steps:

[0013] The predetermined time in the standard task description data is parsed, and a time sensitivity weight is calculated and assigned according to the urgency of the time.

[0014] Identify path nodes in the standard task description data and compare their overlap with historical task paths, marking overlapping path segments as path conflict markers.

[0015] The time sensitivity weight is fused with the path conflict marker, and a unique task identifier corresponding to the original transportation task data is added to generate structured coded data.

[0016] The structured coded data is sorted in descending order according to the time sensitivity weight to generate the task spatiotemporal coded data.

[0017] Preferably, inputting the task spatiotemporal encoded data into a trained deep neural network scheduling optimization model includes the following steps:

[0018] The initial encoding layer of the deep neural network scheduling optimization model receives the task spatiotemporal encoded data;

[0019] The output of the initial encoding layer enters a spatiotemporal attention allocation layer, which dynamically allocates and calculates attention based on the time sensitivity weight and the path conflict marker.

[0020] The output of the spatiotemporal attention allocation layer is input into a multi-objective optimization layer, which simultaneously calculates a combination of objective functions that minimize transportation costs, minimize time delays, and maximize path utilization.

[0021] The output of the multi-objective optimization layer is decoded by a scheme generation layer, which outputs a set of preliminary scheduling schemes including specific vehicle allocation, path order, and time nodes.

[0022] Preferably, inputting the initial scheduling scheme into a multi-scheme game decision module includes the following steps:

[0023] The multi-scheme game decision module receives real-time traffic information streams, which include dynamically updated road segment traffic speeds and traffic event data.

[0024] By treating each preliminary scheduling scheme as a game participant, a non-cooperative game model is established with comprehensive transportation benefits as the payoff function.

[0025] In the non-cooperative game model, the Nash equilibrium point of each preliminary scheduling scheme is calculated iteratively using the real-time traffic information flow;

[0026] The preliminary scheduling scheme associated with the Nash equilibrium point corresponding to the highest payoff function value is selected as the optimal scheduling scheme.

[0027] Preferably, the establishment of the non-cooperative game model with comprehensive transportation benefits as the payoff function includes the following steps:

[0028] The comprehensive transportation benefit is defined as the weighted sum of the economic benefit function and the time benefit function;

[0029] The economic benefit function is calculated based on the estimated fuel consumption, vehicle wear and tear, and route costs of the preliminary scheduling plan.

[0030] The time efficiency function is calculated based on the difference between the estimated completion time and the predetermined time after the application of the real-time traffic information stream in the preliminary scheduling scheme.

[0031] A virtual game strategy space is assigned to each preliminary scheduling scheme, and the game strategy space contains minor adjustment operations on path nodes;

[0032] Multiple rounds of strategy selection and payoff calculation are performed within the game strategy space until the strategy selection of all preliminary scheduling schemes tends to stabilize.

[0033] Preferably, the dynamic environmental simulation of the optimal scheduling scheme includes the following steps:

[0034] Based on the optimal scheduling scheme, extract its planned path sequence and planned time sequence;

[0035] Construct a discrete event deduction environment with time evolution as the axis;

[0036] In the discrete event simulation environment, transportation events in the optimal scheduling scheme are gradually injected according to the planned time series.

[0037] At each simulation time step, a randomly generated external disturbance event is introduced, and the state evolution of the transportation event under the disturbance is calculated;

[0038] The actual path offset and actual time consumption resulting from the state evolution are recorded to form the execution deduction path.

[0039] Preferably, constructing a discrete event deduction environment with time evolution as the axis includes the following steps:

[0040] Define the global clock and time progression mechanism of the discrete event deduction environment;

[0041] Register multiple types of transportation entity models in the discrete event simulation environment, wherein the transportation entity models include at least vehicle models and cargo models;

[0042] Register various types of environmental event generators in the discrete event simulation environment. These environmental event generators can generate events that simulate traffic congestion and weather changes.

[0043] A response rule base for the transportation entity model to the environmental event generator is established, and the response rule base defines the logic of entity state changes.

[0044] Preferably, the method further includes the step of:

[0045] During the actual execution of the simulated execution path, execution process monitoring data is collected;

[0046] The execution process monitoring data is compared with the expected data of the execution simulation path to generate execution deviation data;

[0047] The execution deviation data is fed back into the training process of the deep neural network scheduling optimization model to update the network parameters of the deep neural network scheduling optimization model;

[0048] The deep neural network scheduling optimization model, after updating the network parameters, is used to process the subsequent arriving transportation task data.

[0049] Preferably, the process of feeding back the execution deviation data to the training process of the deep neural network scheduling optimization model includes:

[0050] The execution bias data is converted into the loss function increment required for model training;

[0051] The new training samples containing the increment of the loss function are merged with the historical training samples to form an incremental training dataset.

[0052] The deep neural network scheduling optimization model is trained using an incremental learning approach and the incremental training dataset.

[0053] Freeze the parameters of the basic feature extraction layer in the deep neural network scheduling optimization model, and only update the parameters of the multi-objective optimization layer and the scheme generation layer.

[0054] Preferably, when the processor executes the computer program, it implements the steps of the artificial intelligence-based cargo transportation scheduling method as described in any of the above-mentioned embodiments.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] The task spatiotemporal encoding generates a data structure containing time-sensitivity weights and path conflict markers. This transforms abstract timeliness requirements and spatial conflict risks into quantifiable parameters that the model can directly process. Time-sensitivity weights precisely quantify the urgency of different tasks in the time dimension, enabling the model to assign higher optimization priority to highly sensitive tasks. Path conflict markers pre-identify potential competition points between different tasks in terms of spatiotemporal resources. This provides deep neural networks with input features rich in scheduling semantics, allowing the model to deeply understand complex spatiotemporal constraints and generate preliminary scheduling schemes that perform better in terms of constraint satisfaction.

[0057] The multi-scheme game theory decision-making module interacts with and selects the best option based on preset game rules and real-time traffic information flow. The game rules define the interaction logic and benefit balance objectives of multiple scheduling schemes in the simulated environment. Real-time traffic information flow continuously injects external dynamic changes. By simulating the game process between multiple schemes, the robustness and adaptability of each scheme in the face of dynamic interference and the balancing of multiple interests are evaluated. It can select the scheduling scheme with the best overall performance and strongest anti-interference capability in a dynamically changing environment, surpassing the decision-making mode of traditional single-scheme output or static rule selection, and improving the scheduling system's ability to cope with uncertainty. Attached Figure Description

[0058] Figure 1 This is a schematic diagram illustrating the working principle of the artificial intelligence-based cargo transportation scheduling method described in this invention.

[0059] Figure 2 A flowchart for task spatiotemporal encoding;

[0060] Figure 3 A flowchart for processing deep neural network scheduling optimization models;

[0061] Figure 4 The graph shows the iterative changes of the comprehensive transportation benefit payoff function in a multi-option game decision-making process.

[0062] Figure 5A comparison chart of execution deviations before and after updating the deep neural network scheduling optimization model. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Please see Figure 1 This invention provides an artificial intelligence-based freight transportation scheduling method, comprising: standardizing and formatting the original transportation task data, including cargo attributes, path nodes, and predetermined time, to obtain standard task description data; performing task spatiotemporal encoding on the standard task description data to generate task spatiotemporal encoded data containing time sensitivity weights and path conflict markers; inputting the task spatiotemporal encoded data into a trained deep neural network scheduling optimization model, which outputs a set of preliminary scheduling schemes corresponding to the original transportation task data; inputting the set of preliminary scheduling schemes into a multi-scheme game decision module, and selecting an optimal scheduling scheme from the set of preliminary scheduling schemes based on preset game rules and real-time traffic information flow; and performing dynamic environment simulation on the optimal scheduling scheme to generate an execution simulation path with time and resource tags.

[0065] Example 1: See Figure 2 The process involves performing task spatiotemporal coding on standard task description data, parsing the predetermined time in the standard task description data, calculating and assigning time sensitivity weights based on the urgency of the time; identifying path nodes in the standard task description data and comparing their overlap with historical task paths, marking overlapping path segments as path conflict markers; fusing the time sensitivity weights with the path conflict markers and adding a unique task identifier corresponding to the original transportation task data to generate structured coded data; and sorting the structured coded data in descending order of time sensitivity weights to generate task spatiotemporal coded data.

[0066] In practice, the original transportation task data contains multiple tasks to be scheduled. Task attributes include cargo type, volume, weight, origin, destination, and the customer-required "scheduled time." Standard task description data has already converted this information into a unified digital vector format. The first step in performing spatiotemporal coding on the standard task description data is to parse the "scheduled time" of each task and calculate its time sensitivity weight. The time sensitivity weight is calculated based on the difference between the task's "scheduled time" and the current system time; the smaller the difference, the higher the urgency of the task, and the greater the weight assigned. In some embodiments, the calculation of the time sensitivity weight is quantified using a formula:

[0067]

[0068] Where: characters Indicates the first Time sensitivity weight of each task, character Indicates the first The scheduled timestamp for each task, characters Represents the current system timestamp, character This is a preset normalization constant. Based on this formula, the time sensitivity weight calculated for a task scheduled 2 hours later will be significantly higher than that for a task scheduled 24 hours later. In practice, path nodes in the standard task description data are identified and compared with historical task paths for overlap. Historical task paths refer to the set of path segments occupied by currently scheduled but not yet completed transportation tasks. For a new task, its planned path from start to finish is discretized into a series of continuous road segments. The system compares these segments one by one with the segments in the historical task path set. When the overlap of any road segment exceeds a preset threshold, that road segment is marked as a "path conflict marker." The path conflict marker not only records the geographic identifier of the conflicting road segments but also associates it with the severity level of the conflict. The severity level is determined based on the length of the overlapping road segments and the expected traffic flow density during the overlapping time period.

[0069] Understandably, after calculating the time-sensitivity weights and identifying path conflict markers, it's necessary to fuse the time-sensitivity weights and path conflict markers. The fusion process involves concatenating the single numerical time-sensitivity weight calculated for each task with a structured list of all path conflict markers for that task, and adding a globally unique task identifier to the concatenated data block, thus generating structured coded data. This task identifier ensures that the data identity of that task is clear and traceable in all subsequent processing stages.

[0070] In some embodiments, the generated structured coded data is a multi-dimensional vector whose dimensions include task identifiers, time-sensitivity weights, and a variable-length array for storing path conflict markers. Finally, the structured coded data of all tasks is treated as a single list and sorted in descending order according to the values ​​of their respective time-sensitivity weights, with the task having the highest time-sensitivity weight at the front of the list. This sorted list of structured coded data constitutes the final "task spatiotemporal coded data." This sorting operation ensures that urgent tasks receive priority processing in subsequent scheduling optimization models. In practical implementation, after the above process, an urgent medical supply transportation task, even if its path partially conflicts with several ordinary tasks, will receive higher priority in resource competition due to its earlier position in the task spatiotemporal coded data list.

[0071] Example 2: See Figure 3 The initial encoding layer of the deep neural network scheduling optimization model receives the task's spatiotemporal encoded data. The output of the initial encoding layer enters a spatiotemporal attention allocation layer, which dynamically allocates and calculates attention based on time sensitivity weights and path conflict markers. The output of the spatiotemporal attention allocation layer is input into a multi-objective optimization layer, which simultaneously calculates a combination of objective functions that minimize transportation costs, minimize time delays, and maximize path utilization. The output of the multi-objective optimization layer is decoded by a scheme generation layer, which outputs a set of preliminary scheduling schemes containing specific vehicle allocations, path sequences, and time nodes.

[0072] In practice, the initial encoding layer performs feature mapping and vectorization on the structured encoded data of each task in the list, converting the variable-length arrays of task identifiers, time-sensitivity weight values, and path conflict markers into fixed-dimensional dense vector representations. These initially encoded dense vectors form an initial feature sequence according to the original order of the task's spatiotemporal encoded data and are output to the spatiotemporal attention allocation layer.

[0073] The spatiotemporal attention allocation layer dynamically allocates computational attention based on the time sensitivity weights and path conflict markers embedded in each task feature vector. Task feature vectors with higher time sensitivity weights receive higher attention weights, meaning the model allocates more computational resources to them in subsequent calculations. The presence of path conflict markers also modulates the attention mechanism; for task feature vectors with high-level path conflict markers, the spatiotemporal attention allocation layer enhances the attention interaction strength between them and other task feature vectors with path conflicts, specifically addressing potential resource contention issues. In some embodiments, the spatiotemporal attention allocation layer employs a multi-head attention mechanism, where the generation of its query, key, and value matrices incorporates correction factors for time sensitivity weights and embedded representations of path conflict markers. The output of the spatiotemporal attention allocation layer is a new feature sequence after attention weighting and feature interaction, which emphasizes the relationship between urgent tasks and tasks with path conflicts.

[0074] In practical implementation, the output of the spatiotemporal attention allocation layer is input into a multi-objective optimization layer. This layer simultaneously calculates a combination of objective functions: minimizing transportation costs, minimizing time delay, and maximizing path utilization. Essentially, the multi-objective optimization layer is a fully connected network that receives the aforementioned feature sequences and computes the values ​​of multiple objectives in parallel. The objective function for minimizing transportation costs estimates vehicle fuel consumption, driver labor costs, and road tolls based on the feature sequences. The objective function for minimizing time delay calculates the total deviation between the task completion time and the scheduled time based on the feature sequences. The objective function for maximizing path utilization assesses vehicle load factor and the saturation level of the travel path based on the feature sequences. The multi-objective optimization layer combines these three objective functions into a comprehensive optimization objective. Optionally, the comprehensive optimization objective is expressed as a weighted sum of formulas:

[0075]

[0076] Where: characters This represents the overall optimization objective value that needs to be maximized. Represents the value of the transportation cost sub-function, character Represents the value of the time delay subfunction, character The value of the path utilization subfunction is represented by the character. , , These are the non-negative weight coefficients obtained through pre-training.

[0077] The output features of the multi-objective optimization layer are then decoded by a scheme generation layer, which is typically a sequence-to-sequence decoder structure. In some embodiments, the scheme generation layer operates in an autoregressive manner, progressively generating each decision unit in the scheduling scheme. The scheme generation layer maps the abstract features output by the multi-objective optimization layer into specific, executable scheduling instructions, which are ultimately organized into a set of preliminary scheduling schemes. Each set of preliminary scheduling schemes can be understood as a complete scheduling plan, including specific vehicle allocation, path order, and time nodes. Vehicle allocation specifies the vehicle identifier executing each task, path order specifies the order in which each vehicle visits each path node, and time nodes estimate the start and end times of each operation. In practical implementation, the deep neural network scheduling optimization model, through an end-to-end approach, transforms the input task spatiotemporally encoded data into a set of preliminary scheduling schemes containing specific vehicle allocation, path order, and time nodes, thus completing the transformation from problem features to solutions.

[0078] Example 3: The multi-scheme game decision-making module receives real-time traffic information streams, which include dynamically updated road segment speeds and traffic event data. Each preliminary scheduling scheme is treated as a game participant, and a non-cooperative game model is established with comprehensive transportation benefits as the payoff function. In the non-cooperative game model, the Nash equilibrium point of each preliminary scheduling scheme is iteratively calculated using the real-time traffic information stream. The preliminary scheduling scheme associated with the Nash equilibrium point corresponding to the highest payoff function value is selected as the optimal scheduling scheme. Comprehensive transportation benefits are defined as the weighted sum of an economic benefit function and a time benefit function. The economic benefit function is calculated based on the estimated fuel consumption, vehicle wear and tear, and route costs of the preliminary scheduling scheme. The time benefit function is calculated based on the difference between the estimated completion time and the scheduled time after applying the real-time traffic information stream. A virtual game strategy space is allocated to each preliminary scheduling scheme, containing minor adjustments to path nodes. Multiple rounds of strategy selection and payoff calculations are performed within the game strategy space until the strategy selections of all preliminary scheduling schemes tend to stabilize.

[0079] In practical implementation, the multi-scheme game decision-making module receives real-time traffic information streams from external data interfaces. These streams contain dynamically updated road segment speeds and traffic event data. For example, if the speed on a highway changes from the normal 100 km / h to 40 km / h due to an accident, or if a temporary traffic control event is reported at an intersection, this information is continuously fed into the multi-scheme game decision-making module in the form of structured messages. In practice, the system assigns a set of preliminary scheduling schemes output by the deep neural network scheduling optimization model—for example, three preliminary scheduling schemes numbered A, B, and C—as game participants. Each preliminary scheduling scheme represents a complete resource allocation and path planning strategy.

[0080] Each preliminary scheduling scheme is treated as a game participant. The multi-scheme game decision module establishes a non-cooperative game model with comprehensive transportation efficiency as the payoff function. Comprehensive transportation efficiency is a quantitative indicator for evaluating the merits of a preliminary scheduling scheme under real-time traffic conditions. In some embodiments, comprehensive transportation efficiency is defined as a weighted sum of an economic efficiency function and a time efficiency function. The economic efficiency function is calculated based on the estimated fuel consumption, vehicle wear and tear, and route costs of the preliminary scheduling scheme. The time efficiency function is calculated based on the difference between the estimated completion time and the scheduled time after applying real-time traffic information. It can be understood that a virtual game strategy space is allocated to each preliminary scheduling scheme. This space contains minor adjustments to path nodes, such as replacing a node in the path with an adjacent backup node or swapping the order of two consecutive tasks. These minor adjustments constitute the set of strategies available to each game participant.

[0081] In practical implementation, the Nash equilibrium point of each preliminary scheduling scheme is iteratively calculated using real-time traffic information flow in the non-cooperative game model. The calculation process involves multiple rounds of strategy selection and payoff calculation within the game's strategy space. In each iteration, each preliminary scheduling scheme, based on the current state of other schemes, selects a small adjustment operation from its own game strategy space that improves its overall transportation benefit payoff function value. The calculation of the payoff function value depends entirely on the latest real-time traffic information flow. After multiple rounds of strategy selection and payoff calculation, the strategy selections of all preliminary scheduling schemes tend to stabilize, meaning that no preliminary scheduling scheme can obtain a higher overall transportation benefit payoff function value by unilaterally changing its own strategy. At this point, the system reaches a Nash equilibrium state. Optionally, the overall transportation benefit payoff function value corresponding to the Nash equilibrium point is calculated and compared using the following formula:

[0082]

[0083] Where: characters Indicates the preliminary scheduling plan The value of the comprehensive transportation benefit payment function at the Nash equilibrium point, character This represents the calculated value of the economic benefit function corresponding to the proposed scheme. This represents the calculated value of the time benefit function corresponding to the scheme, in characters. and These are preset weighting coefficients used to balance economic and time factors, and .

[0084] In some embodiments, after the iterative calculation is completed, the multi-scheme game decision module selects the preliminary scheduling scheme associated with the Nash equilibrium point corresponding to the highest payoff function value as the optimal scheduling scheme. For example, after game calculation, the payoff function values ​​of preliminary scheduling schemes A, B, and C at the equilibrium point are... , , With payoff values ​​of 8.5, 9.2, and 7.8 respectively, preliminary scheduling scheme B is selected as the optimal scheduling scheme because of its highest payoff function value.

[0085] See Figure 4 This graph, with the number of iterations on the horizontal axis and the comprehensive transportation benefit payoff function value on the vertical axis, illustrates the changing trends of the benefits of three initial scheduling schemes during the game process. The graph visually presents the iterative process of the non-cooperative game model: each scheme adjusts its strategy based on the comprehensive transportation benefit payoff function, influenced by real-time traffic information flow. Ultimately, scheme A achieves a stable and highest payoff function value, conforming to the core rule of selecting the Nash equilibrium point associated with the highest payoff function value as the optimal scheduling scheme. This clearly demonstrates the game mechanism's role in screening the robustness and adaptability of scheduling schemes.

[0086] Example 4: Based on the optimal scheduling scheme, extract its planned path sequence and planned time sequence; construct a discrete event simulation environment with time evolution as the axis; in the discrete event simulation environment, gradually inject transportation events from the optimal scheduling scheme according to the planned time sequence; at each simulation time step, introduce randomly generated external disturbance events and calculate the state evolution of transportation events under disturbance; record the actual path offset and actual time consumption caused by the state evolution to form the execution simulation path. Define the global clock and time advancement mechanism of the discrete event simulation environment; register multiple types of transportation entity models in the discrete event simulation environment, including at least vehicle models and cargo models; register multiple types of environmental event generators in the discrete event simulation environment, which can generate events simulating traffic congestion and weather changes; establish a response rule base for transportation entity models to environmental event generators, which defines the logic of entity state changes.

[0087] In practical implementation, based on the optimal scheduling scheme selected from the multi-scheme game decision module, its planned path sequence and planned time sequence are extracted. For example, the optimal scheduling scheme instructs vehicle V001 to pass through path nodes A, B, C, and D in sequence, planning to depart from A at timetamp T1, arrive at B at timetamp T2, arrive at C at timetamp T3, and finally arrive at D at timetamp T4. In practical implementation, a discrete event deduction environment with time evolution as the axis is constructed, defining a global clock and time advancement mechanism for the discrete event deduction environment. The initial time of the global clock is set to the planned start time, and the time advancement mechanism adopts an event-driven approach, that is, whenever a transportation event or external interference event occurs, the global clock immediately jumps to the time of occurrence of the event for processing.

[0088] Multiple types of transportation entity models are registered in the discrete event simulation environment. These models include at least vehicle and cargo models. Vehicle models have attributes such as location, speed, load status, and driver fatigue, while cargo models have attributes such as cargo identification, vehicle ownership, and cargo status. Multiple types of environmental event generators are also registered in the environment. These generators can generate events simulating traffic congestion and weather changes. For example, a weather change event generator can generate a "heavy rain" event at a specific time and geographical area, causing the speed coefficient of the affected road segment to decrease to 0.6 of its original value. In implementation, a response rule base for transportation entity models to environmental event generators is established. This rule base defines the logic for entity state changes. The rule base includes rules such as, "When a vehicle model encounters a 'traffic congestion' event, its speed attribute will be updated according to a predefined speed-congestion lookup table based on the congestion level."

[0089] In the discrete event simulation environment, transportation events from the optimal scheduling scheme are injected step-by-step according to the planned time sequence. These transportation events include "vehicle departure," "vehicle arrival at node," "start loading / unloading," and "complete loading / unloading." At each simulation time step, while advancing the global clock and processing planned transportation events, the system introduces randomly generated external disturbance events based on a probabilistic model. In practice, the probabilistic model operates through pre-registered environmental event generators in the discrete event simulation environment. These generators are configured to simulate traffic congestion, weather changes, and other external disturbance events. Each event generator encapsulates dynamic probabilistic logic based on the current state of the simulation environment, such as combining the global clock time period and real-time road segment load for trigger evaluation. At each simulation time step, the system iterates through active generators, determining whether to generate an event based on its probability rules, such as "road segment L1 congestion" or "heavy rain in area A." After triggering, the system injects the event instance into the environment and drives the transportation entity model to evolve its state according to the response rule base, thereby achieving dynamic simulation of the disturbance process of transportation events. It can be understood that the randomly generated external disturbance events originate from the registered environmental event generators, and their occurrence time, location, and intensity are all random. For example, when the global clock advances to T2.5, the system introduces a "temporary traffic control" event occurring on the road segment between nodes B and C. In practice, calculating the state evolution of transportation events under disturbances requires relying on the response rule base and considering the cumulative effect of events. For instance, if a vehicle model is simultaneously affected by both a "heavy rain" event and a "traffic congestion" event, its speed update will incorporate the coefficient decay caused by both events. The calculation of state evolution can be expressed by the following formula:

[0090]

[0091] Where: characters This represents the actual speed of the vehicle model at the current moment after being affected by all active events. This represents the baseline speed of the vehicle model under undisturbed conditions. Indicates the first The impact coefficient of an effective external interference event on the traffic speed of this road segment, character This indicates the number of external interference events that are currently active on the road segment where the vehicle is located.

[0092] Referring to Table 1, the actual path offset and actual time consumption generated by the state evolution are recorded to form the execution simulation path. The execution simulation path not only records the final actual sequence of nodes the vehicle passes through, but also attaches the actual time label and resource label of each key action.

[0093] Table 1: Event Record Table for Dynamic Environment Simulation

[0094]

[0095] The records in Table 1 clearly show that the planned completion time for the optimal scheduling scheme is T4. However, after dynamic environment simulation, the actual completion time is delayed to T4+95 minutes, and the specific events and resource consumption changes that caused the delay are recorded. In some embodiments, the execution simulation path is output as trajectory data with time and resource labels. The time labels mark the time when each actual event occurs, and the resource labels record the actual fuel consumption and additional vehicle wear information for each road segment.

[0096] Example 5: During the actual execution of the simulated path, execution process monitoring data is collected; the execution process monitoring data is compared with the expected data of the simulated path to generate execution deviation data; the execution deviation data is fed back into the training process of the deep neural network scheduling optimization model to update the network parameters of the deep neural network scheduling optimization model; the deep neural network scheduling optimization model with updated network parameters is used to process the subsequent arriving transportation task data. The execution deviation data is converted into the loss function increment required for model training; the new training samples containing the loss function increment are merged with the historical training samples to form an incremental training dataset; the deep neural network scheduling optimization model is trained using the incremental training dataset with an incremental learning approach; the parameters of the basic feature extraction layer in the deep neural network scheduling optimization model are frozen, and only the parameters of the multi-objective optimization layer and the scheme generation layer are updated.

[0097] In practice, the system collects execution process monitoring data through vehicle-mounted GPS terminals, IoT sensors, and the warehouse management system. This monitoring data includes the vehicle's actual driving trajectory sequence, the timestamp for each trajectory point, actual fuel consumption readings, the actual start and end times of loading and unloading operations, and reports of real traffic incidents encountered along the way. For example, a truck assigned to execute the optimal scheduling plan might have GPS records showing that, during the time it was supposed to pass through the highway, it actually traveled on a parallel national highway, and its actual arrival time at the destination was 30 minutes later than the estimated time in the simulated route.

[0098] In practice, execution deviation data is generated by comparing the monitoring data during the execution process with the expected data of the execution simulation path. The expected data comes from the time tags and path tags attached to the execution simulation path. The comparison operation is performed on an event-by-event basis. For the event of "vehicle arriving at a certain node," the system compares its planned time, the estimated time of the execution simulation path, and the actual monitored time to calculate the time deviation value. For "vehicle travel path," the system compares the planned path sequence, the actual path offset sequence of the execution simulation path, and the actual trajectory sequence monitored by GPS to calculate the path deviation degree. It can be understood that the generated execution deviation data is a structured difference record, which not only contains numerical deviations but also associates contextual information that caused the deviations, such as "a traffic accident actually occurred on road segment L1, causing an additional delay of 30 minutes compared to the estimated time in the simulation path."

[0099] The training process that feeds execution bias data back to the deep neural network scheduling optimization model is used to update the network parameters of the deep neural network scheduling optimization model. This process begins by converting the execution bias data into a loss function increment required for model training. The loss function increment is a scalar value calculated based on the bias data, which quantifies the difference between the deep neural network scheduling optimization model's previous predictions and the final actual results. In some embodiments, the loss function increment can be calculated according to the formula:

[0100]

[0101] Where: characters This represents the calculated increment of the loss function, character. Indicates the number of events that caused the deviation, character Indicates the first The actual time of occurrence of each event, characters This indicates that the deep neural network scheduling optimization model optimizes the execution of the inferred path for the first... Predicted time for each event, character It is a function used to calculate the first... The actual path in each event With predicted path Spatial distance measurement between characters and These are preset coefficients used for weighted time deviation and path deviation, respectively.

[0102] In practice, the new training samples containing the loss function increment are merged with historical training samples to form an incremental training dataset. The construction of the new training samples uses the original task spatiotemporal encoded data that triggered the current round of scheduling optimization as input features, and the loss function increment calculated according to the formula as part of the target supervision signal. An incremental learning approach is used to train the deep neural network scheduling optimization model using the incremental training dataset. Incremental learning aims to enable the model to absorb new experiences without forgetting old knowledge. Optionally, to achieve this goal, the training process freezes the parameters of the basic feature extraction layer in the deep neural network scheduling optimization model. The basic feature extraction layer typically includes the initial encoding layer and part of the lower-level network; only the parameters of the multi-objective optimization layer and the scheme generation layer are updated. This partially frozen strategy helps maintain the stability of the model's learned general feature representation capabilities while allowing the decision layer to adapt to new data patterns.

[0103] See Figure 5 This graph, with deviation type on the horizontal axis and deviation value on the vertical axis, compares the deviation performance before and after the model update using grouped bar charts: before the model update, the values ​​of indicators such as time deviation and loading / unloading time deviation were relatively high; while after the update, the values ​​of all deviation types decreased significantly. This graph intuitively verifies the effectiveness of the execution deviation feedback-incremental learning model update mechanism: the actual execution monitoring data is compared with the predicted data to generate deviations, which are then converted into incremental loss function values ​​to update the parameters of the multi-objective optimization layer and the scheme generation layer of the model. The decreasing trend of deviations in this graph is a reflection of the actual effect of this mechanism. After deviation feedback optimization, the scheduling accuracy of transportation time, route, and resource consumption is greatly improved, further enhancing the adaptability of the scheduling scheme in real-world scenarios.

[0104] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A freight transportation scheduling method based on artificial intelligence, characterized in that, Includes the following steps: The raw transportation task data, which includes cargo attributes, route nodes, and scheduled time, is standardized and formatted preprocessed to obtain standard task description data. Perform task spatiotemporal coding on the standard task description data to generate task spatiotemporal coded data that includes time sensitivity weights and path conflict markers; The task spatiotemporal encoded data is input into a trained deep neural network scheduling optimization model, and the deep neural network scheduling optimization model outputs a set of preliminary scheduling schemes corresponding to the original transportation task data. The set of preliminary scheduling schemes is input into a multi-scheme game decision module, and an optimal scheduling scheme is selected from the set of preliminary scheduling schemes based on the preset game rules and real-time traffic information flow. The optimal scheduling scheme is dynamically simulated to generate an execution simulation path with time and resource tags.

2. The artificial intelligence-based freight transportation scheduling method according to claim 1, characterized in that, Performing task spatiotemporal encoding on the standard task description data includes the following steps: The predetermined time in the standard task description data is parsed, and a time sensitivity weight is calculated and assigned according to the urgency of the time. Identify path nodes in the standard task description data and compare their overlap with historical task paths, marking overlapping path segments as path conflict markers. The time sensitivity weight is fused with the path conflict marker, and a unique task identifier corresponding to the original transportation task data is added to generate structured coded data. The structured coded data is sorted in descending order according to the time sensitivity weight to generate the task spatiotemporal coded data.

3. The artificial intelligence-based freight transportation scheduling method according to claim 2, characterized in that, The step of inputting the task spatiotemporal encoded data into a trained deep neural network scheduling optimization model includes the following steps: The initial encoding layer of the deep neural network scheduling optimization model receives the task spatiotemporal encoded data; The output of the initial encoding layer enters a spatiotemporal attention allocation layer, which dynamically allocates and calculates attention based on the time sensitivity weight and the path conflict marker. The output of the spatiotemporal attention allocation layer is input into a multi-objective optimization layer, which simultaneously calculates a combination of objective functions that minimize transportation costs, minimize time delays, and maximize path utilization. The output of the multi-objective optimization layer is decoded by a scheme generation layer, which outputs a set of preliminary scheduling schemes including specific vehicle allocation, path order, and time nodes.

4. The artificial intelligence-based cargo transportation scheduling method according to claim 3, characterized in that, The step of inputting the initial scheduling scheme into a multi-scheme game decision module includes the following steps: The multi-scheme game decision module receives real-time traffic information streams, which include dynamically updated road segment traffic speeds and traffic event data. By treating each preliminary scheduling scheme as a game participant, a non-cooperative game model is established with comprehensive transportation benefits as the payoff function. In the non-cooperative game model, the Nash equilibrium point of each preliminary scheduling scheme is calculated iteratively using the real-time traffic information flow; The preliminary scheduling scheme associated with the Nash equilibrium point corresponding to the highest payoff function value is selected as the optimal scheduling scheme.

5. The artificial intelligence-based cargo transportation scheduling method according to claim 4, characterized in that, The establishment of the non-cooperative game model with comprehensive transportation benefits as the payoff function includes the following steps: The comprehensive transportation benefit is defined as the weighted sum of the economic benefit function and the time benefit function; The economic benefit function is calculated based on the estimated fuel consumption, vehicle wear and tear, and route costs of the preliminary scheduling plan. The time efficiency function is calculated based on the difference between the estimated completion time and the predetermined time after the application of the real-time traffic information stream in the preliminary scheduling scheme. A virtual game strategy space is assigned to each preliminary scheduling scheme, and the game strategy space contains minor adjustment operations on path nodes; Multiple rounds of strategy selection and payoff calculation are performed within the game strategy space until the strategy selection of all preliminary scheduling schemes tends to stabilize.

6. The artificial intelligence-based freight transportation scheduling method according to claim 5, characterized in that, The dynamic environmental simulation of the optimal scheduling scheme includes the following steps: Based on the optimal scheduling scheme, extract its planned path sequence and planned time sequence; Construct a discrete event deduction environment with time evolution as the axis; In the discrete event simulation environment, transportation events in the optimal scheduling scheme are gradually injected according to the planned time series. At each simulation time step, a randomly generated external disturbance event is introduced, and the state evolution of the transportation event under the disturbance is calculated; The actual path offset and actual time consumption resulting from the state evolution are recorded to form the execution deduction path.

7. The artificial intelligence-based freight transportation scheduling method according to claim 6, characterized in that, The construction of a discrete event deduction environment with time evolution as the axis includes the following steps: Define the global clock and time progression mechanism of the discrete event deduction environment; Register multiple types of transportation entity models in the discrete event simulation environment, wherein the transportation entity models include at least vehicle models and cargo models; Register various types of environmental event generators in the discrete event simulation environment. These environmental event generators can generate events that simulate traffic congestion and weather changes. A response rule base for the transportation entity model to the environmental event generator is established, and the response rule base defines the logic of entity state changes.

8. The artificial intelligence-based freight transportation scheduling method according to claim 7, characterized in that, It also includes the following steps: During the actual execution of the simulated execution path, execution process monitoring data is collected; The execution process monitoring data is compared with the expected data of the execution simulation path to generate execution deviation data; The execution deviation data is fed back into the training process of the deep neural network scheduling optimization model to update the network parameters of the deep neural network scheduling optimization model; The deep neural network scheduling optimization model, after updating the network parameters, is used to process the subsequent arriving transportation task data.

9. The artificial intelligence-based freight transportation scheduling method according to claim 8, characterized in that, The process of feeding back the execution deviation data to the training process of the deep neural network scheduling optimization model includes: The execution bias data is converted into the loss function increment required for model training; The new training samples containing the increment of the loss function are merged with the historical training samples to form an incremental training dataset. The deep neural network scheduling optimization model is trained using an incremental learning approach and the incremental training dataset. Freeze the parameters of the basic feature extraction layer in the deep neural network scheduling optimization model, and only update the parameters of the multi-objective optimization layer and the scheme generation layer.

10. An artificial intelligence-based freight transportation scheduling system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based cargo transportation scheduling method as described in any one of claims 1 to 9.

Images