A logistics transportation intelligent scheduling method and system

Abstract

The application discloses a kind of logistics transportation intelligent scheduling method and system, belong to intelligent logistics transportation technical field, this method includes: extracting time sequence characteristics from real-time transportation data;Time sequence characteristics are input into prediction model, obtain predicted default probability, and default loss is calculated based on predicted default probability and default penalty;If default loss is greater than the preset intervention threshold, then based on the current vehicle location, path planning is carried out again, locate the candidate path where the default loss is minimum, if its default loss is less than the intervention threshold, then the candidate path is used as self-healing path, vehicle changes to self-healing path and carries out goods transportation;If self-healing path cannot be located and default loss is greater than scheduling threshold, then based on road network topology and candidate vehicle set, heterogeneous road network graph is constructed;Heterogeneous road network graph is analyzed based on graph neural network and decision model to output the best scheduling scheme. Through the application, transportation timeliness can be improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent logistics transportation technology, specifically relating to an intelligent scheduling method and system for logistics transportation. Background Technology

[0002] In the logistics and transportation sector, technologies such as GPS, GIS, and vehicle-to-everything (V2X) enable real-time vehicle location tracking. Combined with big data and cloud computing, route planning and estimated arrival time calculations significantly improve the visibility and management efficiency of logistics. However, existing technologies, when performing route planning or emergency dispatch, lack comprehensive consideration of dynamic, multi-dimensional factors such as road conditions, vehicle status, and driving behavior. Their models are relatively simple and struggle to output optimal dynamic adjustment or rescue solutions, resulting in bottlenecks in transportation efficiency and timeliness.

[0003] To address the aforementioned issues, Chinese patent document CN117114528A proposes a logistics transportation analysis system and method based on big data. This method comprehensively considers factors such as the distance between the replacement vehicle and the breakdown point, driver fatigue, and familiarity with historical routes, solving the problem of how to quickly and rationally select rescue vehicles in the event of a sudden breakdown and improving the efficiency of emergency response decisions. However, the drawback of this solution is that it is essentially still a reactive measure; it initiates dispatch only after the vehicle has already broken down, making it impossible to anticipate and mitigate risks in advance. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides an intelligent scheduling method and system for logistics transportation, thereby resolving the issues present in the background section.

[0005] To achieve the aforementioned objectives, this invention proposes an intelligent logistics transportation scheduling method, comprising:

[0006] Acquire real-time transportation data of vehicles and extract time-series features from the real-time transportation data;

[0007] The time-series features are input into the prediction model to obtain the predicted default probability, obtain the default penalty for vehicle transportation orders, and calculate the default loss based on the predicted default probability and default penalty.

[0008] If the breach of contract loss exceeds the preset intervention threshold, the route is replanned based on the current vehicle location to obtain multiple candidate routes. The candidate route with the smallest breach of contract loss is located. If the breach of contract loss is less than the intervention threshold, the candidate route is used as the self-healing route and the vehicle is switched to the self-healing route for cargo transportation.

[0009] If a self-healing path cannot be located, then within a preset neighborhood centered on the current vehicle location, count the number of ideal vehicles with a default probability less than a preset threshold and no transportation task, and map the number of ideal vehicles inversely to an emergency resource scarcity between 0 and 1. Multiply the emergency resource scarcity by a preset fixed cost to obtain the scheduling threshold.

[0010] If the default loss exceeds the scheduling threshold, a candidate vehicle set is generated based on the load compatibility check, and a heterogeneous road network map containing vehicle nodes and hub nodes is constructed based on the road network topology and the candidate vehicle set.

[0011] The heterogeneous road network map is analyzed based on graph neural networks and decision models to output the optimal scheduling scheme for decentralized transportation. Furthermore, the graph neural network embeds impedance suppression mechanism and flow direction gating mechanism during the node feature aggregation process to respectively reduce the information weight of high-resistance paths and logically block reverse paths.

[0012] Furthermore, inputting the time-series features into the prediction model to obtain the predicted default probability includes the following steps:

[0013] The intersection of road segments in the electronic map is defined as a physical node. The damping coefficient of the connecting path between two physical nodes is calculated. The damping coefficient is proportional to the resistance of logistics transportation. After obtaining the real-time location in the current transportation path, the damping coefficient of each road segment is generated, and the change gradient of the damping coefficient sequence is calculated.

[0014] The temporal features include the changing gradient. The prediction model is an LSTM model. An inhibition operator is embedded between the output of the forget gate of the LSTM model and the cell state at the previous time step. A slope activation unit is embedded at the output of the cell state. The inhibition operator generates a damping vector based on the damping coefficient at the current position. The cell state at the previous time step after the forget gate is modified based on the damping vector to suppress the transmission of historical inertial features. The slope activation unit maps the changing gradient to the slope factor of the activation function. When the changing gradient is negative, the slope activation unit increases the slope factor to amplify the cell state output signal. When the changing gradient is positive, the slope factor decreases to compress the cell state output signal. The prediction model uses the updated cell state to calculate the predicted default probability.

[0015] Furthermore, calculating the damping coefficient of the connectivity path between two physical nodes includes the following steps:

[0016] Obtain the baseline speed limit and real-time traffic speed of the connected path, and calculate the ratio of the difference between the baseline speed limit and the real-time traffic speed as the damping coefficient.

[0017] Furthermore, the time-series features also include speed features, progress features, and driving behavior features. The speed feature is the average speed of various road segments during transportation, and the progress feature is the ratio of the completed route length in the current transportation route to the total transportation route length. The transportation route is divided into multiple business nodes, and the average time taken for vehicles to reach each business node is calculated based on historical data. The actual time taken for vehicles to reach each business node is obtained based on the real-time transportation data. The ratio of the actual time taken to the corresponding average time taken is used as an efficiency value. An efficiency sequence is generated based on multiple efficiency values, and the efficiency sequence is analyzed to determine the driving behavior features.

[0018] Furthermore, the prediction model includes a boundary loss function. When the implicit average driving speed of the prediction model exceeds the maximum physical speed, the boundary loss function outputs a penalty to correct the prediction model.

[0019] Furthermore, the analysis of the heterogeneous road network graph based on graph neural networks and decision models to output the optimal scheduling scheme includes the following steps:

[0020] The initial weights of edges in the heterogeneous road network graph are determined based on the damping coefficient. The heterogeneous road network graph is then aggregated using a graph neural network to correct the initial weights to dynamic travel costs. The scheduling task is modeled as a path optimization problem with the lowest overall cost on the heterogeneous road network graph. The overall cost of the scheduling scheme is calculated based on the dynamic travel costs and additional costs of each candidate vehicle. The overall cost of different scheduling schemes is calculated using a path planning algorithm, and the scheduling scheme with the lowest overall cost is determined as the optimal scheduling scheme.

[0021] Furthermore, the additional costs include entry potential cost, temporal potential cost, and regression potential cost. The entry potential cost is positively correlated with the topological complexity of the candidate vehicle entering the transfer hub from its current location. The temporal potential cost is positively correlated with the time difference between the candidate vehicle and the target vehicle arriving at the same transfer hub. The angle between the transfer path vector of the candidate vehicle heading to the destination of the relay cargo and its original task path vector is obtained, and the regression potential cost is positively correlated with this angle.

[0022] The present invention also provides an intelligent logistics transportation scheduling system, which is used to implement the above-described method, and the system includes:

[0023] The acquisition module obtains real-time transportation data of vehicles and extracts time-series features from the real-time transportation data;

[0024] The prediction module inputs the time-series features into the prediction model to obtain the predicted default probability, obtain the default penalty for the vehicle transportation order, and calculate the default loss based on the predicted default probability and the default penalty.

[0025] In the first planning module, if the default loss is greater than the preset intervention threshold, the route planning is re-performed based on the current vehicle location to obtain multiple candidate routes. The candidate route with the smallest default loss is located. If the default loss is less than the intervention threshold, the candidate route is used as the self-healing route and the vehicle is switched to the self-healing route for cargo transportation.

[0026] The second planning module, if a self-healing path cannot be located, counts the number of ideal vehicles with default probabilities less than a preset threshold and no transportation tasks within a preset neighborhood centered on the current vehicle location. This ideal vehicle number is then mapped inversely to an emergency resource scarcity between 0 and 1. The emergency resource scarcity is multiplied by a preset fixed cost to obtain a scheduling threshold. If the default loss exceeds the scheduling threshold, a candidate vehicle set is generated based on load compatibility verification. A heterogeneous road network map containing vehicle nodes and hub nodes is constructed based on the road network topology and the candidate vehicle set. The heterogeneous road network map is analyzed using a graph neural network and a decision model to output the optimal scheduling scheme for distributed transportation. Furthermore, the graph neural network embeds an impedance suppression mechanism and a flow gating mechanism during node feature aggregation to respectively reduce the information weight of high-resistance paths and logically block reverse paths.

[0027] The beneficial effects of this invention are as follows:

[0028] This invention first acquires real-time vehicle transportation data and inputs it into a prediction model to obtain the predicted probability of default. Combined with default penalties, it calculates the default loss, thus completing a quantitative assessment of future transportation risks. Then, when the default loss exceeds an intervention threshold, it replans routes based on the current vehicle location and identifies the self-healing path with the minimum default loss. This achieves low-cost risk avoidance through path self-healing at the initial stage of risk. Subsequently, when a self-healing path cannot be located and the default loss exceeds a dynamic scheduling threshold, a heterogeneous road network map containing candidate vehicles and hub nodes is constructed and analyzed using graph neural networks and decision models. Through the risk warning and intervention mechanism of this invention, the problem of existing technologies being unable to predict and avoid risks in advance is solved, improving the timeliness and intelligence of logistics transportation. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the steps of an intelligent logistics transportation scheduling method according to the present invention.

[0030] Figure 2 This is a comparison chart showing the change in the predicted default probability over time between the model of this invention and the traditional model.

[0031] Figure 3 This is a schematic diagram of the structure of an intelligent logistics transportation scheduling system according to the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various elements, but unless otherwise stated, these elements are not limited by these terms. These terms are used only to distinguish one element from another.

[0034] It should be noted that all information acquisition and utilization in this invention comply with the requirements of relevant laws and regulations and are carried out with the authorization of the device access administrator.

[0035] like Figure 1 As shown, a smart scheduling method for logistics transportation includes:

[0036] Acquire real-time transportation data of vehicles and extract time-series features from the real-time transportation data.

[0037] During vehicle transportation, real-time transportation data is continuously acquired through onboard terminals and a back-end system. This data includes vehicle location, real-time speed, and external environmental conditions. This data is used for subsequent calculations and is also displayed through a pre-defined interface for remote monitoring by personnel. The real-time transportation data is then analyzed to extract time-series characteristics.

[0038] The time-series features are input into the prediction model to obtain the predicted default probability, obtain the default penalty for vehicle transportation orders, and calculate the default loss based on the predicted default probability and default penalty.

[0039] In this embodiment, the temporal features, in addition to the speed, progress, and driving behavior features described later, also include the vehicle's current speed, speed limits and lane numbers for each segment of the subsequent path, and the gradient of change, as described later. A pre-trained prediction model is used to analyze the temporal features, outputting the predicted probability of failure to reach the destination on time. Simultaneously, a quantitative value of the penalty for failure is obtained from the transport order's service level agreement. Based on the probability of failure and the penalty, the expected loss due to failure is calculated. In this embodiment, the product of the probability of failure and the penalty is used as the loss due to failure. Therefore, the loss due to failure in this embodiment is a comprehensive value; a larger loss due to failure indicates a greater likelihood of failure or a greater economic loss caused by the failure.

[0040] If the breach of contract loss exceeds the preset intervention threshold, the route is replanned based on the current vehicle location to obtain multiple candidate routes. The candidate route with the smallest breach of contract loss is then selected. If the breach of contract loss is less than the intervention threshold, the candidate route is used as the self-healing route, and the vehicle is switched to the self-healing route for cargo transportation.

[0041] The intervention threshold is a value set according to the importance of the transported goods. The higher the importance level of the goods, the lower the intervention threshold. The rules for setting the intervention threshold can be determined based on actual needs, and will not be described here. By setting the intervention threshold, excessive intervention is avoided when the default loss is low. When the default loss exceeds the intervention threshold, multi-route planning is first performed again based on the vehicle's current location to generate a set of candidate routes. For example, all routes between the current location and the destination are considered as candidate routes. The default loss is then assessed again for each candidate route. If at least one candidate route has an assessed default loss lower than the intervention threshold, the route with the lowest loss is defined as the self-healing route, and an instruction is generated to guide the vehicle to switch to the self-healing route to continue transportation, thereby achieving risk avoidance with minimal operational changes.

[0042] If a self-healing path cannot be located, then within a preset neighborhood centered on the current vehicle location, count the number of ideal vehicles with a default probability less than a preset threshold and no transportation task, and map the number of ideal vehicles inversely to an emergency resource scarcity between 0 and 1. Multiply the emergency resource scarcity by a preset fixed cost to obtain the scheduling threshold.

[0043] If the default loss exceeds the scheduling threshold, a candidate vehicle set is generated based on the load compatibility check, and a heterogeneous road network map containing vehicle nodes and hub nodes is constructed based on the road network topology and the candidate vehicle set.

[0044] If a self-healing path cannot be located, it means that switching paths cannot avoid delays. In this case, other vehicles are used for distributed relay transportation to speed up the cargo transportation. Specifically, the relay transportation referred to in this embodiment is a one-to-many distributed parallel transportation mode. The execution logic of this mode is as follows: at a predetermined cargo transfer hub, the large batch of goods originally in a single transport unit is split into multiple sub-loads, which are then transferred to multiple vehicles with higher average speeds. Combined with the ease of passage of smaller vehicles, the transportation time of the goods is shortened, thereby maximizing the final delivery time of the entire batch of goods.

[0045] To avoid intervention during periods of resource scarcity, this step introduces an emergency resource scarcity level. This level quantifies the current urgency of available emergency transport capacity and dynamically adjusts the trigger threshold for intervention decisions accordingly, ensuring the cost-effectiveness of intervention actions.

[0046] In this embodiment, the scarcity of emergency resources is obtained as follows: First, a search neighborhood with a specific radius is defined centered on the real-time location of the current vehicle. All transport vehicles within this neighborhood are retrieved, and vehicles with a default probability less than a preset threshold or no transport task are selected as ideal vehicles. The preset threshold is, for example, 30%. Then, the number of ideal vehicles is counted and mapped to a value between 0 and 1. This mapped value is defined as the scarcity of emergency resources. In this embodiment, the smaller the number of ideal vehicles, the closer the mapped scarcity of emergency resources is to 1. The scarcity of emergency resources is multiplied by a preset fixed cost to obtain the scheduling threshold. In other words, the fewer ideal vehicles there are, the tighter the scheduling resources are, and the closer the scarcity of emergency resources is to 1. A higher scheduling threshold makes it less likely to trigger the calculation of a scheduling scheme, thus reducing the waste of computing resources. Specifically, if the calculation of a scheduling scheme cannot be triggered, it indicates that the current transport capacity is tight, and distributed relay transport may consume more costs. In this case, the vehicle is instructed to continue along the original route.

[0047] Load compatibility verification involves selecting vehicles with the appropriate transportation qualifications and available loading space from ideal vehicles based on the physical properties, temperature control requirements, and hazard classification of the goods to be transferred, forming a candidate vehicle set. Subsequently, the candidate vehicles, the vehicles to be transferred, and logistics hubs with cargo handover capabilities are used as nodes in a graph, and the actual road connections between nodes are used as edges in the graph, constructing a heterogeneous road network graph containing multiple types of nodes.

[0048] The heterogeneous road network map is analyzed based on graph neural networks and decision models to output the optimal scheduling scheme for decentralized transportation. Furthermore, the graph neural network embeds impedance suppression mechanism and flow direction gating mechanism during the node feature aggregation process to respectively reduce the information weight of high-resistance paths and logically block reverse paths.

[0049] The constructed heterogeneous road network map is input into a pre-trained graph neural network (Graph Neural Network). The Graph Neural Network updates the edge weights within its nodes through its multi-layer information propagation and aggregation mechanism. A larger weight indicates a higher toll cost for the road segment. For example, the weight of a road segment depends not only on its length but also on factors such as whether it connects to a congested road segment downstream or whether there are many slow-moving vehicle nodes in its vicinity. By learning these complex nonlinear relationships, the Graph Neural Network ultimately outputs a weighted road network map.

[0050] Next, the optimal scheduling scheme will be modeled as a path optimization problem. For example, for target transport vehicle B, the optimal transfer vehicle and hub node will be found from the candidate vehicle set and hub nodes. When generating the optimal scheduling scheme, all candidate scheme combinations will be traversed, and the total scheduling cost of the candidate scheme combinations will be calculated on the road network graph based on the path planning algorithm. After calculating the total cost of all candidate schemes, the scheme with the lowest total cost will be determined as the optimal scheduling scheme.

[0051] To enhance the physical perception capabilities and logical rigor of graph neural networks, this scheme incorporates impedance suppression and flow gating mechanisms into the standard graph neural network aggregation process. Impedance suppression quantifies real-time traffic conditions as the network's impedance; when aggregating neighbor node information, it attenuates the weights of high-impedance paths, indicating that these paths have significant travel costs. Flow gating blocks information propagation from reverse topology paths during aggregation, thus preventing the generation of scheduling schemes that violate traffic rules at the algorithm level.

[0052] For example, suppose logistics hub node A is evaluating surrounding roads to update its toll costs. It connects two roads: one leading to node B (low damping coefficient) and the other to node C (high damping coefficient). During information aggregation, an impedance suppression mechanism, through a decay function (e.g., update weight = initial weight * (1 - damping coefficient)), significantly reduces the weight of information from node C. This causes the graph neural network to perceive a sharp increase in toll costs in the C direction, thus prioritizing vehicles traveling in the B direction in the final scheduling decision.

[0053] The path from intersection X to intersection Y is a one-way street. When the graph neural network aggregates information, node Y can receive information from X. However, when node X aggregates its neighbor information, the flow gating mechanism checks the graph's adjacency matrix or edge list to identify that information cannot flow backward from Y to X, thus setting the aggregation weight of this path to zero. This algorithm avoids generating traffic violation schemes like reverse driving, ensuring the executability of the scheduling plan.

[0054] In this embodiment, inputting the time-series features into the prediction model to obtain the predicted default probability includes the following steps:

[0055] The intersections of road segments in the electronic map are defined as physical nodes. The damping coefficient of the connecting path between two physical nodes is calculated. The damping coefficient is proportional to the resistance of logistics transportation. After obtaining the real-time location of each road segment in the current transportation path, the damping coefficient sequence is generated, and the gradient of the change of the damping coefficient sequence is calculated.

[0056] First, the road network in the electronic map is processed into a graph structure, abstracting key geographical locations such as road intersections, transportation hubs, toll stations, and service area entrances / exits as physical nodes. Then, the damping coefficient of the connecting path between any two physical nodes is calculated. The damping coefficient is used to quantify the overall resistance to logistics transportation; its value is directly proportional to the difficulty of traversing the path. The damping coefficient ranges from 0 to 1. A damping coefficient closer to 1 indicates greater resistance to logistics transportation, which reduces transportation speed; a damping coefficient closer to 0 indicates less resistance. The specific calculation method for the damping coefficient will be introduced later.

[0057] After a vehicle begins its transport mission, along the candidate path, starting from the vehicle's current real-time location, the damping coefficients of all subsequent road segments are sequentially acquired and arranged according to the travel order, thus generating a damping coefficient sequence. For example, for a truck traveling from point A to point B, currently located at point C, the system will generate a list of damping coefficients for each road segment along the route from point C to point B, such as [0.8, 0.2, 0.3, 0.6, ...]. This list of damping coefficients depicts the changes in resistance during the future journey. Finally, a difference operation is performed on the list of damping coefficients to calculate the numerical differences between adjacent elements in the sequence, generating a gradient sequence. For example, a change in damping coefficient from 0.8 to 0.2 has a gradient of -0.6, indicating a significant improvement in road conditions ahead; while a change from 0.3 to 0.6 has a gradient of 0.3, suggesting potential congestion ahead.

[0058] The temporal features include the changing gradient. The prediction model is an LSTM model. An inhibition operator is embedded between the output of the forget gate of the LSTM model and the cell state at the previous time step. A slope activation unit is embedded at the output of the cell state. The inhibition operator generates a damping vector based on the damping coefficient at the current position. The cell state at the previous time step after the forget gate is modified based on the damping vector to suppress the transmission of historical inertial features. The slope activation unit maps the changing gradient to the slope factor of the activation function. When the changing gradient is negative, the slope activation unit increases the slope factor to amplify the cell state output signal. When the changing gradient is positive, the slope factor decreases to compress the cell state output signal. The prediction model uses the updated cell state to calculate the predicted default probability.

[0059] The core of the standard LSTM model is the recursive update mechanism of cell states. This mechanism, through the synergistic action of forget gates, input gates, and output gates, selectively retains historical information and incrementally integrates new information. Mathematically, this process is represented by a continuous state transition. However, in the physical scenario of logistics and transportation, there are numerous abrupt events, such as instantaneous congestion caused by traffic accidents or traffic control. When such events occur, the kinetic energy and right-of-way of vehicles approach zero in a very short time, and the correlation between their subsequent states and historical states is forcibly severed by physical conditions. Because the standard LSTM model lacks a direct perception and mandatory intervention mechanism for such external physical constraints, its state updates still follow a smooth transition pattern, causing its internal states to fail to synchronize with physical reality, thus resulting in a significant delay in recognizing risk states caused by abrupt events.

[0060] To address this issue, this embodiment embeds a suppression operator between the forget gate output of the LSTM model and the cell state at the previous time step. Its function is to forcibly correct the state inertia characteristics transmitted within the LSTM based on the damping coefficient of the current road segment, thereby enabling the model to simulate discontinuous abrupt events in the physical world.

[0061] In the specific inference process of the LSTM model, when the time step reaches the current moment, the standard forget gate is first run to determine which information from the cell state at the previous moment is retained. Then, the suppression operator maps the damping coefficient of the current segment to a damping vector through a decay function. The decay function can be a variant of a linear function or an exponential function. For example, the suppression operator uses an exponential decay function to generate the damping vector. The formula for calculating each element v of the damping vector is: v = (1-D). 2 Where D is the damping coefficient of the current road segment normalized to the interval [0, 1]. When D is close to 0, v is close to 1, producing a small inhibition effect. When D is close to 1, v approaches 0, producing a large inhibition effect.

[0062] Multiplying this damping vector element-wise with the previous cell state (after being filtered by the forgetting gate) forces the feature components of the motion trend characteristics in the previous cell state to decay to near zero when multiplied by the near-zero damping vector. It's important to clarify that this doesn't delete factual historical data such as the vehicle's completed transport progress; rather, it specifically clears the model's memory of continuing past positive trends. The direct impact on prediction is that in subsequent cell state synthesis steps, the component representing historical trends contributes almost zero. Simultaneously, the LSTM's input gate functions normally, encoding the new information of the vehicle's current slow speed or stationary state as candidate cell states. Therefore, the newly generated cell states will be almost entirely dominated by the current adverse road conditions, no longer influenced by the smoothness of previous history. Figure 2 As shown, through this mechanism, the predictive model can instantly increase the predicted probability of default, rather than increasing it slowly as in the standard model, thereby achieving rapid early warning of sudden high-risk events.

[0063] Furthermore, in standard LSTM cells, the activation function for the cell state uses a fixed-form nonlinear function, such as the hyperbolic tangent function. The derivative of this function is fixed, meaning that its compression and mapping transformation of the cell state remains constant regardless of the dynamics of the input signal. However, real-world logistics systems exhibit nonlinear dynamic response characteristics. For example, after congestion dissipates, drivers may exhibit catch-up behavior to make up for lost time, while in the early stages of congestion formation, lane-changing disturbances may be amplified through traffic wave propagation, leading to a deterioration of nonlinearity. Fixed-form activation functions cannot adjust their mapping intensity according to the dynamic changes in the input sequence, thus failing to accurately simulate and predict these complex nonlinear response behaviors. This results in deviations between the model's output and the actual dynamic evolution of the system when predicting recovery speed or deterioration trends.

[0064] Therefore, this method transforms the hyperbolic tangent function into tanh(α·x), where α is the slope factor. The slope factor is calculated based on the changing gradient, for example, the same method as mapping the damping coefficient. The slope activation unit uses a linear function or an exponential function to map the changing gradient to a slope factor between 0 and 1.

[0065] The cell states of the LSTM model are used to characterize the virtual speed of the vehicle. When the gradient is negative, the slope factor is increased, making the gradient of the activation function near the origin steeper. This causes the positive signal at the beginning of the vehicle's acceleration to be non-linearly amplified, enabling the model to predict a recovery rate higher than the normal cruising speed. A faster recovery rate means that the estimated time required to complete the remaining transportation will be shortened. Therefore, the predicted default probability calculated by the model will decrease at a faster rate, thus accurately reflecting the dynamic changes in default risk.

[0066] When the gradient is positive, the slope factor is reduced, making the activation function smoother. This means that even if there is an acceleration signal in the current input, its contribution to the cell state will be compressed. This simulates the phenomenon that the acceleration capability of a vehicle is eliminated by congestion in a deteriorating physical environment, which significantly increases the predicted default probability and provides an early warning of impending delays.

[0067] As for the other hyperparameters that need to be determined for the LSTM model, this embodiment uses existing technology to determine them. Finally, the prediction model uses the updated cell state after the dual correction by the above-mentioned suppression operator and slope activation unit to calculate a more accurate predicted default probability through the output layer.

[0068] In this embodiment, calculating the damping coefficient of the connection path between two physical nodes includes the following steps:

[0069] Obtain the baseline speed limit and real-time traffic speed of the connected path, and calculate the ratio of the difference between the baseline speed limit and the real-time traffic speed as the damping coefficient.

[0070] The difference between the baseline speed limit and the real-time traffic speed is calculated. This difference is then divided by the baseline speed limit, and the larger value between the difference and zero is taken to obtain the damping coefficient of the corresponding road segment. Therefore, the damping coefficient represents the degree of traffic congestion within the road segment.

[0071] In this embodiment, the time-series features also include speed features, progress features, and driving behavior features. The speed feature is the average speed of various road segments during transportation, and the progress feature is the ratio of the completed route length in the current transportation route to the total transportation route length. The transportation route is divided into multiple business nodes. The average time taken for a vehicle to reach each business node is calculated based on historical data. The actual time taken for a vehicle to reach each business node is obtained based on the real-time transportation data. The ratio of the actual time taken to the corresponding average time taken is used as an efficiency value. An efficiency sequence is generated based on multiple efficiency values, and the efficiency sequence is analyzed to determine the driving behavior features.

[0072] During transportation, the average speed of vehicles on various road sections is calculated, such as the average speed on highways and national roads. Progress characteristics include the length of the completed route and the total transportation route length, as well as the ratio between the two, and the time consumed in the current transportation.

[0073] The complete transportation route is evenly divided into multiple business nodes. Then, based on historical transportation data, the average time a vehicle takes to reach each business node under normal circumstances is calculated. During actual transportation, the actual time a vehicle takes to reach each business node is recorded in real time. The actual time at a particular business node is compared to the average time for that node; the ratio is the node's efficiency value. An efficiency value greater than 1 indicates that the transportation time for that segment exceeds expectations, while a value less than 1 indicates that the time is less than expected. As vehicles continuously pass through business nodes, an efficiency sequence consisting of multiple efficiency values ​​is generated.

[0074] When the efficiency values ​​of multiple consecutive nodes in the efficiency sequence are all greater than a preset deterioration threshold, these nodes are defined as delay events. The degree of delay is calculated by calculating the excess value of the delay time efficiency value exceeding the deterioration threshold. For example, if the efficiency value is 1.2 and the deterioration threshold is 1, then the excess value of the delay time efficiency value exceeding the deterioration threshold is 0.2. The excess values ​​of all delay events are summed to obtain the degree of delay. After each delay event, a recovery observation window consisting of the next N business nodes is obtained. The average efficiency value of the recovery observation window is calculated. If the average value is less than a preset recovery threshold, a recovery action is determined to have occurred. For each recovery action, its recovery strength is calculated. The calculation principle for recovery strength is the same as for the degree of delay: the value of each node's efficiency value below the recovery threshold is calculated, and this value is summed to obtain the recovery strength.

[0075] For the degree of delay and the corresponding recovery effort, the corresponding net catch-up benefit is calculated. The net catch-up benefit is the difference between the recovery effort and the degree of delay. The driving behavior characteristic is obtained by taking a time-decay weighted average of all historical net catch-up benefits. Using the above method, if a driver has frequently and successfully made up for time after a delay in the past, their driving behavior characteristic will have a large positive value; conversely, if they cannot effectively recover from a delay or the delay is aggravated, the driving behavior characteristic index will be negative or close to zero, indicating that the driver is unable to make up for the time lost due to congestion.

[0076] In this embodiment, the prediction model is equipped with a boundary loss function. When the implicit average driving speed of the prediction model exceeds the maximum physical speed, the boundary loss function outputs a penalty to correct the prediction model.

[0077] The LSTM model outputs predictions of a series of future states based on the current input data, and finally summarizes them into a predicted default probability. For example, during the internal calculation process, the LSTM model implicitly determines the expected remaining time, compares the expected remaining time with the remaining time agreed in the order, and calculates the default probability based on the difference between the two.

[0078] This prediction process implicitly estimates the future travel speed. In logistics transportation scenarios, vehicles cannot violate road speed limits. If the model predicts that a speed far exceeding the road speed limit is required to complete the journey, the predicted probability of default based on this result is invalid in reality. To extract this implicit speed, this embodiment allows the prediction model's output to directly predict the travel time required for the vehicle to complete the remaining distance across various road segment levels. Then, based on the travel time and the length of the road segment, the average travel speed on that segment is calculated. When the model's predicted average travel speed exceeds the road speed limit, the boundary loss is triggered. The calculation formula can adopt the form of a modified linear unit activation function, for example, the boundary loss function is: Loss value = (Average travel speed - Road speed limit) 2 This formula means that if the average driving speed exceeds the speed limit of a road segment, the loss value will increase sharply in the form of the square of the difference, forming a large gradient signal. This forces the model to adjust its internal weights during backpropagation to reduce the predicted driving speed of the corresponding road segment, eventually bringing it back to the speed limit range.

[0079] At this point, the total loss function of the LSTM model is the sum of the main loss function and the boundary loss function. The main loss function is the original loss function of the LSTM model, such as mean squared error or cross-entropy loss. By introducing the boundary loss function, the prediction model can not only accurately predict the probability of default, but also ensure that all its internal prediction processes and final results strictly comply with regulations, thereby generating more reliable early warnings.

[0080] In this embodiment, the analysis of the heterogeneous road network map based on graph neural networks and decision models to output the optimal scheduling scheme includes the following steps:

[0081] The initial weights of edges in the heterogeneous road network graph are determined based on the damping coefficient. The heterogeneous road network graph is then aggregated using a graph neural network to correct the initial weights to dynamic travel costs. The scheduling task is modeled as a path optimization problem with the lowest overall cost on the heterogeneous road network graph. The overall cost of the scheduling scheme is calculated based on the dynamic travel costs and additional costs of each candidate vehicle. The overall cost of different scheduling schemes is calculated using a path planning algorithm, and the scheduling scheme with the lowest overall cost is determined as the optimal scheduling scheme.

[0082] In heterogeneous road network graphs, the initial weight of an edge is set as the product of the physical length of the corresponding road segment and its damping coefficient. This not only considers distance factors but also incorporates local dynamic information reflecting the current congestion situation. The heterogeneous road network graph is then input into a graph neural network (Graph Neural Network). Through its unique multi-layer information propagation and aggregation mechanism, the Graph Neural Network enables each node and edge in the graph to perceive the global topology and real-time traffic situation of the entire road network. For example, the toll cost of a road segment depends not only on its own damping coefficient but also on complex contextual factors such as whether it connects to a congested hub downstream or whether there are many slow-moving vehicle nodes in its vicinity. By learning these complex nonlinear relationships, the Graph Neural Network's final output is not a node classification or score, but rather a correction of the initial weight of each edge, thereby generating a context-aware dynamic toll cost.

[0083] The scheduling decision problem is modeled as a path optimization problem to find the lowest comprehensive cost path in a modified heterogeneous road network graph, where the comprehensive cost is the sum of dynamic passage cost, entry potential cost, temporal potential cost, and regression potential cost. Finally, an improved A* path planning algorithm is used to solve this optimization problem. During the search process, the path cost g(n) in the evaluation function g(n) + h(n) is the comprehensive cost defined above, and h(n) is (the straight-line distance from node n to the destination) / (the maximum physical speed set by the system). The system traverses all compliant candidate scheme combinations, calculates the comprehensive cost for each combination, and determines the scheduling scheme with the lowest comprehensive cost as the optimal scheduling scheme. Specifically, based on the above description, this embodiment aims to achieve distributed transportation. Therefore, an iterative growth strategy is adopted when forming candidate schemes. Specifically, under the premise of meeting transportation timeliness and full cargo transshipment requirements, the possibility of a two-vehicle relay combination is first evaluated. If this requirement cannot be met, a three-vehicle combination is evaluated, and so on. After finding the first combination size that meets the requirements, comparisons are made only within that size (such as among all two-vehicle combinations), and the scheduling scheme with the lowest overall cost is determined as the final optimal scheduling scheme.

[0084] In this embodiment, the additional costs include entry potential energy cost, temporal potential energy cost, and regression potential energy cost. The entry potential energy cost is positively correlated with the topological complexity of the candidate vehicle entering the transfer hub from its current location. The temporal potential energy cost is positively correlated with the time difference between the candidate vehicle and the target vehicle arriving at the same transfer hub. The angle between the transfer path vector of the candidate vehicle heading to the destination of the relay goods and its original task path vector is obtained, and the regression potential energy cost is positively correlated with this angle.

[0085] The cut-in potential cost represents the topological complexity of a rescue vehicle deviating from its current main lane and entering a designated transfer hub, prioritizing the most easily accessible transfer hubs. The calculation process is as follows: First, the system calls the path planning engine to calculate the navigation path from the candidate vehicle's current location to the transfer hub. Then, the navigation path is parsed to obtain the maneuvers it includes. Based on a preset "maneuver cost table," maneuvers such as going straight, turning right, turning left, making a U-turn, entering a roundabout, or crossing an overpass are assigned complexity scores. For example, merging into a lane in the same direction has a complexity score of 0, while turning right has a score of 1. The cut-in potential cost is the sum of the complexity scores of all maneuvers on the navigation path.

[0086] Time-series potential cost is used to evaluate the collaborative efficiency of time connection between candidate vehicles and target vehicles (vehicles requiring cargo transfer) at a transit hub. Specifically, it first calculates the estimated arrival time of the candidate vehicle and the target vehicle at the transit hub, and then determines the time difference Δt. When the time difference is not equal to 0, it means that either the target vehicle or the candidate vehicle needs to wait at the transit hub, indicating a capacity vacancy and incurring waiting costs, which can be calculated using the function C=Cw*(Δt). 2 The time-series potential cost C is calculated, where Cw is a preset capacity idle cost coefficient, the value of which can be set according to factors such as vehicle operating costs and driver wages. Based on this function, the larger the time difference, the greater the calculated time-series potential cost.

[0087] The regression potential cost is used to measure the degree of disturbance that the current task causes to the candidate vehicle's original transportation plan. Specifically, the transit path vector is defined as the direction from the transit hub to the final destination of the dispersed goods, and the original path vector is defined as the direction from the candidate vehicle's current position to the destination of its original task. Then, the angle θt between these two vectors is calculated. The smaller the angle θt, the more convenient the relay transportation task is. The function P=θw*(θt) is used. 2 Calculate the regression potential cost P, where θw is the preset route coefficient, the value of which can also be set according to factors such as vehicle operating costs and driver wages.

[0088] Specifically, the calculated entry potential cost, time-series potential cost, and return potential cost need to be normalized before being added to the overall cost to eliminate dimensional differences. For example, after unifying all additional costs into time units, the entry potential cost is converted to time by multiplying it by a preset time consumption coefficient; the time-series potential cost is itself a function of time difference, and its unit is considered as equivalent time. The return potential cost is converted to the extra time incurred due to deviating from the main route by multiplying it by a detour equivalent time coefficient.

[0089] like Figure 3As shown, the present invention also provides an intelligent logistics transportation scheduling system, which is used to implement the above-described method. The system includes:

[0090] The data acquisition module obtains real-time transportation data of vehicles and extracts time-series features from the real-time transportation data.

[0091] The prediction module inputs the time-series features into the prediction model to obtain the predicted default probability, acquires the default penalty for vehicle transportation orders, and calculates the default loss based on the predicted default probability and the default penalty.

[0092] In the first planning module, if the default loss exceeds the preset intervention threshold, the route planning is re-performed based on the current vehicle location to obtain multiple candidate routes. The candidate route with the smallest default loss is located. If the default loss is less than the intervention threshold, the candidate route is used as the self-healing route, and the vehicle is switched to the self-healing route for cargo transportation.

[0093] The second planning module, if a self-healing path cannot be located, counts the number of ideal vehicles with default probabilities less than a preset threshold and no transportation tasks within a preset neighborhood centered on the current vehicle location. This ideal vehicle number is then mapped inversely to an emergency resource scarcity between 0 and 1. The emergency resource scarcity is multiplied by a preset fixed cost to obtain a scheduling threshold. If the default loss exceeds the scheduling threshold, a candidate vehicle set is generated based on load compatibility verification. A heterogeneous road network map containing vehicle nodes and hub nodes is constructed based on the road network topology and the candidate vehicle set. The heterogeneous road network map is analyzed using a graph neural network and a decision model to output the optimal scheduling scheme for distributed transportation. Furthermore, the graph neural network embeds an impedance suppression mechanism and a flow gating mechanism during node feature aggregation to respectively reduce the information weight of high-resistance paths and logically block reverse paths.

[0094] It should be understood that the above-described embodiments are merely preferred examples of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make various combinations, modifications, or equivalent substitutions of the technical features in the above solutions without departing from the core ideas and principles of the present invention. For example, the prediction model is not limited to LSTM and its variants, but can also be any other machine learning or deep learning model capable of effectively processing time-series data and performing probability prediction. The graph analysis model is not limited to a specific graph neural network structure, but can also be other graph computing techniques capable of aggregating and reasoning about heterogeneous graph information. The various thresholds, coefficients, weights, and cost calculation formulas involved in the solutions can all be set and optimized by those skilled in the art according to different business scenarios, data characteristics, and performance requirements through conventional experiments or adaptive learning methods.

[0095] The implementation order and specific combination of the various modules or steps described in the solution can also be adjusted, as long as the overall technical effect to be achieved by the present invention can be realized.

Claims

1. A logistics transportation intelligent scheduling method, characterized in that, include: Acquire real-time transportation data of vehicles and extract time-series features from the real-time transportation data; The intersections of road segments in the electronic map are defined as physical nodes. The damping coefficient of the connecting path between two physical nodes is calculated. The damping coefficient is proportional to the resistance of logistics transportation. After obtaining the real-time location of each road segment in the current transportation path, the damping coefficient sequence is generated. The change gradient of the damping coefficient sequence is calculated. The time series feature includes the change gradient. The time series feature is input into the prediction model to obtain the predicted default probability. The default penalty of the vehicle transportation order is obtained. The default loss is calculated based on the predicted default probability and the default penalty. If the breach of contract loss exceeds the preset intervention threshold, the route is replanned based on the current vehicle location to obtain multiple candidate routes. The candidate route with the smallest breach of contract loss is located. If the breach of contract loss is less than the intervention threshold, the candidate route is used as the self-healing route and the vehicle is switched to the self-healing route for cargo transportation. If a self-healing path cannot be located, then within a preset neighborhood centered on the current vehicle location, count the number of ideal vehicles with a default probability less than a preset threshold and no transportation task, and map the number of ideal vehicles inversely to an emergency resource scarcity between 0 and 1. Multiply the emergency resource scarcity by a preset fixed cost to obtain the scheduling threshold. If the default loss exceeds the scheduling threshold, a candidate vehicle set is generated based on the load compatibility check, and a heterogeneous road network map containing vehicle nodes and hub nodes is constructed based on the road network topology and the candidate vehicle set. The heterogeneous road network map is analyzed based on graph neural networks and decision models to output the optimal scheduling scheme for distributed transportation. The graph neural network embeds an impedance suppression mechanism and a flow gating mechanism during the node feature aggregation process. The impedance suppression mechanism quantifies real-time traffic conditions into the road network impedance. When aggregating neighbor node information, the impedance suppression mechanism attenuates the weight of high-impedance paths to reduce the weight of high-impedance path information. The flow gating mechanism identifies paths where information cannot flow backward by checking the adjacency matrix or edge list of the graph during the aggregation process and sets the aggregation weight of the path to zero to achieve logical blocking of the reverse path. The analysis of the heterogeneous road network graph based on graph neural networks and decision models to output the optimal scheduling scheme includes the following steps: The initial weights of edges in the heterogeneous road network graph are determined based on the damping coefficient. A graph neural network is used to aggregate the heterogeneous road network graph to correct the initial weights to dynamic travel costs. The scheduling task is modeled as a path optimization problem with the lowest comprehensive cost on the heterogeneous road network graph. The comprehensive cost of the scheduling scheme is calculated based on the dynamic travel costs and additional costs of each candidate vehicle. The decision model is an improved A* path planning algorithm. The decision model calculates the comprehensive cost of different scheduling schemes through the path planning algorithm and determines the scheduling scheme with the lowest comprehensive cost as the optimal scheduling scheme.

2. The method according to claim 1, characterized in that, The steps involved in inputting the time-series features into the prediction model to obtain the predicted default probability are as follows: The prediction model is an LSTM model. A suppression operator is embedded between the forget gate output of the LSTM model and the cell state at the previous time step. A slope activation unit is embedded at the cell state output. The suppression operator generates a damping vector based on the damping coefficient at the current position. The cell state at the previous time step after the forget gate is modified based on the damping vector to suppress the transmission of historical inertial features. The slope activation unit maps the changing gradient to the slope factor of the activation function. When the changing gradient is negative, the slope activation unit increases the slope factor to amplify the cell state output signal. When the changing gradient is positive, the slope factor decreases to compress the cell state output signal. The prediction model uses the updated cell state to calculate the predicted default probability.

3. The method according to claim 2, characterized in that, Calculating the damping coefficient of the connectivity path between two physical nodes involves the following steps: Obtain the baseline speed limit and real-time traffic speed of the connected path, and calculate the ratio of the difference between the baseline speed limit and the real-time traffic speed as the damping coefficient.

4. The method according to claim 1, characterized in that, The time-series features also include speed features, progress features, and driving behavior features. The speed feature is the average speed of various road segments during transportation. The progress feature is the ratio of the completed route length in the current transportation route to the total transportation route length. The transportation route is divided into multiple business nodes. The average time taken for a vehicle to reach each business node is calculated based on historical data. The actual time taken for a vehicle to reach each business node is obtained based on the real-time transportation data. The ratio of the actual time taken to the corresponding average time taken is used as an efficiency value. An efficiency sequence is generated based on multiple efficiency values. The efficiency sequence is analyzed to determine the driving behavior features.

5. The method according to claim 2, characterized in that, The prediction model includes a boundary loss function. When the implicit average driving speed of the prediction model exceeds the maximum physical speed, the boundary loss function outputs a penalty to correct the prediction model.

6. The method according to claim 5, characterized in that, Additional costs include entry potential cost, temporal potential cost, and regression potential cost. The entry potential cost is positively correlated with the topological complexity of the candidate vehicle entering the transfer hub from its current location. The temporal potential cost is positively correlated with the time difference between the candidate vehicle and the target vehicle arriving at the same transfer hub. The angle between the transfer path vector of the candidate vehicle to the destination of the relay cargo and its original task path vector is obtained, and the regression potential cost is positively correlated with this angle.

7. A logistics transportation intelligent scheduling system, used to implement the method as described in any one of claims 1-6, characterized in that, The system includes: The acquisition module obtains real-time transportation data of vehicles and extracts time-series features from the real-time transportation data; The prediction module inputs the time-series features into the prediction model to obtain the predicted default probability, obtain the default penalty for the vehicle transportation order, and calculate the default loss based on the predicted default probability and the default penalty. In the first planning module, if the default loss is greater than the preset intervention threshold, the route planning is re-performed based on the current vehicle location to obtain multiple candidate routes. The candidate route with the smallest default loss is located. If the default loss is less than the intervention threshold, the candidate route is used as the self-healing route and the vehicle is switched to the self-healing route for cargo transportation. The second planning module, if a self-healing path cannot be located, counts the number of ideal vehicles with default probabilities less than a preset threshold and no transportation tasks within a preset neighborhood centered on the current vehicle location. This ideal vehicle number is then mapped inversely to an emergency resource scarcity between 0 and 1. The emergency resource scarcity is multiplied by a preset fixed cost to obtain a scheduling threshold. If the default loss exceeds the scheduling threshold, a candidate vehicle set is generated based on load compatibility verification. A heterogeneous road network map containing vehicle nodes and hub nodes is constructed based on the road network topology and the candidate vehicle set. The heterogeneous road network map is analyzed using a graph neural network and a decision model to output the optimal scheduling scheme for distributed transportation. Furthermore, the graph neural network embeds an impedance suppression mechanism and a flow gating mechanism during node feature aggregation to respectively reduce the information weight of high-resistance paths and logically block reverse paths.

Images