Intelligent logistics coordination and dynamic scheduling optimization method and system based on human-vehicle-goods field full-link data

Abstract

This invention belongs to the field of logistics and supply chain management technology, specifically disclosing an intelligent logistics collaboration and dynamic scheduling optimization method and system based on full-link data of people, vehicles, goods, and yards. The method includes: real-time collection of full-link data of people, vehicles, goods, and yards to construct vehicle status data and network point status data; abstracting the logistics network into a spatiotemporal graph structure, and using a spatiotemporal graph convolutional network model to predict cargo volume distribution; constructing an optimization problem covering the future rolling time domain, the optimization problem including coupling constraints to characterize the relationship between vehicle status and travel path; using a search algorithm to solve and generate an initial scheduling scheme; dividing the initial scheduling scheme into multiple vehicle groups according to geographical regions, and initiating a negotiation mechanism within each vehicle group to generate a revised scheduling scheme; issuing and executing the scheme and rolling the time domain window. This invention achieves deep coupling of prediction and optimization, and a hybrid scheduling of centralized planning and local self-organization, improving the dynamic adaptability and system robustness of logistics scheduling.

Description

Technical Field

[0001] This invention relates to the field of logistics and supply chain management technology, and in particular to a method and system for intelligent logistics collaboration and dynamic scheduling optimization based on full-link data of people, vehicles, goods and sites. Background Technology

[0002] In the field of logistics and supply chain management technology, with the popularization of IoT, 5G, and big data technologies, enterprises are now able to collect massive amounts of data on people, vehicles, goods, and locations in real time. The logistics industry is undergoing a profound transformation from digitalization to intelligent digitalization. Currently, the core of industry development has shifted from data collection and visualization to intelligent decision-making based on real-time data. However, most existing intelligent scheduling systems rely on preset rule engines or simple integer programming models. When dealing with high-frequency, strongly coupled, and nonlinear dynamic changes, they often exhibit slow response, low optimization efficiency, and difficulty in achieving deep collaborative optimization of multiple elements.

[0003] The shortcomings of existing technologies are as follows: First, traditional scheduling schemes cannot respond in real time to dynamic events such as sudden surges in cargo volume and unexpected vehicle malfunctions, and the optimization lag leads to the rapid failure of the schemes; second, existing methods treat people, vehicles, goods, and locations as independent variables, ignoring the nonlinear coupling relationship between vehicle remaining energy, loading rate, and travel path, resulting in low resource utilization; third, traditional centralized scheduling models lack system self-correction capabilities and have weak anti-interference capabilities. This invention solves the above technical problems by constructing a closed-loop framework that couples spatiotemporal graph convolutional prediction with rolling temporal domain optimization and introducing a distributed negotiation mechanism, thereby achieving dynamic adaptation and deep collaboration in logistics scheduling.

[0004] Therefore, this invention proposes an intelligent logistics collaboration and dynamic scheduling optimization method and system based on full-link data of people, vehicles, goods and yards. Summary of the Invention

[0005] This invention provides an intelligent logistics collaboration and dynamic scheduling optimization method and system based on full-link data of people, vehicles, goods and yards. By constructing a rolling time-domain optimization framework that couples prediction and optimization in a closed loop and combining it with a regional division group negotiation mechanism, it solves the technical defects of existing technologies such as delayed response of scheduling schemes, insufficient multi-element collaboration and weak anti-interference ability of the system, and realizes dynamic self-adaptation and deep resource collaboration of logistics scheduling.

[0006] This invention provides an intelligent logistics collaboration and dynamic scheduling optimization method based on full-link data of people, vehicles, goods, and yards, comprising the following steps: Step S1: Collect real-time data on the entire chain of people, vehicles, goods, and yards to construct vehicle status data and network status data; Step S2: Abstract the logistics network into a spatiotemporal graph structure, and use the spatiotemporal graph convolutional network model to process the network point status data and dynamically predict the future cargo volume distribution in the time domain; Step S3: At each decision moment, construct an optimization problem covering the future rolling time domain. The constraints of the optimization problem include coupling constraints, which are used to characterize the relationship between the vehicle state and the driving path. Step S4: Use a search algorithm to solve the optimization problem and generate an initial scheduling scheme; Step S5: Divide the initial scheduling plan into multiple vehicle groups according to geographical regions, and initiate a negotiation mechanism within each vehicle group to generate a revised scheduling plan; Step S6: Send the revised scheduling plan to the corresponding vehicle for execution, and roll the time domain window forward by one step, repeating steps S1 to S5.

[0007] Furthermore, vehicle status data includes real-time vehicle location, remaining vehicle loading capacity, remaining vehicle energy, and remaining driver working hours, while branch status data is the branch's cargo volume status data.

[0008] Furthermore, in the step of dynamically predicting the distribution of goods volume at network points in the future time domain using a spatiotemporal graph convolutional network model, the spatiotemporal graph convolutional network model is trained using a prediction-optimization guided loss function. The prediction-optimization guided loss function includes a prediction error term and a scheduling cost gap term. The prediction error term is the error between the predicted goods volume value and the historical actual goods volume value, and the scheduling cost gap term is the cost difference between the scheduling scheme generated based on the predicted goods volume value and the optimal scheduling scheme generated by posterior calculation based on the historical actual goods volume value.

[0009] Furthermore, in the step of dynamically predicting the distribution of goods volume at network points in the future time domain using a spatiotemporal graph convolutional network model, the spatiotemporal graph convolutional network model includes a dynamic attention module. The dynamic attention module generates attention weights for each network point based on time features, weather features, and holiday features. The attention weights are used to dynamically adjust the contribution of different network points in the goods volume prediction.

[0010] Furthermore, the objective function of the optimization problem includes travel distance cost, energy consumption cost, unused capacity penalty, and network backlog penalty. The constraints of the optimization problem also include vehicle capacity constraints, vehicle driver working hours constraints, and task spatiotemporal consistency constraints.

[0011] Furthermore, the coupling constraints are in the form of nonlinear functions, used to characterize the nonlinear relationship between vehicle load rate, vehicle remaining energy, and vehicle travel path. The coupling constraints are constructed based on the physical energy consumption laws during vehicle travel, including joint modeling of the nonlinear physical relationship between load rate and rolling resistance, the electrochemical characteristics of battery discharge efficiency decreasing with remaining charge, and the physical influence of road slope on vehicle traction. The functional expression of the coupling constraints is: in, Based on energy consumption, This represents the current load capacity of the vehicle. This is the vehicle's maximum loading capacity. For nonlinear coefficients greater than 1, The residual energy penalty function, For slope factor, As driving behavior factors, The usable energy after deducting reserved energy; When the function value of the coupling constraint is greater than the preset threshold, it is determined that the current vehicle and driving path allocation is not feasible.

[0012] Furthermore, the search algorithm is a multi-operator search algorithm, which includes multiple destruction operators and multiple repair operators. The steps for solving the optimization problem using the multi-operator search algorithm include: The current solution is iteratively optimized using multiple destruction operators and multiple repair operators. The destruction operators are used to remove some tasks from the current solution, and the repair operators are used to reinsert the removed tasks into the solution. The selection probability of each operator is dynamically adjusted based on its historical performance, enabling the search process to converge adaptively. The specific method for dynamically adjusting the selection probability of each operator is as follows: maintain a weight value for each operator, score the operators used based on the improvement of the objective function generated in each iteration, periodically calculate the cumulative score of each operator, update the weight value of each operator based on the cumulative score, and then convert the weight value into a selection probability through normalization.

[0013] Furthermore, the negotiation mechanism is a distributed negotiation mechanism, and the steps to initiate the distributed negotiation mechanism within each vehicle group include: Within each vehicle group, a coordinating node is elected through a heartbeat mechanism. The coordinating node collects scheduling adjustment proposals from each vehicle in the vehicle group. Each vehicle calculates the marginal contribution of its scheduling adjustment proposal to the local objective function of the vehicle group; The optimal scheduling adjustment proposal is selected based on the marginal contribution value, and a consensus is reached within the vehicle group; If consensus is not reached after the maximum number of iterations in the negotiation rounds, the coordinating node will directly select the proposal with the best improvement effect as the mandatory consensus solution based on the marginal contribution value and issue it for execution. When resource conflicts arise in the negotiation results of multiple vehicle groups, a cross-group arbitration mechanism is initiated. The upper-level central node determines the final allocation of conflicting resources based on the global objective function value and issues arbitration instructions to the relevant groups.

[0014] Furthermore, the time domain length of the rolling time domain is a preset duration, and the rolling step size is less than the time domain length.

[0015] This invention provides an intelligent logistics collaboration and dynamic scheduling optimization system based on full-link data of people, vehicles, goods, and yards, comprising: The data acquisition module is used to collect real-time data on the entire process of people, vehicles, goods, and yards. The status construction module is used to construct vehicle status data and network status data based on the full-link data of people, vehicles, goods and yards collected by the data acquisition module; The cargo volume prediction module is used to abstract the logistics network into a spatiotemporal graph structure. It uses a spatiotemporal graph convolutional network model to process the network point state data constructed by the state construction module and dynamically predict the cargo volume distribution in the future time domain. The rolling time domain optimization module is used to construct an optimization problem covering the future rolling time domain at each decision time, and to solve the optimization problem using a multi-operator search algorithm to generate an initial scheduling scheme. The constraints of the optimization problem include coupling constraints, which are used to characterize the relationship between vehicle state and driving path. The distributed negotiation module is used to divide the initial scheduling scheme generated by the rolling time-domain optimization module into multiple vehicle groups according to geographical regions, and to start a negotiation mechanism within each vehicle group to generate a revised scheduling scheme. The instruction issuance module is used to issue the revised scheduling scheme generated by the distributed negotiation module to the corresponding vehicles for execution.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: By constructing a closed-loop framework that couples spatiotemporal graph convolutional prediction with rolling temporal optimization, prediction and optimization are deeply integrated. Furthermore, by introducing regional partitioning and group negotiation mechanisms, a hybrid scheduling system combining centralized planning and local self-organization is achieved, resulting in the following benefits: High dynamic adaptability—the system can perceive dynamic events such as sudden changes in cargo volume and vehicle malfunctions in real time and adaptively adjust the scheduling scheme to avoid scheme failure; Deep resource collaboration—through coupling constraints, vehicle loading rate, remaining energy, and driving path are nonlinearly correlated, achieving global optimization of multiple elements such as people, vehicles, cargo, and sites, significantly improving resource utilization; Strong system robustness—the distributed group negotiation mechanism partially decentralizes decision-making power, enabling the system to respond quickly to local disturbances and significantly enhancing its anti-interference capabilities.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is the core flowchart of the intelligent logistics collaboration and dynamic scheduling optimization method in the embodiments of the present invention; Figure 2 This is a schematic diagram of the coupling constraint function in an embodiment of the present invention; Figure 3 This is a flowchart of the distributed negotiation mechanism in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] refer to Figure 1 , Figure 2 , Figure 3 This invention provides an embodiment of an intelligent logistics collaboration and dynamic scheduling optimization method based on full-link data of people, vehicles, goods, and yards, comprising the following steps: Step S1: Collect real-time data on the entire chain of people, vehicles, goods, and yards to construct vehicle status data and network status data; Step S2: Abstract the logistics network into a spatiotemporal graph structure, and use the spatiotemporal graph convolutional network model to process the network point status data and dynamically predict the future cargo volume distribution in the time domain; Step S3: At each decision moment, construct an optimization problem covering the future rolling time domain. The constraints of the optimization problem include coupling constraints, which are used to characterize the relationship between the vehicle state and the driving path. Step S4: Use a search algorithm to solve the optimization problem and generate an initial scheduling scheme; Step S5: Divide the initial scheduling plan into multiple vehicle groups according to geographical regions, and initiate a negotiation mechanism within each vehicle group to generate a revised scheduling plan; Step S6: Send the revised scheduling plan to the corresponding vehicle for execution, and roll the time domain window forward by one step, repeating steps S1 to S5.

[0022] In this embodiment, vehicle status data includes the vehicle's real-time location, remaining loading capacity, remaining energy, and remaining driver hours. Network status data refers to the network's cargo volume status. For example, vehicle V001's real-time location is 116.397 degrees East longitude and 39.908 degrees North latitude, with a remaining loading capacity of 4.2 cubic meters, remaining energy of 98 kilowatt-hours, and 8 hours of remaining driver hours; the network's central warehouse currently has 800 orders awaiting shipment, and the predicted cargo volume increase is 35% in the next 2 hours.

[0023] In this embodiment, the spatiotemporal graph convolutional network model treats each point in the logistics network as a node in the spatiotemporal graph and the logistics routes between points as edges. It uses graph convolution operations to capture the mutual influence relationships of cargo volume changes between adjacent points and temporal convolution operations to capture the periodic patterns of cargo volume changes over time. For example, when the cargo volume in the central warehouse increases, the model can learn the spatial transmission relationship that this upward trend will drive a synchronous increase in the cargo volume of surrounding micro-warehouses. Simultaneously, combined with the temporal periodic pattern of cargo volume surges during the morning peak hours, it outputs the predicted cargo volume values ​​for each point in the next two hours.

[0024] In this embodiment, the spatiotemporal graph convolutional network model is trained using a prediction-optimization guided loss function, which includes both a prediction error term and a scheduling cost gap term. The prediction error term minimizes the deviation between the predicted cargo volume and the historical actual cargo volume, while the scheduling cost gap term minimizes the cost difference between the scheduling scheme generated based on the predicted values ​​and the optimal scheduling scheme generated posteriorly based on the actual cargo volume. Through this training method, the model not only pursues the accuracy of the predicted values ​​but also the value of the prediction results for subsequent scheduling optimization.

[0025] In this embodiment, the spatiotemporal graph convolutional network model also includes a dynamic attention module. This module generates attention weights for each location based on current time features, weather features, and holiday features, which are used to dynamically adjust the contribution of different locations in freight volume prediction. For example, during Monday morning rush hour, the model automatically increases the attention weight for locations in commercial areas, significantly increasing their predicted freight volume; during off-peak hours on weekends, it increases the attention weight for locations in residential areas.

[0026] In this embodiment, rolling time domain refers to optimizing only a fixed-length time domain window at each decision point, rather than optimizing all tasks indefinitely. For example, if the time domain length is set to 3 hours, the system will only optimize delivery tasks between 8:00 AM and 11:00 AM at 8:00 AM. When the time progresses to 8:15 AM, the window rolls to the period between 8:15 AM and 11:15 AM, and optimization is performed again. This rolling optimization method allows the system to continuously absorb the latest real-time data and constantly revise subsequent scheduling plans.

[0027] In this embodiment, coupling constraints are used to simultaneously consider the nonlinear relationship between vehicle load factor, remaining energy, and driving route. For example, when a vehicle with 42% battery power and a 26% load factor is assigned to a 60-kilometer suburban route, the system calculates through coupling constraints that the base energy consumption is 12 kWh, the load factor is 1.07, the remaining energy factor is 1.38, the gradient factor is 1.0, the driving behavior factor is 0.93, and the total energy consumption is 16.5 kWh. However, the available energy after deducting reserved energy is only 16 kWh. Therefore, the allocation is deemed infeasible, and the system automatically reassigns the task to a vehicle with more battery power.

[0028] In this embodiment, the search algorithm is a multi-operator search algorithm, comprising multiple destructive operators and multiple repair operators. Destructive operators are used to remove some tasks from the current solution, such as random removal, cost-based removal, or cluster removal; repair operators are used to reinsert the removed tasks into the solution, such as greedy insertion or regret-based insertion. The system dynamically adjusts the probability of each operator being selected based on its past performance, enabling the search process to adaptively converge to a high-quality solution.

[0029] In this embodiment, the negotiation mechanism is a distributed negotiation mechanism. Vehicles are divided into multiple groups based on geographical location, and a coordinating node is elected within each group through a heartbeat mechanism. For example, 20 vehicles in the Northeast region form a group. Each vehicle in the group can propose a scheduling adjustment, such as transferring three orders from vehicle V012 to vehicle V015. Each vehicle calculates the marginal contribution of the proposal to the total mileage of the group. After multiple rounds of asynchronous negotiation, the proposal with the greatest improvement is selected to reach a consensus, generating a revised local scheduling scheme that reduces the total mileage of the group by 8.2 kilometers and the total energy consumption by 6.4 kilowatt-hours.

[0030] In this embodiment, when an abnormality in the status of transportation resources, a change in tasks, or a sudden change in road conditions is detected, the system immediately initiates a negotiation mechanism within the vehicle group to make rapid adjustments. For example, when vehicle V012 experiences a motor failure while driving, the system automatically reports the failure status, and the affected group completes local negotiation within 30 seconds, reassigning V012's remaining tasks to other vehicles in the group and automatically notifying the customer of a 20-minute delay in delivery time. The entire process does not require triggering a global rescheduling.

[0031] Furthermore, vehicle status data includes real-time vehicle location, remaining vehicle loading capacity, remaining vehicle energy, and remaining driver working hours, while branch status data is the branch's cargo volume status data.

[0032] For example, the real-time location of vehicle V001 is 116.397 degrees east longitude and 39.908 degrees north latitude, indicating that the vehicle is currently near the central warehouse; the remaining loading capacity of the vehicle is 4.2 cubic meters, meaning that the vehicle can currently load 4.2 cubic meters of goods, with a maximum loading capacity of 6 cubic meters, i.e., the current loading rate is 30%; the remaining energy of the vehicle is 98 kWh, meaning that the vehicle currently has 98% of its total energy of 100 kWh remaining, possessing long-distance delivery capabilities; the driver's remaining working hours are 8 hours, meaning that the driver can work 8 hours that day, complying with labor regulations. The network central warehouse's inventory status data includes current inventory, number of orders awaiting pickup, and number of orders awaiting delivery. For example, if there are currently 800 orders awaiting outbound shipment and 500 orders awaiting delivery, with an inventory pressure index of 0.75, it indicates that the network has significant inventory pressure and delivery vehicles need to be prioritized for evacuation.

[0033] Furthermore, in the step of dynamically predicting the distribution of goods volume at network points in the future time domain using a spatiotemporal graph convolutional network model, the spatiotemporal graph convolutional network model is trained using a prediction-optimization guided loss function. The prediction-optimization guided loss function includes a prediction error term and a scheduling cost gap term. The prediction error term is the error between the predicted goods volume value and the historical actual goods volume value, and the scheduling cost gap term is the cost difference between the scheduling scheme generated based on the predicted goods volume value and the optimal scheduling scheme generated by posterior calculation based on the historical actual goods volume value.

[0034] In this embodiment, the spatiotemporal graph convolutional network model is trained using a prediction-optimization guided loss function, which includes a prediction error term and a scheduling cost gap term. The prediction error term represents the error between the predicted cargo volume and the historical actual cargo volume, ensuring that the model can accurately fit historical data patterns. The scheduling cost gap term represents the cost difference between the scheduling scheme generated based on the predicted cargo volume and the optimal scheduling scheme generated through posterior calculation based on the historical actual cargo volume, guiding the model to learn the most valuable prediction results for subsequent optimization. Taking a city's logistics network as an example, if the model simply pursues minimizing the prediction error, it might adjust the predicted cargo volume of a certain branch from 100 orders to 101 orders to reduce the error by 1%. However, this adjustment might lead to an additional vehicle being allocated to that branch in the scheduling scheme, increasing the total transportation cost by 500 yuan. After training with the prediction-optimization guided loss function, the model will prioritize learning the prediction pattern that minimizes scheduling costs, i.e., it would rather retain 1% prediction error than waste transportation capacity due to prediction deviation. This training method enables deep coupling between the prediction model and the downstream optimization task, fundamentally solving the problems of separation between prediction and optimization and error propagation amplification in traditional technologies.

[0035] To further illustrate the superiority of the prediction-optimization guided loss function, this invention provides a set of comparative experimental data. In a city's logistics network, historical data from 30 consecutive days were selected, and spatiotemporal graph convolutional network models were trained using the following three methods: Method 1 (Traditional Prediction-Oriented): Only the prediction error term is used as the loss function; Method 2 (the method of this invention): Simultaneously use the prediction error term and the scheduling cost gap term as the loss function; Method 3 (Ideal Posterior): Calculate the optimal scheduling scheme using historical real cargo volume values ​​as the theoretical upper bound.

[0036] Experimental results show that the model trained using Method 1 has an average absolute percentage error of 8.7% between its predicted cargo volume and the actual value, but the cost difference between the scheduling scheme generated based on its predictions and the actual optimal scheduling scheme is as high as 12.3%. The model trained using Method 2 (the method of this invention) has a slightly higher prediction error of 9.2%, but the scheduling cost difference is significantly reduced to 3.8%, and the difference with the ideal posterior scheme decreases from 12.3% to 3.8%, representing a relative improvement of 69%. This result verifies the core idea of ​​this invention: the purpose of prediction is optimization, not simply pursuing numerical accuracy. By introducing a scheduling cost difference term, the model learns "which prediction deviations have a small impact on scheduling decisions and which deviations have a large impact," thus achieving an optimal balance between prediction accuracy and decision value.

[0037] Furthermore, in the step of dynamically predicting the distribution of goods volume at network points in the future time domain using a spatiotemporal graph convolutional network model, the spatiotemporal graph convolutional network model includes a dynamic attention module. The dynamic attention module generates attention weights for each network point based on time features, weather features, and holiday features. The attention weights are used to dynamically adjust the contribution of different network points in the goods volume prediction.

[0038] In this embodiment, the spatiotemporal graph convolutional network model includes a dynamic attention module. This module generates attention weights for each network point based on time features, weather features, and holiday features. These attention weights are used to dynamically adjust the contribution of different network points to cargo volume prediction. Specifically, time features include the current hour of the day and whether it is a weekday or weekend; for example, 8:00 AM to 10:00 AM is the morning peak period, and 5:00 PM to 7:00 PM is the evening peak period. Weather features include the impact of different weather types such as sunny, rainy, snowy, and foggy conditions on logistics and delivery; for example, rainy or snowy weather can lead to delivery delays and cargo backlogs. Holiday features include the impact of statutory holidays and shopping festivals on consumer behavior; for example, during the "Double Eleven" shopping festival, cargo volume at commercial area network points surges. Taking a city's logistics network as an example, during Monday morning rush hour, the dynamic attention module automatically calculates the attention weight for commercial area outlets as 0.8 and for residential area outlets as 0.2, making the model focus more on the freight volume trends of commercial area outlets during prediction. During weekend afternoon hours, the dynamic attention module automatically adjusts the weight of commercial area outlets to 0.3 and the weight of residential area outlets to 0.7, making the model focus more on the delivery needs of residential area outlets. Through this dynamic attention mechanism, the model can adaptively adjust its prediction focus according to the spatiotemporal context, significantly improving the accuracy of freight volume prediction and its adaptability to real-world scenarios.

[0039] Furthermore, the objective function of the optimization problem includes travel distance cost, energy consumption cost, unused capacity penalty, and network backlog penalty. The constraints of the optimization problem also include vehicle capacity constraints, vehicle driver working hours constraints, and task spatiotemporal consistency constraints.

[0040] In this embodiment, the objective function of the optimization problem includes travel distance cost, energy consumption cost, unused capacity penalty, and network backlog penalty. Travel distance cost measures the fuel consumption, vehicle wear and tear, and time costs incurred by the vehicle's mileage, for example, set at 2.5 yuan per kilometer; energy consumption cost measures the cost of electricity or fuel consumed by the vehicle, for example, set at 0.8 yuan per kilowatt-hour, taking into account peak and off-peak electricity prices; unused capacity penalty encourages vehicles to operate at full capacity to avoid wasting transport capacity, for example, set at 50 yuan per vehicle per hour for unused capacity; network backlog penalty incentivizes the system to prioritize network points with severe backlogs to avoid prolonged cargo delays, for example, set at 10 yuan per order per hour for backlog. By weighted summing of these four costs, the system can comprehensively balance multiple objectives such as transportation efficiency, energy consumption, capacity utilization, and customer satisfaction.

[0041] In this embodiment, the constraints of the optimization problem also include vehicle capacity constraints, driver working hour constraints, and task spatiotemporal consistency constraints. Vehicle capacity constraints ensure that the total load of tasks assigned to a vehicle does not exceed the vehicle's maximum loading capacity. For example, if the maximum loading capacity is 6 cubic meters, the total volume of all goods assigned to that vehicle must not exceed 6 cubic meters. Driver working hour constraints ensure that the driver's working hours do not exceed the legal working hour limit and the driver's remaining working hours. For example, if a driver has worked 4 hours that day and has 4 hours remaining, the estimated travel time assigned to that vehicle must not exceed 4 hours. Task spatiotemporal consistency constraints ensure that multiple tasks performed by the same vehicle are consecutively feasible in time and space. For example, after completing a delivery task at point A, the vehicle's arrival time at point B must be later than the completion time at point A, and it must meet the task time window requirements of point B, without any task sequence disorder or time conflict.

[0042] Furthermore, the coupling constraints are in the form of nonlinear functions, used to characterize the nonlinear relationship between vehicle load rate, vehicle remaining energy, and vehicle travel path. The coupling constraints are constructed based on the physical energy consumption laws during vehicle travel, including joint modeling of the nonlinear physical relationship between load rate and rolling resistance, the electrochemical characteristics of battery discharge efficiency decreasing with remaining charge, and the physical influence of road slope on vehicle traction. The functional expression of the coupling constraints is: in, Based on energy consumption, This represents the current load capacity of the vehicle. This is the vehicle's maximum loading capacity. For nonlinear coefficients greater than 1, The residual energy penalty function, For slope factor, As driving behavior factors, The usable energy after deducting reserved energy; When the function value of the coupling constraint is greater than the preset threshold, it is determined that the current vehicle and driving path allocation is not feasible.

[0043] In this embodiment, the coupling constraint is in the form of a nonlinear function, used to characterize the nonlinear correlation between vehicle load rate, vehicle remaining energy, and vehicle travel path. The coupling constraint is constructed based on the physical energy consumption patterns during vehicle operation, including joint modeling of the nonlinear physical relationship between load rate and rolling resistance, the electrochemical characteristics of battery discharge efficiency decreasing with remaining charge, and the physical influence of road slope on vehicle traction. Specifically, the functional expression of the coupling constraint is: f = Σ[base energy consumption multiplied by load rate factor multiplied by remaining energy penalty factor multiplied by slope factor multiplied by driving behavior factor] minus available energy. The base energy consumption depends on the length of the driving route and the vehicle type. For example, the base energy consumption of an electric light truck is 0.2 kWh per kilometer. The loading rate factor is 1 plus the k-th power of the loading rate, where k is a non-linear coefficient greater than 1. For example, if the value is 2, the loading rate factor is 1.64 when the loading rate is 80%, meaning that a high loading rate will lead to a non-linear increase in energy consumption. The remaining energy penalty function is used to reflect the phenomenon of a sharp increase in energy consumption when the remaining energy is low. For example, when the remaining energy is less than 30%, the penalty factor increases linearly from 1 to 3. The gradient factor is used to reflect the amplification effect of uphill sections on energy consumption. For example, a 5-degree gradient corresponds to a factor of 1.2. The driving behavior factor is used to reflect the impact of the driver's driving habits on energy consumption. Aggressive driving corresponds to a factor greater than 1, and stable driving corresponds to a factor less than 1. Available energy is the remaining energy of the vehicle after deducting the reserved energy for air conditioning, refrigeration, and insulation.

[0044] It should be noted that directly solving the aforementioned nonlinear coupling constraints in large-scale logistics scheduling scenarios may lead to non-convexity, difficulty in solving, or even non-convergence of the mixed-integer programming problem. To balance the model's expressive accuracy and solution efficiency, this invention further provides linearization and relaxation strategies for the coupling constraints.

[0045] Specifically, the load factor The piecewise linearization method approximates the load rate interval [0,1] by dividing it into M equal-length subintervals. Within each subinterval, a linear function replaces the original nonlinear function, and a 0-1 variable is introduced to indicate the current load rate segment. This transforms the original nonlinear constraint into a mixed-integer linear constraint. (For the remaining energy penalty function...) and slope factor A similar piecewise linearization process is used.

[0046] Furthermore, when the problem size exceeds a preset threshold, the system relaxes coupling constraints into a penalty term added to the objective function: a penalty coefficient much larger than the normal cost (e.g., set to 10 times the normal energy consumption cost) is applied to allocation schemes that violate coupling constraints, enabling the solver to automatically avoid infeasible allocations during the search process while ensuring the numerical stability of the solution process. Actual engineering verification shows that, after adopting the above linearization and relaxation strategies, a scheduling problem involving 200 vehicles and 1500 tasks can obtain a feasible solution within 60 seconds, with a coupling constraint violation rate of less than 0.5%.

[0047] In practical engineering verification, after adopting the coupling constraints of this invention, 100 electric delivery vehicles of a logistics company did not experience a single breakdown due to inaccurate power estimation during a three-month trial operation. In contrast, the breakdown rate of the control group (using the traditional linear energy consumption model) was 3.2 times per 100 vehicles per month. This robust technical effect directly proves that the coupling constraints accurately characterize the physical energy consumption patterns, rather than merely focusing on cost optimization.

[0048] It is important to emphasize that the coupling constraints of this invention are not a simple mathematical model of economic benefits, but rather a technical characterization of the physical energy consumption patterns during vehicle operation. Specifically, The term characterizes the nonlinear physical relationship between vehicle load rate and rolling resistance (the higher the load rate, the greater the tire deformation and the exponential increase in energy consumption). This term characterizes the electrochemical property that the battery discharge efficiency decreases as the remaining charge decreases; The term characterizes the physical effect of changes in gravitational potential energy on vehicle traction. The terms characterize the physical perturbation of driving behavior on energy conversion efficiency. These parameters together constitute a complete model of the vehicle's "oil / electricity-power-motion" physical link. Therefore, the coupling constraints of this invention constitute a technical intervention in the physical world, rather than a simple economic management rule.

[0049] Taking an electric light truck as an example, its maximum loading capacity is 6 cubic meters, the current loading is 4.8 cubic meters (80% loading rate), the remaining power is 42 kWh, the reserved air conditioning energy consumption is 5 kWh, and the available energy is 37 kWh. If the system attempts to assign this vehicle to a suburban route with a total length of 60 kilometers, including a 5-kilometer uphill section with a gradient of 5 degrees, and a driving behavior factor of 0.93, the calculation process is as follows: the base energy consumption is 60 kilometers multiplied by 0.2 kWh, which equals 12 kWh per kilometer; the loading rate factor is 1 plus 0.8 squared, which equals 1.64; the remaining energy penalty factor is 1; the gradient factor is 1 on flat roads and 1.2 on uphill sections; the driving behavior factor is 0.93. After segmented calculation, the total energy consumption is approximately 16.5 kWh. Since 16.5 kWh is less than the available energy of 37 kWh, the coupling constraint function value is negative, and the assignment is deemed feasible. Conversely, if the vehicle's remaining battery power is only 20 kWh, and the usable energy after deducting the reserve is 15 kWh, then 16.5 kWh is greater than 15 kWh, the coupling constraint function value is positive, and the allocation is determined to be infeasible. The system automatically reassigns the task to a vehicle with more battery power. Through this nonlinear coupling constraint, the present invention achieves deep coordination between vehicle loading status, energy status, and travel path, effectively avoiding the risk of low-battery vehicles being assigned to heavy-load long-distance routes, and significantly improving the feasibility of the scheduling scheme and resource utilization efficiency.

[0050] Furthermore, the search algorithm is a multi-operator search algorithm, which includes multiple destruction operators and multiple repair operators. The steps for solving the optimization problem using the multi-operator search algorithm include: The current solution is iteratively optimized using multiple destruction operators and multiple repair operators. The destruction operators are used to remove some tasks from the current solution, and the repair operators are used to reinsert the removed tasks into the solution. The selection probability of each operator is dynamically adjusted based on its historical performance, enabling the search process to converge adaptively. The specific method for dynamically adjusting the selection probability of each operator is as follows: maintain a weight value for each operator, score the operators used based on the improvement of the objective function generated in each iteration, periodically calculate the cumulative score of each operator, update the weight value of each operator based on the cumulative score, and then convert the weight value into a selection probability through normalization.

[0051] In this embodiment, the search algorithm is a multi-operator search algorithm, which includes multiple destructive operators and multiple repair operators. Destructive operators are used to remove some tasks from the current solution, including random removal operators, cost-based removal operators, and clustering removal operators. The random removal operator randomly selects a specified number of tasks from all tasks to increase solution diversity; the cost-based removal operator prioritizes removing tasks with the highest marginal cost, i.e., tasks with the greatest cost contribution in the current allocation scheme; the clustering removal operator selects a group of geographically close tasks for removal to break the local clustering allocation structure. Repair operators are used to re-insert the removed tasks into the solution, including greedy insertion operators and regret-value insertion operators. The greedy insertion operator inserts each task into the position that minimizes its cost increase; the regret-value insertion operator comprehensively considers the current insertion cost and the potential loss of abandoning insertion, prioritizing the insertion of tasks whose subsequent costs would increase significantly if not inserted immediately.

[0052] The steps for solving the optimization problem using a multi-operator search algorithm include: First, generating an initial solution using a greedy algorithm; then, in each iteration, selecting a destructive operator and a repair operator based on the current selection probabilities of each operator using a roulette wheel. The destructive operator removes some tasks from the current solution, and the repair operator reinserts the removed tasks, generating a new solution; calculating the objective function value of the new solution, updating the optimal solution if it is better than the global optimum, and accepting it if it is better than the current solution or satisfies the simulated annealing acceptance criterion; scoring the selected destructive and repair operators based on their performance in the current iteration, with higher-performing operators receiving higher scores; and dynamically adjusting the selection probabilities of each operator based on their cumulative scores at regular intervals, increasing the probability of selecting operators with high scores and decreasing the probability of selecting operators with low scores. Through this adaptive mechanism, the search process can automatically select the most effective combination of operators for the current problem during iterations, achieving a balance between exploration and utilization, and quickly converging to a high-quality solution. Taking a scheduling optimization as an example, the initial solution had a total cost of 8,500 yuan. After 800 iterations, the system found that the clustering removal operator combined with the regret value insertion operator performed better than other combinations. Its selection probability gradually increased from the initial 0.2 to 0.45. Finally, the optimized solution with a total cost of 7,200 yuan was obtained, which was more than 15% lower than the initial solution.

[0053] The above-mentioned "dynamically adjusting the selection probability of operators based on the historical performance of each operator" is implemented through the following logical flow: The first step is to initialize the operator weights: assign the same initial weight value to each destruction operator and each repair operator. The weight value is used to characterize the degree to which the operator is selected in subsequent searches.

[0054] The second step is to perform iterations and record operator performance: In each iteration, record the combination of destruction and repair operators selected in the current round; calculate the improvement of the objective function of the new solution generated in this iteration relative to the solution in the previous round. A positive improvement indicates that the new solution is better, and a negative improvement indicates that the new solution is worse; according to the magnitude of the improvement, assign a score to the combination of operators used in this round according to the preset scoring rules, where the combination of operators that produces positive improvement receives a positive score, and the combination of operators that produces negative improvement receives a zero or negative score.

[0055] The third step is to periodically update the operator scores: After completing a preset number of iterations (e.g., every 100 rounds), calculate the cumulative score of each operator in the current statistical period; the cumulative score is the sum of the scores obtained by the operator in all selected rounds.

[0056] The fourth step is to update the operator weights based on the cumulative score: update the current weight value of each operator to the weighted sum of its original weight and cumulative score, where a certain proportion of the original weight is retained (for example, retaining 80% of the historical weight and adding 20% ​​of the periodic score), so that the weight update can reflect recent performance while maintaining a certain degree of stability.

[0057] Fifth step, update the selection probability: normalize the weight values ​​of all operators, that is, the selection probability of each operator is equal to the weight value of the operator divided by the sum of the weight values ​​of all operators; the higher the weight of the operator, the greater the probability of it being selected by the roulette wheel in subsequent iterations.

[0058] Step 6, repeat execution: Repeat steps 2 to 5 above to continuously tilt the operator selection probability in the search process towards the operator with better performance, thereby achieving adaptive convergence of the search process.

[0059] Through the above logical process, the system can automatically identify and prioritize the most effective combination of operators for the current problem during the iteration process. Taking a scheduling optimization as an example, in the initial stage, the weight of each operator is 1.0, and the selection probability is equal. After 200 iterations, the clustering removal operator has the highest cumulative score, its weight increases from 1.0 to 2.3, and its selection probability increases from 0.2 to 0.45; while the random removal operator has a lower score, its weight decreases to 0.6, and its selection probability decreases to 0.12. This adaptive mechanism enables the algorithm to achieve a balance between exploration and utilization, effectively improving the solution efficiency and the quality of the solution.

[0060] Furthermore, the negotiation mechanism is a distributed negotiation mechanism, and the steps to initiate the distributed negotiation mechanism within each vehicle group include: Within each vehicle group, a coordinating node is elected through a heartbeat mechanism. The coordinating node collects scheduling adjustment proposals from each vehicle in the vehicle group. Each vehicle calculates the marginal contribution of its scheduling adjustment proposal to the local objective function of the vehicle group; The optimal scheduling adjustment proposal is selected based on the marginal contribution value, and a consensus is reached within the vehicle group; If consensus is not reached after the maximum number of iterations in the negotiation rounds, the coordinating node will directly select the proposal with the best improvement effect as the mandatory consensus solution based on the marginal contribution value and issue it for execution. When resource conflicts arise in the negotiation results of multiple vehicle groups, a cross-group arbitration mechanism is initiated. The upper-level central node determines the final allocation of conflicting resources based on the global objective function value and issues arbitration instructions to the relevant groups.

[0061] In this embodiment, the negotiation mechanism is a distributed negotiation mechanism, and the specific steps for initiating the distributed negotiation mechanism within each vehicle group are as follows.

[0062] First, a coordinating node is elected within each vehicle group using a heartbeat mechanism. The heartbeat mechanism involves each vehicle periodically sending a liveness signal to other vehicles. If a vehicle fails to receive a heartbeat signal from another vehicle multiple times consecutively, it is considered offline. When there is no coordinating node in the group or the coordinating node is offline, an election is initiated, and the vehicle receiving more than half the votes becomes the new coordinating node. The coordinating node is responsible for maintaining the group state, collecting proposals, and promoting consensus. For example, in a group of 20 vehicles in the Northeast region, vehicle V015, with sufficient battery power and strong computing capabilities, is elected as the coordinating node through the heartbeat mechanism.

[0063] Secondly, the coordinating node collects scheduling adjustment proposals from each vehicle within the vehicle group. Each vehicle, based on its real-time status and local information, proposes improvements to the group's scheduling plan. The proposal includes the affected order identifier, the vehicle currently assigned to the order, the target vehicle for the proposed transfer, and the expected improvement effect. For example, vehicle V012, due to low battery and nearing full capacity, proposes transferring three orders in area A to vehicle V015; vehicle V015 accepts the transfer but requires minor adjustments to its original route.

[0064] Next, each vehicle calculates the marginal contribution of its scheduling adjustment proposal to the local objective function of the vehicle group. The local objective function can be a weighted sum of the total mileage, total energy consumption, or total completion time of all vehicles in the group. After receiving the proposal list broadcast by the coordinating node, each vehicle calculates the marginal contribution of each proposal to the local objective function based on its own status and task allocation. The marginal contribution is the change in the local objective function after adopting the proposal; a negative value indicates a cost reduction, and a positive value indicates a cost increase. For example, if vehicle V012 proposes to transfer an order to V015, the total mileage of the group is expected to decrease by 8.2 kilometers, and the marginal contribution is -8.2 kilometers, indicating that the proposal can reduce the group's cost.

[0065] Specifically, the calculation logic for the marginal contribution value is as follows: Let the local objective function of the current vehicle group be the weighted sum of the total mileage, total energy consumption, or total completion time of all vehicles in the group. The local objective function value before adopting the proposal is denoted as the first value. After simulating the adoption of the proposal, the task allocation and driving paths of the affected vehicles are updated, and the local objective function value of the group is recalculated to obtain the second value. The marginal contribution value is the difference between the second value and the first value. When the marginal contribution value is negative, it indicates that adopting the proposal can reduce the local objective function value, i.e., it has a positive improvement effect on the group. When the marginal contribution value is positive, it indicates that adopting the proposal will increase the local objective function value, i.e., it has a negative effect. The larger the absolute value of the marginal contribution value, the greater the impact of the proposal on the group. Taking the total mileage as the local objective function as an example, if the current total mileage of the group is 320 kilometers, and a vehicle proposes to transfer an order from vehicle A to vehicle B, after the simulation of adoption, the total mileage of the group becomes 311.8 kilometers. Then the marginal contribution value is -8.2 kilometers, indicating that the proposal can reduce the total mileage of the group by 8.2 kilometers.

[0066] Finally, the optimal scheduling adjustment proposal is selected based on the marginal contribution value, and consensus is reached within the vehicle group. After collecting the marginal contribution values ​​from all vehicles, the coordinating node selects the proposal with the best improvement effect as a candidate solution and broadcasts the candidate solution to all vehicles in the group for confirmation. If a majority of vehicles agree, consensus is reached, the proposal is adopted, and the group scheduling plan is updated. If they disagree, the next round of proposal and negotiation continues. For example, after three rounds of asynchronous iteration, the vehicles in the group reached a consensus on the transfer proposal for V012, while fine-tuning the routes of V015 and V018, ultimately reducing the total mileage of the Northeast group by 8.2 kilometers and the total energy consumption by 6.4 kWh. Through this distributed negotiation mechanism, the system partially delegates decision-making power to the vehicle group, enabling local disturbances to be resolved quickly through negotiation within the group, significantly improving the system's response speed and robustness.

[0067] To ensure the determinism and convergence of the distributed negotiation mechanism in engineering applications, this invention further introduces the following safeguard strategies: First, the system controls the maximum number of negotiation rounds. The system presets a maximum number of negotiation iterations (e.g., 10 rounds). When the number of negotiation rounds within a group reaches this limit without reaching a consensus, the coordinating node no longer waits for further negotiation. Instead, based on the marginal contribution values ​​of each proposal, it directly selects the proposal with the best improvement effect as the mandatory consensus solution and distributes it to all vehicles in the group for execution. This mechanism ensures that the negotiation process terminates within a finite number of steps, avoiding livelock or deadlock problems.

[0068] Second, cross-group conflict arbitration. When the negotiation results of multiple vehicle groups involve resource allocation conflicts for the same vehicle or order (e.g., two groups simultaneously requesting the use of the same spare vehicle), the group coordination node reports the conflict information to the upper-level central node. The central node re-evaluates the allocation scheme of the conflicting resources based on the global objective function value (rather than the local objective function), determines the final ownership, and issues an arbitration instruction to the relevant groups. Upon receiving the arbitration instruction, the relevant groups unconditionally adjust their local schemes. This mechanism ensures the coordination and consistency between global optimization and local negotiation.

[0069] Third, negotiation timeout protection. The system sets a timeout timer (e.g., 5 seconds) for each round of negotiation. If no response to a proposal is received from a vehicle within the timeout period, the coordinating node considers the vehicle offline, reassigns its task to other vehicles in the group, and excludes the vehicle from subsequent negotiations. This mechanism ensures the robustness of the negotiation process against sudden vehicle failures or communication interruptions.

[0070] Furthermore, the time domain length of the rolling time domain is a preset duration, and the rolling step size is less than the time domain length.

[0071] In this embodiment, the time domain length of the rolling time domain is a preset duration, and the rolling step size is smaller than the time domain length. The time domain length refers to the future time range covered by each optimization. For example, if it is set to 3 hours, the system will only optimize delivery tasks within the next 3 hours at each decision moment, rather than planning all tasks indefinitely. The rolling step size refers to the time interval by which the time domain window advances after each optimization. For example, if it is set to 15 minutes, the system will trigger a new optimization every 15 minutes. By setting the rolling step size to be smaller than the time domain length, there is an overlap between the time domain windows of two adjacent optimizations, ensuring that the scheduling scheme can transition smoothly. Taking the urban delivery scenario as an example, the system constructs an optimization problem covering 8:00 AM to 11:00 AM and generates a scheduling scheme at 8:00 AM. When the time advances to 8:15 AM, the system receives the latest vehicle location, remaining battery power, real-time traffic conditions, and new order data, rolls the time domain window forward to 8:15 AM to 11:15 AM, reconstructs and solves the optimization problem, and generates an updated scheduling scheme. This rolling optimization approach enables the system to continuously absorb real-time dynamic data and constantly correct subsequent scheduling decisions, avoiding the problem of traditional static scheduling schemes failing rapidly due to their inability to perceive dynamic changes. Simultaneously, by reasonably setting the time domain length and rolling step size—for example, a time domain length of 3 hours and a rolling step size of 15 minutes—it ensures both a controllable solution scale for the optimization problem and a rapid response to dynamic events.

[0072] This invention provides an embodiment of an intelligent logistics collaboration and dynamic scheduling optimization system based on full-link data of people, vehicles, goods, and yards, including: The data acquisition module is used to collect real-time data on the entire process of people, vehicles, goods, and yards. The status construction module is used to construct vehicle status data and network status data based on the full-link data of people, vehicles, goods and yards collected by the data acquisition module; The cargo volume prediction module is used to abstract the logistics network into a spatiotemporal graph structure. It uses a spatiotemporal graph convolutional network model to process the network point state data constructed by the state construction module and dynamically predict the cargo volume distribution in the future time domain. The rolling time domain optimization module is used to construct an optimization problem covering the future rolling time domain at each decision time, and to solve the optimization problem using a multi-operator search algorithm to generate an initial scheduling scheme. The constraints of the optimization problem include coupling constraints, which are used to characterize the relationship between vehicle state and driving path. The distributed negotiation module is used to divide the initial scheduling scheme generated by the rolling time-domain optimization module into multiple vehicle groups according to geographical regions, and to start a negotiation mechanism within each vehicle group to generate a revised scheduling scheme. The instruction issuance module is used to issue the revised scheduling scheme generated by the distributed negotiation module to the corresponding vehicles for execution.

[0073] In this embodiment, the system includes a data acquisition module, a status construction module, a cargo volume prediction module, a rolling time domain optimization module, a distributed negotiation module, and an instruction issuance module. The modules work together to achieve intelligent logistics collaboration and dynamic scheduling optimization.

[0074] The data acquisition module is used to collect real-time data across the entire data chain, including people, vehicles, goods, and storage. Specifically, this module acquires real-time data such as vehicle location, remaining energy, speed, fuel consumption, and electricity consumption through GPS and onboard terminals installed in vehicles; it acquires real-time data on warehouse and branch network inventory, pending pickup orders, and pending delivery orders through cameras, sensors, and warehouse management systems; and it acquires driver order status, working hours, scheduling preferences, and driving behavior data through driver mobile terminals. The data acquisition module continuously collects the above data at a frequency of seconds or minutes and transmits the raw data to the status construction module in real time.

[0075] The state construction module is used to construct vehicle status data and network status data based on the full-link data of people, vehicles, goods, and yards collected by the data acquisition module. Vehicle status data includes the vehicle's real-time location, remaining loading capacity, remaining energy, and remaining driver hours. For example, the real-time location of vehicle V001 is constructed as 116.397 degrees East longitude and 39.908 degrees North latitude, with a remaining loading capacity of 4.2 cubic meters, remaining energy of 98 kilowatt-hours, and 8 hours of remaining driver hours. Network status data includes network cargo volume status data; for example, the current pending outbound orders in the central warehouse are constructed as 800 orders, and pending delivery orders as 500 orders. The state construction module converts the collected raw data into structured state vectors, providing a unified data format for subsequent prediction and optimization.

[0076] The cargo volume prediction module abstracts the logistics network into a spatiotemporal graph structure. It processes the network point state data constructed by the state construction module using a spatiotemporal graph convolutional network model to dynamically predict the cargo volume distribution of network points in the future time domain. This module treats each network point as a node in the spatiotemporal graph and the logistics routes between network points as edges. It uses graph convolution operations to capture the mutual influence relationships of cargo volume changes between adjacent network points and temporal convolution operations to capture the periodic patterns of cargo volume changes over time. The cargo volume prediction module is trained using a prediction-optimization guided loss function, which simultaneously minimizes the difference between prediction error and scheduling cost, making the prediction results more conducive to subsequent optimization. The cargo volume prediction module outputs the predicted cargo volume values ​​for each network point in the future time domain, which serve as input to the rolling time domain optimization module.

[0077] The rolling time-domain optimization module constructs an optimization problem covering the future rolling time domain at each decision moment and solves it using a multi-operator search algorithm to generate an initial scheduling scheme. The objective function of the optimization problem includes travel distance cost, energy consumption cost, unused capacity penalty, and network backlog penalty. Constraints include vehicle capacity constraints, driver working hours constraints, task spatiotemporal consistency constraints, and coupling constraints. Coupling constraints characterize the relationship between vehicle state and travel path, specifically in the form of a nonlinear function, and are used to simultaneously consider the coupled effects of vehicle loading rate, remaining energy, and travel path. The rolling time-domain optimization module uses a multi-operator search algorithm to solve the problem. This algorithm includes multiple destruction operators and multiple repair operators, and dynamically adjusts the selection probability of each operator based on its historical performance, enabling adaptive convergence of the search process. After solving the problem, the rolling time-domain optimization module generates an initial global scheduling scheme.

[0078] The distributed negotiation module divides the initial scheduling scheme generated by the rolling temporal optimization module into multiple vehicle groups based on geographical regions. Within each vehicle group, a negotiation mechanism is initiated to generate a revised local scheduling scheme. This module uses a location-based clustering algorithm to divide vehicles into several groups based on their geographical location; for example, 20 vehicles from the Northeast region are grouped into one group. Within each group, a coordinating node is elected through a heartbeat mechanism. The coordinating node collects scheduling adjustment proposals from each vehicle within the group. Each vehicle calculates the marginal contribution of its proposal to the group's local objective function. Based on the marginal contribution, the optimal proposal is selected, and consensus is reached within the group. After multiple rounds of negotiation, the distributed negotiation module generates the revised local scheduling scheme.

[0079] The instruction issuance module distributes the revised scheduling plan generated by the distributed negotiation module to the corresponding vehicles for execution. This module sends scheduling instructions to the vehicle's onboard terminal or the driver's mobile terminal via a wireless communication network. The instructions include a list of tasks the vehicle needs to perform, the task order, the estimated arrival time at each delivery point, and suggested routes. After receiving the instructions, the vehicle executes the delivery task according to the scheduling plan. Changes in the vehicle's status during execution are captured again by the data acquisition module, forming a complete closed-loop feedback loop.

[0080] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent logistics coordination and dynamic scheduling optimization method based on human-vehicle-goods yard full-link data, characterized in that, Includes the following steps: Step S1: Collect real-time data on the entire chain of people, vehicles, goods, and yards to construct vehicle status data and network status data; Step S2: Abstract the logistics network into a spatiotemporal graph structure, and use the spatiotemporal graph convolutional network model to process the network point status data and dynamically predict the future cargo volume distribution in the time domain; Step S3: At each decision moment, construct an optimization problem covering the future rolling time domain. The constraints of the optimization problem include coupling constraints, which are used to characterize the relationship between the vehicle state and the driving path. Step S4: Use a search algorithm to solve the optimization problem and generate an initial scheduling scheme; Step S5: Divide the initial scheduling plan into multiple vehicle groups according to geographical regions, and initiate a negotiation mechanism within each vehicle group to generate a revised scheduling plan; Step S6: Send the revised scheduling plan to the corresponding vehicle for execution, and roll the time domain window forward by one step, repeating steps S1 to S5. 2.The intelligent logistics coordination and dynamic scheduling optimization method based on human-vehicle-goods field full-link data according to claim 1, characterized in that, Vehicle status data includes real-time vehicle location, remaining vehicle loading capacity, remaining vehicle energy, and remaining driver working hours. Network status data is the status data of the goods volume at the network. 3.The intelligent logistics coordination and dynamic scheduling optimization method based on human-vehicle-goods field full-link data according to claim 1, characterized in that, In the step of dynamically predicting the distribution of goods volume at network points in the future time domain using a spatiotemporal graph convolutional network model, the spatiotemporal graph convolutional network model is trained using a prediction-optimization guided loss function. The prediction-optimization guided loss function includes a prediction error term and a scheduling cost gap term. The prediction error term is the error between the predicted goods volume value and the historical actual goods volume value. The scheduling cost gap term is the cost difference between the scheduling scheme generated based on the predicted goods volume value and the optimal scheduling scheme generated by posterior calculation based on the historical actual goods volume value. 4.The intelligent logistics coordination and dynamic scheduling optimization method based on human-vehicle-goods field full-link data according to claim 1, characterized in that, In the step of dynamically predicting the distribution of goods volume at network points in the future time domain using a spatiotemporal graph convolutional network model, the spatiotemporal graph convolutional network model includes a dynamic attention module. The dynamic attention module generates attention weights for each network point based on time features, weather features, and holiday features. The attention weights are used to dynamically adjust the contribution of different network points in the goods volume prediction. 5.The intelligent logistics coordination and dynamic scheduling optimization method based on human-vehicle-goods field full-link data according to claim 1, characterized in that, The objective function of the optimization problem includes travel distance cost, energy consumption cost, unused capacity penalty, and network backlog penalty. The constraints of the optimization problem also include vehicle capacity constraints, vehicle driver working hours constraints, and task spatiotemporal consistency constraints. 6.The intelligent logistics coordination and dynamic scheduling optimization method based on human-vehicle-goods field full-link data according to claim 1, characterized in that, The coupling constraints are in the form of nonlinear functions, used to characterize the nonlinear relationship between vehicle load rate, vehicle remaining energy, and vehicle travel path. The coupling constraints are constructed based on the physical energy consumption laws during vehicle travel, including joint modeling of the nonlinear physical relationship between load rate and rolling resistance, the electrochemical characteristics of battery discharge efficiency decreasing with remaining charge, and the physical influence of road slope on vehicle traction. The functional expression of the coupling constraints is: wherein, is the base energy consumption, is the current load of the vehicle, is the maximum load capacity of the vehicle, is a non-linear coefficient greater than 1, is a residual energy penalty function, is a slope factor, is a driving behavior factor, is the available energy after deduction of the reserved energy; When the function value of the coupling constraint is greater than the preset threshold, it is determined that the current vehicle and driving path allocation is not feasible. 7.The intelligent logistics coordination and dynamic scheduling optimization method based on human-vehicle-goods field full-link data according to claim 1, characterized in that, The search algorithm is a multi-operator search algorithm, which includes multiple destruction operators and multiple repair operators. The steps for solving the optimization problem using a multi-operator search algorithm include: The current solution is iteratively optimized using multiple destruction operators and multiple repair operators. The destruction operators are used to remove some tasks from the current solution, and the repair operators are used to reinsert the removed tasks into the solution. The selection probability of each operator is dynamically adjusted based on its historical performance, enabling the search process to converge adaptively. The specific method for dynamically adjusting the selection probability of each operator is as follows: a weight value is maintained for each operator, and after each iteration, the operators used are scored based on the improvement of the objective function generated in this iteration. The cumulative score of each operator is periodically calculated, and the weight value of each operator is updated based on the cumulative score. Then, the weight value is converted into a selection probability through normalization. 8.The intelligent logistics coordination and dynamic scheduling optimization method based on human-vehicle-goods field full-link data according to claim 1, characterized in that, The negotiation mechanism is a distributed negotiation mechanism. The steps to initiate a distributed negotiation mechanism within each vehicle group include: Within each vehicle group, a coordinating node is elected through a heartbeat mechanism. The coordinating node collects scheduling adjustment proposals from each vehicle in the vehicle group. Each vehicle calculates the marginal contribution of its scheduling adjustment proposal to the local objective function of the vehicle group; The optimal scheduling adjustment proposal is selected based on the marginal contribution value, and a consensus is reached within the vehicle group; If consensus is not reached after the maximum number of iterations in the negotiation rounds, the coordinating node will directly select the proposal with the best improvement effect as the mandatory consensus solution based on the marginal contribution value and issue it for execution. When resource conflicts arise in the negotiation results of multiple vehicle groups, a cross-group arbitration mechanism is initiated. The upper-level central node determines the final allocation of conflicting resources based on the global objective function value and issues arbitration instructions to the relevant groups. 9.The intelligent logistics coordination and dynamic scheduling optimization method based on human-vehicle-goods field full-link data according to claim 1, characterized in that, The time domain length of the rolling time domain is a preset duration, and the rolling step size is less than the time domain length.

10. An intelligent logistics coordination and dynamic scheduling optimization system based on human-vehicle-goods yard full-link data, characterized in that, include: The data acquisition module is used to collect real-time data on the entire process of people, vehicles, goods, and yards. The status construction module is used to construct vehicle status data and network status data based on the full-link data of people, vehicles, goods and yards collected by the data acquisition module; The cargo volume prediction module is used to abstract the logistics network into a spatiotemporal graph structure. It uses a spatiotemporal graph convolutional network model to process the network point state data constructed by the state construction module and dynamically predict the cargo volume distribution in the future time domain. The rolling time domain optimization module is used to construct an optimization problem covering the future rolling time domain at each decision time, and to solve the optimization problem using a multi-operator search algorithm to generate an initial scheduling scheme. The constraints of the optimization problem include coupling constraints, which are used to characterize the relationship between vehicle state and driving path. The distributed negotiation module is used to divide the initial scheduling scheme generated by the rolling time-domain optimization module into multiple vehicle groups according to geographical regions, and to start a negotiation mechanism within each vehicle group to generate a revised scheduling scheme. The instruction issuance module is used to issue the revised scheduling scheme generated by the distributed negotiation module to the corresponding vehicles for execution.

Images