Method and system for dynamic optimization of postal logistics network based on digital twinning and adaptive learning

Abstract

This application relates to a method and system for dynamic optimization of postal logistics networks based on digital twins and adaptive learning, including: acquiring branch information, vehicle information, and delivery demand information; constructing a virtual simulation model; acquiring historical branch data and obtaining the expected parcel inflow volume of target branches through a demand prediction model; acquiring real-time traffic data, identifying road congestion events, and marking corresponding areas in the virtual simulation model; constructing and solving an objective function to generate an optimized execution plan; collecting feedback data, updating the demand prediction model, and correcting the virtual simulation model. In summary, this application, by constructing a digital twin virtual simulation model and combining it with an adaptive learning mechanism, achieves dynamic optimization of the postal logistics network, effectively improving the dynamic response capability of the postal logistics network to real-time traffic changes and fluctuations in parcel inflow volume. It also optimizes resource allocation efficiency and achieves synergistic optimization of delivery timeliness, transportation costs, and resource utilization.

Description

Technical Field

[0001] This application relates to the technical field of logistics information technology, and in particular to a method and system for dynamic optimization of postal logistics networks based on digital twins and adaptive learning. Background Technology

[0002] With the rapid development of e-commerce, postal logistics networks face multiple challenges, including a surge in parcel volume, increased demands for delivery timeliness, and rising operating costs. Traditional postal logistics network optimization methods mainly rely on static planning and manual experience-based scheduling, which are insufficient to cope with dynamically changing operating environments.

[0003] Existing logistics optimization solutions typically employ offline optimization algorithms, generating fixed delivery routes and resource allocation plans based on historical data before the start of each day's deliveries. The main problems with this approach are: First, offline optimization methods cannot respond in real time to sudden events such as changes in road conditions and fluctuations in parcel inflow, leading to significant discrepancies between actual performance and expectations. Second, these methods lack dynamic monitoring of network capacity and vehicle operating status, easily resulting in uneven resource allocation, network overload, or empty vehicles. Third, existing solutions usually have a single optimization objective, making it difficult to balance multiple objectives such as delivery timeliness, transportation costs, and resource utilization.

[0004] In existing technologies, some solutions introduce real-time scheduling mechanisms to dynamically adjust delivery routes through GPS positioning and road condition monitoring. However, these solutions typically employ simple rule-triggered or localized adjustment strategies, lacking a global understanding of the overall network operation status. Furthermore, adjustment decisions still rely on human intervention, resulting in slow response times and unstable decision quality. In addition, existing solutions generally lack deep learning and utilization of historical operational data, failing to extract patterns from past experience and continuously optimize decision-making models.

[0005] In recent years, although digital twin technology has been successfully applied to the real-time mapping and simulation prediction of physical systems in fields such as industrial manufacturing and smart cities, there are still significant obstacles to its application in the optimization of postal logistics networks. These obstacles include: how to build a virtual model that accurately reflects key parameters such as the sorting rate and capacity threshold of the network points; how to achieve real-time mapping of dynamic attributes such as vehicle load and driving speed; and how to combine demand forecasting results with real-time traffic information to support dynamic optimization of decision-making. Summary of the Invention

[0006] To address the aforementioned shortcomings, this application provides a method and system for dynamic optimization of postal logistics networks based on digital twins and adaptive learning.

[0007] The above-mentioned objective of this application is achieved through the following technical solution:

[0008] A dynamic optimization method for postal logistics networks based on digital twins and adaptive learning is applied to a target postal logistics network, which includes several target outlets. The method includes:

[0009] Obtain the network information of the target network point and the vehicle information of the transport vehicles associated with the target network point, and obtain the delivery demand information of the delivery packages associated with the transport vehicles.

[0010] A virtual simulation model is constructed based on network information, vehicle information, and delivery demand information. The virtual simulation model maps the operating status of each target network point and transport vehicle in real time.

[0011] Obtain historical data of the target outlets; based on the outlet information and historical data, use the trained demand prediction model to obtain the expected parcel inflow of the target outlets.

[0012] Acquire real-time traffic data for the target postal logistics network's associated area, identify road congestion events based on the real-time traffic data, and mark the corresponding areas in the virtual simulation model based on the road congestion events;

[0013] The optimization objective is determined, an objective function is constructed based on the optimization objective, and the objective function is solved by running an optimization algorithm through a virtual simulation model in combination with the expected package inflow, thereby generating an optimized execution plan;

[0014] Collect feedback data after the implementation of the optimized execution plan, update the demand prediction model based on the feedback data, and correct the running state estimation parameters in the virtual simulation model.

[0015] The second objective of this invention is achieved through the following technical solution:

[0016] A dynamic optimization system for postal logistics networks based on digital twins and adaptive learning includes:

[0017] The information acquisition module is used to acquire the network information of the target network point and the vehicle information of the transport vehicles associated with the target network point, and to acquire the delivery demand information of the delivery packages associated with the transport vehicles.

[0018] The virtual simulation module is used to construct a virtual simulation model based on network information, vehicle information, and delivery demand information. The virtual simulation model maps the operating status of each target network point and transport vehicle in real time.

[0019] The data prediction module is used to obtain historical data of the target outlet. Based on the outlet information and historical data, the expected parcel inflow of the target outlet is obtained through the trained demand prediction model.

[0020] The traffic marking module is used to acquire real-time traffic data of the target postal logistics network associated area, identify road congestion events based on the real-time traffic data, and mark the corresponding area in the virtual simulation model based on the road congestion events.

[0021] The scheme generation module is used to determine the optimization objective, construct an objective function based on the optimization objective, and solve the objective function by running an optimization algorithm through a virtual simulation model in combination with the expected package inflow, thereby generating an optimized execution scheme;

[0022] The feedback update module is used to collect feedback data after the execution of the optimized execution plan, update the demand prediction model based on the feedback data, and correct the running state estimation parameters in the virtual simulation model.

[0023] In summary, the postal logistics network dynamic optimization method and system based on digital twins and adaptive learning provided in this application can achieve dynamic optimization of the postal logistics network by constructing a digital twin virtual simulation model and combining it with an adaptive learning mechanism. This effectively improves the dynamic response capability of the postal logistics network to real-time road condition changes and fluctuations in parcel inflow, optimizes resource allocation efficiency, and achieves multi-objective collaborative optimization of delivery timeliness, transportation costs, and resource utilization. Attached Figure Description

[0024] Figure 1 This is a flowchart of an embodiment of a dynamic optimization method for postal logistics networks based on digital twins and adaptive learning, as proposed in this application.

[0025] Figure 2 This is a flowchart of step S20 in an embodiment of a dynamic optimization method for postal logistics networks based on digital twins and adaptive learning, as described in this application.

[0026] Figure 3 This is a flowchart of step S30 in an embodiment of a dynamic optimization method for postal logistics networks based on digital twins and adaptive learning, as described in this application. Detailed Implementation

[0027] The following is in conjunction with the appendix Figures 1-3 This application will be described in further detail.

[0028] In one embodiment, such as Figure 1 As shown, this application discloses a dynamic optimization method for postal logistics networks based on digital twins and adaptive learning, applied to a target postal logistics network, which includes several target outlets. The method specifically includes the following steps:

[0029] S10: Obtain the network information of the target network point and the vehicle information of the transport vehicles associated with the target network point, and obtain the delivery demand information of the delivery packages associated with the transport vehicles.

[0030] In this embodiment, digital twin technology refers to constructing a digital model in virtual space that completely corresponds to a physical entity, thereby enabling real-time monitoring, simulation, analysis, and optimization of the physical entity. In the postal logistics field of this embodiment, digital twin technology can map the operational status of various outlets, vehicles, and parcels in the logistics network in real time, providing data support for decision-making. Adaptive learning refers to automatically adjusting the parameters or structure of an internal model or system based on new data or environmental changes to improve performance and accuracy. In the postal logistics network optimization of this embodiment, the adaptive learning mechanism allows the demand forecasting model and virtual simulation model to continuously learn from actual operational data, thereby continuously improving the accuracy of prediction and simulation. The target postal logistics network refers to a system consisting of multiple outlets (such as sorting centers, business outlets, delivery stations, etc.), transportation routes connecting the outlets (such as highways, railways, aviation, etc.), and transportation vehicles, covering a specific area and responsible for logistics operations such as parcel collection, sorting, transportation, and distribution. Target outlets refer to physical nodes in the postal logistics network, including postal processing centers, sorting centers, and delivery stations. Each target outlet has specific attributes such as geographical location, processing capacity, and capacity limits, and is responsible for operations such as receiving, sorting, temporarily storing, and transferring parcels. Outlet information refers to a set of data describing the characteristics of a target outlet, specifically including geographical location information, outlet type information, processing capacity parameters, capacity parameters, and information related to surrounding outlets. Vehicle information refers to a set of data describing the characteristics of transport vehicles, specifically including vehicle identification, vehicle type, load parameters, speed parameters, and current status. Delivery demand information describes the demand characteristics of parcels to be delivered, including parcel attributes, destination attributes, time requirements, and priority.

[0031] Specifically, network point information can include basic data such as the network point's geographical location, processing capacity, and opening hours. Vehicle information can include vehicle type, load capacity, current location, and availability. Delivery demand information can include the package's origin, destination, dimensions, weight, and expected delivery time. This information can be obtained through manual input, sensor data collection, and database queries. For example, network point information can be exported from the logistics management system, vehicle information can be uploaded in real time by onboard GPS devices, and delivery demand information can be extracted from the order management system or package scanning data.

[0032] S20: Construct a virtual simulation model based on network information, vehicle information, and delivery demand information. The virtual simulation model maps the operating status of each target network and transport vehicle in real time.

[0033] In this embodiment, the virtual simulation model refers to a virtual mapping system of the postal logistics network built based on digital twin technology. By replicating the structure, parameters, and operational rules of the real logistics network in a computer, it can synchronize the state changes of the physical network in real time and support simulation prediction and optimization decisions. Real-time mapping refers to the dynamic synchronization mechanism between the virtual simulation model and the real logistics network. By continuously collecting operational data from the physical network, the state parameters of corresponding entities in the virtual simulation model are updated in real time, ensuring that the virtual simulation model always reflects the current state of the real network. Operational status refers to the current working status of each entity in the logistics network, specifically including the operational status of branch outlets, vehicle operations, and parcel flow.

[0034] Specifically, the virtual simulation model aims to map the operational status of each target network point and transport vehicle in real time. When constructing the virtual simulation model, network points can be abstracted as nodes with processing capabilities, vehicles as units with mobility capabilities, and packages as objects flowing within the network. These abstract entities interact through preset rules, thereby simulating the real logistics operation process. For example, the package processing rate of network points, the vehicle speed on road segments, and the transfer logic of packages between different network points can be set.

[0035] S30: Obtain historical data of the target outlet; based on the outlet information and historical data, use the trained demand prediction model to obtain the expected parcel inflow of the target outlet.

[0036] In this embodiment, historical network point data refers to historical data accumulated during the past operation of the target network point, specifically including historical parcel inflow sequences, historical timestamps, historical holiday identifiers, historical traffic data, etc. The demand forecasting model is a prediction algorithm model trained using machine learning techniques, used to predict the parcel inflow volume of each network point in future periods. The expected parcel inflow volume refers to the number of parcels that each target network point is expected to receive in future periods, output by the demand forecasting model. This is used as a basis for advance planning of resource allocation and route scheduling to avoid network point overload or resource idleness.

[0037] Specifically, historical network data can include statistical data such as parcel processing volume, inbound volume, and outbound volume for each network point over a past period. The demand forecasting model can be trained using various machine learning methods, such as statistical regression and time series analysis. By learning from historical data, the demand forecasting model can identify periodicity and trends in parcel inflow, thereby predicting future parcel inflow. For example, based on daily parcel volume data from the past few weeks or months, the parcel volume received by a sorting center in the next 24 hours can be predicted.

[0038] S40: Obtain real-time traffic data for the target postal logistics network associated area, identify road congestion events based on the real-time traffic data, and mark the corresponding area in the virtual simulation model based on the road congestion events;

[0039] In this embodiment, real-time traffic data refers to road traffic condition information obtained from traffic monitoring systems or third-party platforms, specifically including average vehicle speed, traffic density, and congestion status of each road segment. A road segment congestion event refers to road congestion identified through analysis of real-time traffic data, including information such as the location of the congested road segment, congestion level, and impact range. Marking the corresponding area refers to the operation of visually identifying and adjusting parameters of the identified congested road segments in the virtual simulation model, specifically including visual marking and parameter adjustments.

[0040] Specifically, real-time traffic data can come from traffic management departments, map service providers, or vehicle sensors. Congestion events can be identified by comparing real-time vehicle speed with historical average vehicle speed, and real-time traffic density with historical average traffic density. Once a congested section is identified, it can be marked in a virtual simulation model, for example, by changing the section's color or adding special markings to reflect its current traffic conditions.

[0041] S50: Determine the optimization objective, construct the objective function based on the optimization objective, and solve the objective function by running the optimization algorithm through the virtual simulation model in combination with the expected package inflow, thereby generating an optimized execution plan;

[0042] In this embodiment, the optimization objective refers to the goal pursued in optimizing the postal logistics network, such as minimizing total delivery time, minimizing total transportation cost, and maximizing network resource utilization. The objective function is the mathematical expression of the optimization objective, composed of multiple optimization terms, such as delivery time, transportation cost, and resource utilization. The objective function can typically integrate multiple optimization terms into a single evaluation index through weighted summation. The goal of the optimization algorithm is to find the optimal solution that optimizes the objective function. The optimization algorithm is the computational method used to find the optimal solution to the objective function. Common types include heuristic algorithms, exact algorithms, and hybrid algorithms. The optimization algorithm runs in a virtual simulation model, iteratively searching to generate a solution that satisfies the constraints and optimizes the objective function. Furthermore, the optimization process usually involves constraints that must be met, which may include vehicle load constraints, network capacity constraints, and time window constraints. The optimization execution plan refers to the specific implementation plan obtained by the optimization algorithm, specifically including a combination of path planning schemes and resource scheduling schemes.

[0043] Specifically, optimization objectives can include, but are not limited to, minimizing total transportation time, minimizing total operating costs, and maximizing vehicle utilization. The objective function is the mathematical expression of the optimization objective. The optimization algorithm can be a heuristic, metaheuristic, or exact algorithm. During the solution process, a virtual simulation model can provide a dynamic experimental environment, enabling the optimization algorithm to consider real-time changes in package inflow and road conditions. For example, the optimization objective can be set as minimizing the total travel distance while ensuring all packages are delivered on time.

[0044] S60: Collect feedback data after the execution of the optimized execution plan, update the demand prediction model based on the feedback data, and correct the running state estimation parameters in the virtual simulation model.

[0045] In this embodiment, feedback data refers to the actual operational data collected after the implementation of the optimized execution plan. This data is used to evaluate the plan's effectiveness and update the model. Specifically, it includes actual delivery time, actual transportation costs, actual parcel inflow, and the actual load of the target network points. Updating the demand forecasting model refers to the process of retraining the demand forecasting model based on the feedback data. When the prediction error exceeds a preset threshold, new feedback data is used to incrementally train the demand forecasting model, adjusting its parameters to improve prediction accuracy. Operational status estimation parameters refer to the parameters used in the virtual simulation model to estimate the operational status of network points and vehicles. These specifically include parcel sorting rate parameters, vehicle speed parameters, and network point capacity threshold parameters. Incremental learning is a machine learning model update method. Specifically, it involves using newly added training data to locally adjust the model while retaining the original model parameters. Compared to restarting training, incremental training is more computationally efficient and suitable for online learning and continuous optimization scenarios.

[0046] Specifically, feedback data can include actual delivery times, actual transportation costs, actual parcel inflows, and the actual load on delivery points. By comparing predicted and actual values, and simulation results with actual results, prediction errors and simulation deviations can be calculated. When these errors and deviations exceed preset thresholds, the demand forecasting model can be retrained or incrementally trained using new feedback data to improve its prediction accuracy. Simultaneously, parameters in the virtual simulation model used to estimate vehicle speeds, delivery point processing capacity, and other operational statuses can be corrected to better reflect reality. For example, if the actual traffic speed on a certain road segment consistently falls below the preset value in the virtual simulation model, the traffic speed parameters for that segment in the virtual simulation model can be adjusted.

[0047] For example, suppose that in a city’s postal logistics network there are multiple target outlets, such as sorting center A, delivery station B and delivery station C, and multiple transport vehicles D and E, which are facing challenges such as large fluctuations in parcel volume, frequent urban traffic congestion and high delivery time requirements.

[0048] First, the network information of sorting center A, delivery station B, and delivery station C is obtained, including their geographical location, package processing capacity, and operating hours. Simultaneously, vehicle information for transport vehicles D and E is obtained, such as their maximum load capacity, current location, and availability. Furthermore, delivery demand information for all packages awaiting delivery is acquired, including the origin, destination (e.g., delivery station B or C), volume, weight, and expected delivery time for each package. Further, based on the obtained network information, vehicle information, and delivery demand information, a virtual simulation model is constructed. In the virtual simulation model, sorting center A, delivery station B, and delivery station C are abstracted as nodes with specific processing capabilities, transport vehicles D and E are abstracted as units with movement and load-bearing capabilities, and packages awaiting delivery are simulated as objects flowing between nodes and units. At this point, the virtual simulation model can reflect in real-time the package accumulation at sorting center A, the package processing progress at delivery stations B and C, and the real-time location and load status of transport vehicles D and E.

[0049] Simultaneously, historical data from sorting center A, delivery station B, and delivery station C are acquired, such as daily parcel inflow and processing volume over the past few weeks or months. Based on this historical data and information about the stations' geographical location and type, a trained demand prediction model is used to calculate the expected parcel inflow to sorting center A in the coming hours, as well as the expected parcel processing volume to be handled by delivery stations B and C. For example, it is predicted that the expected parcel inflow to sorting center A will peak between 9:00 AM and 12:00 PM.

[0050] In addition, real-time traffic data for urban areas is continuously acquired, such as average vehicle speed and traffic density on major roads. Based on this real-time traffic data, congested road segments are identified, such as a road connecting sorting center A and delivery station B. Once a congestion event is identified, the congested road segment is marked in the virtual simulation model, for example, displayed in red, and the traffic speed parameters for that congested road segment in the model are adjusted to reflect the real traffic conditions.

[0051] Furthermore, the optimization objective is defined. For example, the optimization objective can be set as minimizing the total transportation cost and total delivery time while ensuring that all packages are delivered within the expected time. Based on this optimization objective, a corresponding objective function is constructed, and combined with the previously predicted expected package inflow, an optimization algorithm is run in a virtual simulation model. The optimization algorithm comprehensively considers the processing capacity of the network points, vehicle load limits, real-time traffic conditions on road sections, and the delivery time requirements of the packages to solve the objective function. Based on this, an optimized execution plan is generated, such as planning the optimal delivery route for transport vehicles D and E, and assigning package processing tasks to sorting center A, delivery station B, and delivery station C. For example, the optimization plan might instruct transport vehicle D to avoid congested road sections and choose a slightly longer but smoother alternative route to delivery station B.

[0052] Continuously, after the optimized implementation plan is actually executed, feedback data is collected. This feedback data may include the actual travel time and fuel consumption (reflecting transportation costs) of transport vehicles D and E, the actual parcel inflow at sorting center A, and the actual load conditions at delivery stations B and C. At this point, the feedback data is compared with previous predictions and simulation results. If a significant error is found in the demand forecasting model's prediction of the parcel inflow at sorting center A, or if there is a significant deviation between the estimated travel time and the actual travel time for a certain road segment in the virtual simulation model, the demand forecasting model can be incrementally trained using the new feedback data to better adapt to the new parcel inflow pattern. Simultaneously, the speed parameters of relevant road segments or the processing capacity parameters of network points in the virtual simulation model are corrected to improve the model's accuracy. Through this adaptive learning mechanism, the dynamic optimization method for the postal logistics network can continuously improve and optimize itself.

[0053] Compared to existing technologies that rely on static planning and manual experience-based scheduling, this embodiment achieves real-time mapping of the target network points and the operational status of transport vehicles by constructing a virtual simulation model. In the example above, the virtual simulation model can reflect the parcel accumulation at sorting center A and the real-time locations of transport vehicles D and E, thereby achieving dynamic perception of the overall status of the logistics network, rather than relying solely on a preset static plan.

[0054] To address the issue that existing offline optimization algorithms cannot respond in real-time to changes in road conditions and fluctuations in parcel inflow, this embodiment introduces a demand prediction model and a mechanism for recognizing real-time road condition data. In the example above, the demand prediction model can predict the expected parcel inflow to sorting center A. Combined with real-time road condition data, it can promptly identify road congestion events and update the virtual simulation model. Therefore, when generating execution plans, the optimization algorithm can fully consider future parcel demand and current traffic conditions. For example, it can guide transport vehicle D to avoid congested sections and choose a slightly longer but smoother alternative route to delivery station B, thereby avoiding deviations between actual execution results and expectations due to unforeseen circumstances.

[0055] Furthermore, existing solutions generally suffer from the limitation of having a single optimization objective. This embodiment identifies multiple optimization objectives, including minimizing total transportation costs and total delivery time, and constructs an objective function based on these objectives. This allows the generated optimization implementation plan to be weighed and optimized in a more comprehensive way, such as ensuring timeliness while taking cost-effectiveness into account, thereby improving the comprehensiveness of decision-making.

[0056] More importantly, this embodiment introduces an adaptive learning mechanism. Specifically, it collects feedback data after the optimized execution plan is implemented, such as actual delivery time and actual parcel inflow, and calculates prediction error and simulation deviation based on the feedback data. When the prediction error and simulation deviation exceed preset thresholds, the operating state estimation parameters in the demand forecasting model and the virtual simulation model are updated and corrected. In the example above, when it is found that the demand forecasting model is inaccurate in predicting parcel inflow, or that the road speed in the virtual simulation model does not match the actual speed, incremental training and parameter correction can be performed using new feedback data. Through this continuous self-learning and iterative optimization process, it can maintain continuous adaptation to real-time changes in the logistics network environment and improve the accuracy of optimization decisions.

[0057] In one embodiment, such as Figure 2 As shown, step S20 includes:

[0058] S21: Construct a network processing node based on network information, wherein the network processing node includes a parcel sorting rate parameter and a network capacity threshold parameter;

[0059] In this embodiment, a branch processing node refers to an abstract representation of a postal branch in the physical world within the virtual simulation model. The purpose of constructing branch processing nodes is to digitize the key attributes and processing capabilities of physical branches. Specifically, branch processing nodes include parcel sorting rate parameters and branch capacity threshold parameters. The parcel sorting rate parameter describes the efficiency of the branch in processing parcels, such as the number of parcels that can be sorted per hour. The branch capacity threshold parameter limits the total number of parcels that the branch can accommodate or process at a given time, preventing infinite backlog in the virtual simulation model and thus more realistically reflecting the physical limitations of the branch. Furthermore, the above parameters can be initialized based on historical operational data, equipment performance indicators, or human experience, and adjusted according to actual feedback data during the operation of the virtual simulation model.

[0060] S22: Construct a vehicle transportation unit based on vehicle information, wherein the vehicle transportation unit includes vehicle load parameters and vehicle speed parameters;

[0061] In this embodiment, a vehicle transportation unit refers to an abstract representation of a transportation vehicle in the physical world within a virtual simulation model. The purpose of constructing a vehicle transportation unit is to incorporate the key physical attributes and operational characteristics of the vehicle into the virtual simulation model. Specifically, the vehicle transportation unit includes vehicle load parameters and vehicle speed parameters. The vehicle load parameter represents the maximum volume or weight of packages the vehicle can carry, directly reflecting its transportation capacity and influencing the number of packages transported in a single trip. The vehicle speed parameter describes the average speed of the vehicle under different road conditions or traffic conditions. Furthermore, the settings of these parameters can be initialized and dynamically updated based on vehicle model, engine performance, historical GPS data, or real-time traffic information.

[0062] S23: Construct a parcel flow object based on delivery demand information, wherein the parcel flow object includes parcel volume attributes and destination attributes;

[0063] In this embodiment, the package flow object refers to the abstract representation of packages awaiting delivery in the physical world within the virtual simulation model. The purpose of constructing the package flow object is to incorporate the personalized characteristics and delivery requirements of each package into the virtual simulation model. The package flow object includes package volume attributes and destination attributes. The package volume attribute describes the physical dimensions of the package and is crucial for calculating parameters such as vehicle load capacity and network point capacity occupancy. The destination attribute clarifies the final delivery location of the package, serving as the basis for route planning and sorting decisions. Furthermore, these attributes can be directly obtained from order information and logistics waybill data, and serve as the basic unit information for package flow in the virtual simulation model.

[0064] S24: Establish the relationships between network processing nodes, vehicle transportation units, and parcel flow objects to form the topology of the virtual simulation model.

[0065] In this embodiment, establishing relationships refers to clarifying the interactions and dependencies between three core entities—network processing nodes, vehicle transportation units, and parcel flow objects—within the virtual simulation model. These relationships define how parcels flow between network points, how they are loaded and transported by vehicles, and how vehicles interact with network points. For example, a parcel flow object needs to be loaded by a vehicle transportation unit from one network processing node and then transported to another. These established relationships collectively constitute the topology of the virtual simulation model—an abstract representation of the entities and their interconnections—which determines the flow paths and rules of information and parcels within the virtual simulation model.

[0066] For example, as a specific implementation, the system first reads the network information of each target network point from the database, such as the network point's geographical coordinates, type, and historical processing volume. Based on the read network point information, a network point processing node object is created for each network point, and a package sorting rate parameter is set for it. For example, a large sorting center may be set to process 10,000 packages per hour, while a small delivery station may be set to process 500 packages per hour. Simultaneously, based on the network point's physical space and equipment configuration, a network point capacity threshold parameter is set. For example, the maximum package capacity for a sorting center is 50,000, and for a delivery station, it is 2,000. Further, vehicle information of the transport vehicles is obtained from the vehicle management system, including the vehicle model, maximum load capacity, and average fuel consumption. Based on this, a vehicle transport unit object is created for each vehicle, and a vehicle load parameter is set for it. For example, a heavy-duty truck is set to have a load capacity of 20 tons or 100 cubic meters, and a light delivery vehicle is set to have a load capacity of 2 tons or 10 cubic meters. Simultaneously, based on vehicle type and historical driving data, vehicle speed parameters are set; for example, highway vehicles are set to an average of 80 km / h, and urban delivery vehicles to an average of 30 km / h. Then, delivery demand information for packages to be delivered is obtained from the order system, including each package's unique identifier, volume, weight, origin, and destination. A package flow object is created for each package, assigning it volume and destination attributes; for example, a package might have a volume attribute of 0.1 cubic meters and a destination attribute of "address X". Finally, in the virtual simulation model, the relationships between these objects are established through programming interfaces or data structures. For example, a "load" operation can be defined to transfer package flow objects from network point processing nodes to vehicle transport units; an "unload" operation can be defined to transfer package flow objects from vehicle transport units to network point processing nodes. Furthermore, by defining virtual paths between network point processing nodes and specifying that vehicle transport units can move along these virtual paths, a complete topology is formed. For example, a directed graph can be constructed where nodes represent network processing nodes, edges represent road segments that vehicle transport units can travel on, and package flow objects flow along these edges and nodes.

[0067] By employing the aforementioned technical solutions, the elements of the postal logistics network, such as service points, vehicles, and parcels, are structurally modeled, and the relationships between these elements are established, thereby constructing a virtual simulation model with inherent logic. This virtual simulation model accurately reflects the actual operating status and dynamic changes of the physical logistics network, thus reducing simulation errors caused by insufficient model abstraction or disorganized data. Specifically, by setting parcel sorting rate parameters and service point capacity threshold parameters, the virtual simulation model can realistically simulate the processing capacity and bottlenecks of service points; by setting vehicle load parameters and vehicle speed parameters, the virtual simulation model can accurately reflect the transportation efficiency and limitations of vehicles; and by setting parcel volume attributes and destination attributes, the virtual simulation model can finely manage the flow of each parcel. Through structured modeling and the establishment of relationships, the virtual simulation model provides a high-fidelity simulation environment, offering a reliable input and verification platform for demand forecasting, road condition identification, and optimization algorithm solutions, thereby improving the accuracy and effectiveness of the entire dynamic optimization process.

[0068] In one embodiment, the demand prediction model includes a feature extraction layer, a temporal modeling layer, and a prediction output layer, such as... Figure 3 As shown, step S30 includes:

[0069] S31: Obtain historical branch data and branch information of the target branch. The historical branch data includes historical parcel inflow sequence, historical timestamps, and historical holiday identifiers. The branch information includes geographical location information, branch type information, and surrounding branch association information.

[0070] In this embodiment, the demand forecasting model is used to predict future parcel inflows by analyzing historical data. The demand forecasting model has a hierarchical structure including a feature extraction layer, a temporal modeling layer, and a prediction output layer. The feature extraction layer extracts useful features from the raw data, the temporal modeling layer captures the temporal dependencies of the data, and the prediction output layer generates the final prediction result based on the extracted features and dependencies. This hierarchical architecture helps decompose the complex prediction task into manageable and optimizable sub-tasks, thereby improving the prediction accuracy and generalization ability of the demand forecasting model. For example, the feature extraction layer can be implemented using a convolutional neural network or a multilayer perceptron, the temporal modeling layer can be implemented using a recurrent neural network or a Transformer model, and the prediction output layer can be implemented using a fully connected layer or a linear regression model.

[0071] To accurately predict parcel inflow at a target delivery point, it's necessary to collect historical operational data and static attribute information related to that point. Historical delivery point data provides trends and patterns in past parcel inflows, while delivery point information reveals inherent characteristics that may influence parcel inflows. For example, historical delivery point data can be periodically exported from the postal logistics network database, while delivery point information can be manually entered or obtained from a geographic information system. Historical delivery point data includes historical parcel inflow sequences, historical timestamps, and historical holiday identifiers. Historical parcel inflow sequences record the number of parcels entering the delivery point over a past period and are crucial for predicting future inflows. Historical timestamps mark the time each parcel inflow data point occurs, aiding in time series pattern analysis. Historical holiday identifiers identify special dates, such as public holidays and promotional days, as these often cause significant fluctuations in parcel inflows. For instance, historical parcel inflow sequences can be statistically analyzed by hour, day, or week; historical timestamps can be accurate to the minute; and historical holiday identifiers can be Boolean values ​​or enumeration types. The network information includes geographical location information, network type information, and surrounding network linking information. Geographical location information reflects the regional characteristics of the network, such as whether it is located in a city center, suburb, or rural area, and may be related to the parcel inflow volume. Network type information determines the network's functional positioning and processing capacity, thus affecting its parcel inflow pattern. Surrounding network linking information describes the topological relationships or business connections between the target network and other network points, helping the demand forecasting model understand the flow patterns of parcels in the network, thereby improving the accuracy of the inflow prediction for the target network. For example, geographical location information can be represented using latitude and longitude coordinates, network type information can use predefined classification labels, and surrounding network linking information can be represented using an adjacency matrix or graph structure.

[0072] S32: Arrange the historical parcel inflow sequence according to historical timestamps to construct time-series input data;

[0073] In this embodiment, by matching the historical parcel inflow sequence with its corresponding historical timestamps and arranging them chronologically, a continuous time series can be formed, which serves as the time-series input data. This arrangement ensures that the demand forecasting model can capture the inherent time dependencies between data during learning, such as seasonal variations, periodic fluctuations, or long-term trends. For example, the historical parcel inflow sequence can be organized into a multi-dimensional array, where each row represents a time step and contains the parcel inflow volume for that time step.

[0074] S33: The feature extraction layer encodes the geographic location information, branch type information, and historical holiday identifiers to generate a static feature vector;

[0075] In this embodiment, the feature extraction layer is used to convert the original branch information and historical holiday identifiers into numerical static feature vectors that the model can process. Specifically, geographic location information can be converted into a vector representation using embedding techniques or one-hot encoding; branch type information can also be encoded using one-hot encoding or learned embedding; and historical holiday identifiers can also be encoded using one-hot encoding or binary encoding. The static feature vectors generated after feature encoding can capture the impact of inherent branch attributes and external events on package inflow, providing important contextual information for the time-series modeling layer. For example, geographic location information can be encoded as an embedding vector of geographic regions, branch type information can be encoded as a one-hot vector, and historical holiday identifiers can be encoded as binary vectors.

[0076] S34: Extract temporal dependencies from temporal input data through a temporal modeling layer, and generate temporal feature representations by combining static feature vectors;

[0077] In this embodiment, the time-series modeling layer is used to identify and learn the dynamic changes in parcel inflow from time-series input data, which may include capturing short-term fluctuations, long-term trends, periodic patterns, and the impact of sudden events. By combining static feature vectors, the time-series modeling layer can incorporate the inherent attributes of network points and external event information into the time series analysis, thereby generating a more comprehensive time-series feature representation. For example, the time-series modeling layer can use long short-term memory networks or gated recurrent units to extract time-series dependencies, or it can use attention mechanisms to capture the importance of different time steps.

[0078] S35: Calculate the expected parcel inflow to the target network point by predicting the output layer based on time-series feature representation and surrounding network point association information.

[0079] In this embodiment, the prediction output layer receives the temporal feature representation generated by the temporal modeling layer and, combined with surrounding network information, calculates the expected parcel inflow to the target network. The temporal feature representation includes the target network's own historical dynamic information, while the surrounding network information provides information on network-level interactions. By integrating this information, the prediction output layer can consider the flow and interaction of parcels throughout the logistics network, thus enabling more accurate predictions. For example, the prediction output layer can be one or more fully connected layers, whose output is the expected parcel inflow to the target network at a future time period, or a probability distribution representing the possible range of parcel inflow.

[0080] For example, as a specific implementation, suppose a postal logistics network needs to predict the parcel inflow volume of a sorting center located in a city center over the next 24 hours. First, it obtains historical parcel inflow data for the sorting center over the past year, including the number of parcels per day and hour, along with the corresponding dates and timestamps. Simultaneously, it collects historical holiday information for the sorting center, such as New Year's Day, Spring Festival, and National Day. Furthermore, it obtains the sorting center's network information, including its specific latitude and longitude coordinates, network type information, and association information with several adjacent delivery stations and transfer centers. Further, the historical parcel inflow sequence over the past year is arranged chronologically to form a continuous time-series input data. Simultaneously, the feature extraction layer geocodes the sorting center's latitude and longitude information, converting it into a numerical vector; it performs one-hot encoding on the sorting center's network type; and it performs binary encoding on historical holiday identifiers, for example, 1 for holidays and 0 for non-holidays. This encoded information is then further combined into a static feature vector. Subsequently, the time-series modeling layer, such as a Transformer-based model, receives time-series input data and static feature vectors. It analyzes daily, weekly, and seasonal fluctuations in historical parcel inflow sequences and, combined with the static feature vectors, understands the impact of holidays on parcel inflows, as well as the fundamental impact of the sorting center's geographical location and type on its business volume. In this way, the time-series modeling layer can generate a time-series feature representation that incorporates both dynamic and static information. Finally, the prediction output layer, such as a multilayer perceptron, receives the time-series feature representation and combines it with the correlation information between the sorting center and its surrounding delivery stations and transfer centers. For example, if the parcel volume at a nearby delivery station has recently surged, the prediction output layer will consider this correlation and adjust its prediction of future parcel inflows to the target sorting center accordingly. Ultimately, the prediction output layer calculates and outputs the expected parcel inflow for each hour over the next 24 hours.

[0081] Through the above technical solution, this application can improve the prediction accuracy of parcel inflow to target outlets in the postal logistics network. Specifically, by refining the demand prediction model into a multi-layered architecture including a feature extraction layer, a time-series modeling layer, and a prediction output layer, and combining various influencing factors such as historical parcel inflow sequences, historical timestamps, historical holiday markers, geographical location information, outlet type information, and surrounding outlet association information, the demand prediction model can deeply understand the complex patterns of parcel inflow. Specifically, the feature extraction layer can transform heterogeneous data into features that the model can process, the time-series modeling layer can capture the dynamic changes and time dependencies of parcel inflow, and the prediction output layer can make predictions by integrating multi-dimensional information. Through this multi-layered prediction mechanism, the expected parcel inflow can more accurately reflect the actual situation, thus providing reliable input for the virtual simulation model to run optimization algorithms. This effectively reduces problems such as resource misallocation and low transportation efficiency caused by inaccurate predictions, thereby improving the dynamic optimization effect and operational efficiency of the postal logistics network.

[0082] In one embodiment, the temporal modeling layer includes a first path unit, a second path unit, a feature fusion unit, and a fully connected layer arranged in parallel. Step S34 includes:

[0083] S341: Extract short-term fluctuation characteristics of historical parcel inflow sequence through the first path unit;

[0084] In this embodiment, the main function of the time series modeling layer is to identify and learn the dependencies and patterns within the time series from the input historical time series data, providing time series features for subsequent prediction. The time series modeling layer includes a first path unit, a second path unit, a feature fusion unit, and a fully connected layer, all configured in parallel. The first path unit is used to capture relatively frequent, short-duration local change trends in the historical parcel inflow sequence, such as intraday fluctuations and weekly periodic changes. Its implementation can include using a convolutional neural network to extract multi-scale features from the sequence, or using a recurrent neural network such as a long short-term memory network or a gated recurrent unit to capture the short-term dependencies of the sequence.

[0085] S342: Calculate the importance weight of each time step in the historical parcel inflow sequence through the second path unit, and extract key time node features based on the importance weight;

[0086] In this embodiment, the second path unit is used to identify specific time points or time periods with significant influence in the historical parcel inflow sequence, such as holiday peaks or sudden events, and to quantify the importance of these time points or time periods. This can be achieved through an attention mechanism, which assigns a weight to each time step in the sequence, thereby highlighting key time nodes that have a significant impact on the prediction results.

[0087] S343: The feature fusion unit fuses short-term fluctuation features and key time node features to generate time-dependent features;

[0088] In this embodiment, the feature fusion unit integrates the short-term fluctuation features extracted by the first path unit with the key time node features extracted by the second path unit to form more comprehensive and richer time-series dependent features. The fusion method may include nonlinear transformation after feature concatenation, or adaptive fusion using a gating mechanism.

[0089] S344: Concatenates static feature vectors with temporally dependent features and generates temporally dependent feature representations through a fully connected layer.

[0090] In this embodiment, the fully connected layer is used to perform further nonlinear transformation and dimensional mapping on the fused temporal dependency features and static feature vectors, and finally generate a unified temporal feature representation, which will be used as the input of the demand forecasting model prediction output layer.

[0091] For example, as a specific implementation, the temporal modeling layer can be implemented as follows: The first path unit can employ a one-dimensional convolutional neural network containing multiple convolutional kernels to slide across the historical package inflow sequence and extract local patterns at different scales. For instance, multiple convolutional layers can be set, each followed by an activation function and a pooling layer to capture short-term fluctuations within different time windows. The second path unit can employ a self-attention mechanism to generate attention weights by calculating the correlation between any two time steps in the sequence, thereby identifying and highlighting the key time nodes that have the greatest impact on the prediction results. For instance, a multi-head self-attention layer in a Transformer encoder can be used. The feature fusion unit can concatenate the short-term fluctuation features output by the first path unit and the key time node features output by the second path unit, and then perform a nonlinear transformation through one or more fully connected layers to learn the complex interaction between the short-term fluctuation features and the key time node features. Finally, the fused temporal dependency features are concatenated with the static feature vector from the feature extraction layer and passed through a fully connected layer, such as a two-layer perceptron containing a ReLU activation function, to generate the final temporal feature representation.

[0092] Through the aforementioned technical solutions, the time-series modeling layer can simultaneously capture and effectively fuse short-term fluctuations and key time node features in historical parcel inflow sequences, thereby generating a more accurate time-series feature representation. This approach enhances the demand forecasting model's ability to understand and predict changes in parcel inflow, enabling the postal logistics network to more accurately anticipate parcel processing needs, thereby optimizing resource allocation and transportation scheduling, and improving overall operational efficiency and responsiveness.

[0093] In one embodiment, step S40 includes:

[0094] S41: Obtain real-time traffic data and historical traffic data for each road segment in the target postal logistics network associated area. The real-time traffic data includes the average vehicle speed and traffic density of the road segment, and the historical traffic data includes the historical average vehicle speed and historical average traffic density.

[0095] In this embodiment, real-time traffic data refers to the traffic condition information of each road segment at the current moment, typically including the average vehicle speed and the number of vehicles passing through a certain point per unit time or the density of vehicles per unit length. Real-time traffic data can be obtained from various sources, such as traffic sensors, vehicle-mounted GPS devices, mobile communication network data, or API interfaces provided by traffic management departments. Historical traffic data refers to the statistical information of traffic conditions on each road segment over a past period, such as the average vehicle speed and traffic density during specific time periods (e.g., morning rush hour, off-peak hour, evening rush hour). Historical traffic data can serve as a benchmark to assess the degree of anomaly in current traffic conditions. It can be obtained through statistical analysis of long-term accumulated real-time data or through prediction using traffic simulation models. Average vehicle speed of a road segment refers to the average speed of all vehicles on a specific road segment over a certain time period, which can be calculated from traffic flow detectors or floating car data. Traffic density of a road segment refers to the number of vehicles per unit length on a specific road segment, which can be obtained from devices such as loop detectors and video detectors. Historical average vehicle speed and historical average traffic density refer to the average values ​​obtained from historical data within the same road segment and time period, and are used as a reference benchmark to judge whether the current road conditions are abnormal.

[0096] S42: Calculate the speed deviation between the average speed of each road segment and the historical average speed, and calculate the density deviation between the traffic density of each road segment and the historical average traffic density.

[0097] In this embodiment, speed deviation refers to the degree of difference between the real-time average vehicle speed and the historical average vehicle speed, which can be expressed as a percentage or an absolute value. Density deviation refers to the degree of difference between the real-time traffic density and the historical average traffic density, which can also be expressed as a percentage or an absolute value.

[0098] S43: When the speed deviation exceeds the preset first threshold and the density deviation exceeds the preset second threshold, the corresponding road segment is marked as a congested road segment, and the congestion level of the congested road segment is recorded.

[0099] In this embodiment, the preset first threshold and the preset second threshold are critical values ​​used to determine whether a road segment is congested. Both the first and second thresholds can be obtained based on practical experience, traffic engineering standards, or through training a machine learning model. A congested road segment refers to a road segment identified as having abnormal traffic flow, slow vehicle movement, or stagnation. The congestion level is a quantification of the severity of congestion, which can be divided into different levels such as mild, moderate, and severe, or expressed numerically.

[0100] S44: Based on the geographic coordinates of congested road segments, locate the corresponding road segment nodes in the virtual simulation model;

[0101] In this embodiment, a road segment node refers to an abstract entity in the virtual simulation model that represents an actual road segment and contains the attribute information of the corresponding road segment.

[0102] S45: Mark congested road segments in the virtual simulation model according to the congestion level, and adjust the traffic speed parameters and traffic cost parameters of the road segment nodes according to the congestion level.

[0103] In this embodiment, the traffic speed parameter refers to the key parameter in the virtual simulation model used to simulate the time required for a vehicle to pass through a corresponding road segment, and its specific value directly affects the path planning result. The traffic cost parameter refers to the parameter in the virtual simulation model used to measure the resources (such as time, fuel consumption, toll fees, etc.) required to pass through a corresponding road segment, and is often used in the objective function of the optimization algorithm.

[0104] For example, as a specific implementation, real-time traffic data for the target postal logistics network's associated area can be continuously obtained from the traffic management department's API interface, for instance, updated every 5 minutes. Simultaneously, a historical traffic database can be maintained, storing the average vehicle speed and traffic density for each road segment during different time periods (e.g., weekday morning rush hour, off-peak, evening rush hour, weekends, etc.) over the past year. Upon receiving real-time data, the current average vehicle speed for each road segment can be compared with the historical average speed for the corresponding time period to calculate the speed deviation. For example, if the historical average speed for a road segment is 60 km / h and the real-time average speed is 30 km / h, the speed deviation is 50%. Furthermore, the density deviation between the real-time traffic density and the historical average traffic density can be calculated. If the preset first threshold is a speed deviation exceeding 30%, and the preset second threshold is a density deviation exceeding 20%, then when both conditions are met, the road segment will be marked as a congested road segment. Its congestion level can be further subdivided based on the specific deviation value. For example, a speed deviation of 30%-50% indicates light congestion, 50%-70% indicates moderate congestion, and above 70% indicates severe congestion. Once a congested road segment and its congestion level are identified, the corresponding road segment node can be found in the virtual simulation model based on the road segment's GPS coordinates. The traffic speed and cost parameters of this road segment node can then be adjusted according to the congestion level. For example, for light congestion, the traffic speed parameter might decrease by 20%, and the cost parameter might increase by 10%; for severe congestion, the traffic speed parameter might decrease by 50%, and the cost parameter might increase by 50%. Simultaneously, the adjustments will be immediately reflected in the virtual simulation model, providing accurate real-time traffic information for route planning and resource scheduling.

[0105] Through the above technical solutions, this application can more accurately identify and quantify road congestion in the postal logistics network. By comparing real-time traffic data with historical traffic data, and combining a dual judgment mechanism of speed deviation and density deviation, serious misjudgments that may be caused by relying on a single indicator can be avoided, thereby improving the accuracy of congestion identification. More importantly, by dynamically adjusting the traffic speed parameters and traffic cost parameters of road segment nodes in the virtual simulation model according to the congestion level, the virtual simulation model can reflect the impact of real-world traffic conditions on logistics operations in real time and accurately. This provides more realistic and reliable input for the optimization algorithm, ensuring that the generated optimization execution plan can effectively avoid congested areas, reduce transportation time and lower operating costs, thereby improving the overall operational efficiency and responsiveness of the postal logistics network.

[0106] In one embodiment, step S45 includes:

[0107] S451: Assign corresponding visual signs to congested road sections according to the congestion level, wherein the visual signs include color signs and congestion level labels;

[0108] In this embodiment, visual identifiers refer to graphic or textual information used to visually represent the state of congested road segments in a virtual simulation model. By assigning different visual identifiers to different congestion levels, operators or systems can more quickly and accurately identify the degree of congestion in a road segment. For example, color identifiers can adopt the color system of traffic lights, such as yellow for light congestion, orange for moderate congestion, and red for heavy congestion. Congestion level labels can be text descriptions, such as "light congestion," "moderate congestion," and "heavy congestion," or represented by numbers 1, 2, 3, etc. In addition, color identifiers can also use a gradient color spectrum, transitioning continuously from green to red, to more subtly reflect the degree of congestion. Congestion level labels can be combined with icons for visual representation; for example, one exclamation mark indicates light congestion, two exclamation marks indicate moderate congestion, and three exclamation marks indicate heavy congestion.

[0109] S452: Visually mark congested road sections in the virtual simulation model so that congested road sections are displayed with corresponding visual signs;

[0110] In this embodiment, visual markers refer to displaying visual identifiers assigned to congested road segments on the graphical interface of the virtual simulation model. This allows users to intuitively observe the congestion status of the road segments, enhancing the interactivity and user experience of the virtual simulation model and enabling decision-makers to quickly grasp the network's operational status. Specifically, the virtual simulation model can use a rendering engine to directly modify the line colors of congested road segments to the corresponding color identifiers and overlay congestion level labels above or beside the road segments. Alternatively, the virtual simulation model can employ 3D modeling technology to replace the surface textures or materials of congested road segments with textures containing visual identifiers, or generate floating text labels and icons above the road segments.

[0111] S453: Obtain the baseline traffic speed and baseline traffic cost of the road segment nodes corresponding to the congested road segment, and calculate the traffic speed adjustment coefficient and traffic cost adjustment coefficient according to the congestion level;

[0112] In this embodiment, the baseline traffic speed and baseline traffic cost are traffic parameters for a road segment under ideal or normal traffic conditions. To accurately simulate the impact of congestion on the logistics network, the baseline traffic speed and baseline traffic cost need to be adjusted according to the congestion level, and the traffic speed adjustment coefficient and traffic cost adjustment coefficient are factors that quantify the degree of adjustment. The baseline traffic speed and baseline traffic cost can be obtained from geographic information system data, historical traffic flow data, or road design standards. The traffic speed adjustment coefficient can be a multiplier less than 1, for example, 0.8 for light congestion, 0.5 for moderate congestion, and 0.2 for heavy congestion. The traffic cost adjustment coefficient can be a multiplier greater than 1, for example, 1.2 for light congestion, 1.5 for moderate congestion, and 2.0 for heavy congestion. Furthermore, the baseline traffic speed and baseline traffic cost can also be dynamically estimated through expert experience or machine learning models. Furthermore, the adjustment coefficient can also be a function that takes the congestion level as input and outputs the corresponding adjustment ratio. For example, it can be implemented using a piecewise linear function or an exponential function.

[0113] S454: Adjust the traffic speed parameters of road segment nodes based on the traffic speed adjustment coefficient, and adjust the traffic cost parameters of road segment nodes based on the traffic cost adjustment coefficient.

[0114] In this embodiment, by applying the calculated adjustment coefficients to adjust the actual traffic parameters of road segment nodes in the virtual simulation model, the virtual simulation model can more realistically reflect the impact of congestion on vehicle speed and transportation costs, thereby improving the accuracy of the simulation results. Generally, the adjusted traffic speed parameter equals the base traffic speed multiplied by the traffic speed adjustment coefficient, and the adjusted traffic cost parameter equals the base traffic cost multiplied by the traffic cost adjustment coefficient. In this case, the adjusted parameters can be directly used in the path planning and resource scheduling algorithms within the virtual simulation model. Furthermore, the adjustment process can be implemented by updating the road segment attribute table or the edge weights in the graph structure within the virtual simulation model. For example, in a graph theory model, the weights of road segments (representing travel time or cost) are updated in real time according to the adjustment coefficients to reflect the current congestion situation.

[0115] For example, as a specific implementation, when a road segment is identified as congested, for instance, it is judged to be "moderately congested" based on speed and density deviation. In this case, a specific visual identifier can be assigned to the "moderately congested" road segment according to preset rules. For example, the road segment can be displayed in orange in the virtual simulation model, with the text label "moderately congested" superimposed. Simultaneously, the baseline traffic speed (e.g., 50 km / h) and baseline toll cost (e.g., 1 yuan / km) for the road segment are obtained from the road segment node database. Furthermore, based on the congestion level of "moderately congested," a traffic speed adjustment coefficient (e.g., 0.6) and a toll cost adjustment coefficient (e.g., 1.5) can be calculated. Subsequently, the virtual simulation model updates the traffic speed parameter of the road segment node to 30 km / h and the toll cost parameter to 1.5 yuan / km. Based on this, in the virtual simulation model, the road segment is not only visually displayed in orange with a "moderate congestion" label, but its traffic speed and cost parameters have also been adjusted to reflect the real impact of congestion.

[0116] Through the above technical solution, this application can provide intuitive visual markings for congested road sections, enabling operators to quickly identify and understand different levels of congestion, thereby improving the readability of the virtual simulation model and the efficiency of decision support. Simultaneously, by adjusting the traffic speed and cost parameters of road section nodes based on the congestion level, the virtual simulation model can more realistically and accurately simulate the impact of actual traffic conditions on the operation of the postal logistics network. This combination of visualization and parameter adjustment effectively solves the problems of unintuitive congestion information presentation and coarse parameter adjustments in traditional methods, thus providing reliable input for the optimization algorithm. This allows the generated optimization execution plan to be closer to reality, thereby improving the accuracy and effectiveness of dynamic optimization of the postal logistics network.

[0117] In one embodiment, step S50 includes:

[0118] S51: Determine the optimization objective, which includes at least one of minimizing total delivery time, minimizing total transportation cost, and maximizing network resource utilization.

[0119] In this embodiment, the optimization objective refers to the specific performance indicators or states that are desired to be achieved during the dynamic optimization of the postal logistics network. These optimization objectives are often interrelated and may conflict. For example, minimizing total delivery time aims to improve efficiency and customer satisfaction, which can be achieved by shortening transportation routes and reducing waiting times. Minimizing total transportation costs focuses on economic benefits and may involve optimizing vehicle loading rates, reducing empty mileage, and selecting lower-cost transportation methods. Maximizing the utilization rate of network resources aims to improve the efficiency of existing infrastructure and personnel, such as by rationally allocating parcel processing tasks and avoiding network congestion. The optimization objectives can be selected and combined according to actual business needs and strategy priorities to guide the optimization algorithm in solving the problem.

[0120] S52: Construct an objective function based on the optimization objective, the objective function including delivery time, transportation cost, and resource utilization rate;

[0121] In this embodiment, the objective function refers to the mathematically quantified expression of the optimization objective. It transforms one or more optimization objectives into a computable mathematical expression, and the optimization algorithm finds the optimal solution by minimizing or maximizing the value of the objective function. For example, the delivery time term can be represented as the total or average time taken for all packages from pickup to delivery; the transportation cost term can be represented as the sum of fuel consumption, labor costs, and vehicle depreciation for all transportation vehicles; and the resource utilization rate term can be represented as the average utilization rate of network equipment, personnel, or vehicles. These terms can be assigned different weights according to their importance in the optimization objectives, thereby balancing multiple objectives. The construction of the objective function is the core of the optimization problem, directly determining the solution direction of the optimization algorithm and the quality of the solution.

[0122] S53: Determine the parcel processing requirements of each target network point based on the expected parcel inflow, and set constraints in the virtual simulation model, including vehicle load constraints, network point capacity constraints, and time window constraints.

[0123] In this embodiment, constraints refer to the restrictive conditions that must be met during the optimization process, reflecting the physical, resource, and temporal limitations of the postal logistics network in actual operation. These constraints include vehicle load constraints, network point capacity constraints, and time window constraints. Vehicle load constraints refer to the maximum weight or volume of packages that each transport vehicle can carry; for example, a truck cannot carry packages exceeding its designed load capacity. Network point capacity constraints refer to the number of packages or storage space that each target network point can handle within a specific time period; for example, a sorting center can only handle a certain number of packages per hour. Time window constraints refer to the restriction that package delivery or pickup must be completed within a specific time period; for example, some courier services promise delivery before 10:00 AM. These constraints ensure that the solution generated by the optimization algorithm is feasible and effective in actual operation.

[0124] S54: Run optimization algorithms in the virtual simulation model to solve the objective function and obtain the route planning scheme for each transport vehicle and the resource scheduling scheme for each target network point;

[0125] In this embodiment, an optimization algorithm refers to a mathematical method or computational process used to solve an objective function and satisfy constraints. Its purpose is to find the solution that optimizes the objective function among all feasible solutions. Common optimization algorithms include heuristic algorithms (such as genetic algorithms and ant colony algorithms) or exact algorithms (such as linear programming and integer programming). A route planning scheme refers to determining the specific travel routes for each transport vehicle from its origin to its destination to achieve the optimization objective and satisfy the constraints. This includes determining the vehicle's travel order, the network points it passes through, the travel route, and the estimated arrival time. A resource scheduling scheme refers to the rational allocation and time arrangement of various resources in the postal logistics network (such as network personnel, sorting equipment, and vehicles) to support the execution of the route planning scheme and achieve the optimization objective. This includes determining the amount of manpower and equipment required for each network point in different time periods.

[0126] S55: Generate an optimized execution plan based on path planning and resource scheduling schemes.

[0127] In this embodiment, the optimized execution plan refers to the detailed operational instructions formed by integrating the route planning plan and the resource scheduling plan, which can directly guide the actual operation of the postal logistics network. It usually includes the specific driving routes, stops, estimated arrival / departure times, loading and unloading tasks of each transport vehicle, as well as information such as the allocation of parcel processing tasks, personnel and equipment configurations for each target network point.

[0128] For example, as a specific implementation, when determining the optimization objective, it can be set to simultaneously minimize the total delivery time and minimize the total transportation cost, and assign different weights to the two. For instance, the delivery time weight is 0.6, and the transportation cost weight is 0.4, to reflect a higher priority for timeliness. Based on this, the objective function can be constructed as a weighted sum model, where the delivery time term can be the average delivery time of all packages, and the transportation cost term can be the total fuel consumption of all vehicles and the sum of driver wages. When setting constraints, the vehicle load constraint can be specifically set as a maximum load of 5 tons or a maximum volume of 30 cubic meters for each transport vehicle; the network capacity constraint can be set as a maximum package handling capacity of 10,000 pieces per hour for each sorting center, and a maximum temporary storage area of ​​5,000 pieces; the time window constraint can be set as all same-day delivery packages must be delivered before 20:00 on the same day, and the pickup time window for some specific customers is from 14:00 to 16:00 daily. When running optimization algorithms in a virtual simulation model, a mixed-integer linear programming algorithm or a reinforcement learning-based optimization algorithm can be used. For example, when using a mixed-integer linear programming algorithm, the path planning and resource scheduling problem can be modeled as a series of linear equations and inequalities, and solved using commercial solvers (such as CPLEX or Gurobi). This algorithm calculates the optimal vehicle routes, stopping sequences, package loading and unloading plans, and staff shift arrangements based on the expected package inflow, while satisfying constraints such as vehicle load capacity, branch capacity, and time windows. Finally, the calculation results are integrated into a detailed optimization execution plan. For example, an electronic task sheet containing a detailed route map, stopping points, estimated arrival / departure times, and a package loading list is generated for each vehicle, and an operational guide containing staff scheduling, equipment usage plans, and expected package processing volume is generated for each branch.

[0129] Through the above technical solution, this application can quantify the complex problem of postal logistics network optimization. Specifically, clarifying the optimization objective and constructing the objective function makes the optimization process based on evidence, effectively balancing multiple requirements such as timeliness, economy, and resource utilization. Combining the expected parcel inflow and constraints, it can be ensured that the generated route planning and resource scheduling schemes are not only theoretically optimal but also highly feasible in actual operation. Running the optimization algorithm in a virtual simulation model can fully utilize the precise mapping capability of digital twin technology to the real world, rapidly iterating and verifying multiple optimization strategies in a virtual environment, thereby generating high-quality optimization execution schemes. This can significantly improve the responsiveness of the logistics network to dynamic changes and the scientific nature of decision-making, effectively reducing operating costs while improving delivery efficiency and customer satisfaction, and avoiding resource waste and inefficiency caused by blind or experience-based decision-making.

[0130] In one embodiment, after step S55, the method further includes:

[0131] S551: During the execution of the optimized execution plan, continuously monitor the changes in the running status of the virtual simulation model;

[0132] In this embodiment, continuous monitoring of the virtual simulation model's operational status refers to the uninterrupted acquisition and analysis of real-time operational data of each target network point and transport vehicle within the virtual simulation model, such as the vehicle's real-time location, speed, parcel loading status, and the length of the parcel processing queue at the network point. This monitoring can be based on periodic data collection or implemented through an event-driven mechanism to ensure that the virtual simulation model always reflects the latest status of the postal logistics network.

[0133] S552: When a triggering event is detected, calculate the degree of impact of the triggering event on the optimized execution plan. The triggering events include road congestion events, sudden changes in parcel inflow events, and vehicle malfunction events.

[0134] In this embodiment, a triggering event refers to a sudden situation that may significantly impact the current optimized execution plan. Triggering events include road congestion events, sudden changes in parcel inflow, and vehicle malfunction events. Road congestion events can be traffic jams identified through real-time traffic data analysis; sudden changes in parcel inflow can be abnormal fluctuations discovered by comparing real-time parcel scanning data with expected parcel inflow; and vehicle malfunction events can be abnormal vehicle conditions obtained through onboard sensors or manual reports. The degree of impact can be calculated by simulating the scenario after the triggering event occurs in a virtual simulation model to assess its potential impact on key indicators such as total delivery time, total transportation cost, and network resource utilization rate, or by using a preset impact assessment model to quantify its deviation from the current plan based on parameters such as event type, severity, and location.

[0135] S553: ​​Determine the optimization response strategy based on the degree of impact, specifically including: when the degree of impact is lower than the preset first impact threshold, adopt the local path adjustment strategy; when the degree of impact is higher than the preset first impact threshold but lower than the preset second impact threshold, adopt the regional replanning strategy; when the degree of impact is higher than the preset second impact threshold, adopt the global re-optimization strategy.

[0136] In this embodiment, the optimized response strategy is determined by comparing the calculated impact level with preset thresholds. The preset first and second impact thresholds are quantitative standards used to distinguish different impact levels; for example, they can be set as a percentage increase in total delivery time or an absolute increase in total transportation costs. The local route adjustment strategy involves fine-tuning a small number of affected vehicles or road segments, such as changing vehicle routes or adjusting the delivery order of a few packages, to address minor disturbances with minimal computational overhead. The regional replanning strategy involves replanning vehicle routes and resource allocation within the affected area, such as re-optimizing the routes of all vehicles within a delivery area to address moderate disturbances. The global re-optimization strategy involves comprehensively re-optimizing all vehicle and branch resources across the entire postal logistics network to address severe disturbances and ensure the efficiency of the entire network.

[0137] S554: Based on the optimized response strategy, rerun the optimization algorithm in the virtual simulation model to generate an updated optimized execution plan.

[0138] In this embodiment, based on the selected optimization response strategy, the updated network state in the virtual simulation model is used as input to restart the corresponding optimization algorithm, so as to generate a better path planning scheme and resource scheduling scheme that can adapt to the current new situation.

[0139] For example, as a specific implementation, suppose a postal logistics network is delivering packages according to a pre-generated optimized execution plan. During the execution of the optimized plan, a virtual simulation model is continuously monitored, reflecting in real time the GPS location, speed, package load, and package processing progress of all transport vehicles and each target point. Suddenly, a serious traffic accident is detected on a critical road segment through real-time traffic data, causing complete congestion on that segment, i.e., a road segment congestion event is identified. At this time, the congestion event is immediately simulated in the virtual simulation model, and it is calculated that the event will cause the five transport vehicles originally scheduled to pass through this segment to be delayed by more than two hours, thereby increasing the total delivery time by 15%, and potentially causing two of the vehicles to be unable to complete delivery within the scheduled time window. According to the preset first impact threshold (e.g., a 5% increase in total delivery time) and second impact threshold (e.g., a 10% increase in total delivery time), it is known that the 15% impact level is higher than the second impact threshold, therefore it is determined that a global re-optimization strategy is required. Subsequently, the updated network status (including congested road information, affected vehicle locations, and package status) in the virtual simulation model is used as input to rerun the global optimization algorithm. This global optimization algorithm considers all available vehicles, network resources, and package delivery needs, recalculates the route planning schemes for all transport vehicles and the resource scheduling schemes for each target network point, and generates an optimized execution scheme that can avoid congested road sections and rearrange the delivery order of affected packages.

[0140] Through the above technical solutions, the postal logistics network can transform from static planning to dynamic adaptive optimization, thereby enhancing its ability to respond to emergencies. Specifically, the solution in this embodiment ensures that, even in the face of unpredictable external disturbances, it can quickly adjust and generate the optimal execution plan, effectively reducing delivery delays and operational cost losses caused by emergencies, and improving the overall resilience and service level of the logistics network. Simultaneously, the hierarchical optimization response strategy allows the optimization process to rationally allocate computing resources according to the severity of the event, avoiding unnecessary global re-optimization and improving system operating efficiency.

[0141] In one embodiment, step S60 includes:

[0142] S61: Collect feedback data after the optimized execution plan is implemented. The feedback data includes actual delivery time, actual transportation cost, actual parcel inflow, and actual load of the network points.

[0143] In this embodiment, feedback data refers to various operational indicators and status information collected from the real physical world after the optimized execution plan has been implemented. This feedback data includes actual delivery time, actual transportation costs, actual parcel inflow, and the actual load of the delivery point. Actual delivery time refers to the actual time it takes for a parcel to be delivered, which can be obtained through vehicle GPS trajectory data, parcel scanning records, etc. Actual transportation costs refer to the actual fuel consumption, labor costs, etc., incurred by the vehicle during the delivery task, which can be statistically obtained through vehicle sensors, financial systems, etc. Actual parcel inflow refers to the number of parcels actually received by the target delivery point within a specific time period, which can be obtained through the delivery point's parcel scanning system, warehousing records, etc. The actual load of the delivery point refers to the actual occupancy and utilization of the delivery point's resources, such as sorting equipment, personnel, and storage space, at a certain moment or within a certain period, which can be monitored through the delivery point's internal management system, sensor data, etc.

[0144] S62: Calculate the prediction error between the expected parcel inflow and the actual parcel inflow. When the prediction error exceeds the preset error threshold, perform incremental training on the demand prediction model based on the feedback data and update the model parameters of the demand prediction model.

[0145] In this embodiment, prediction error refers to the difference between the expected parcel inflow volume given by the demand forecasting model and the actual observed parcel inflow volume. It can be obtained by calculating the absolute difference, relative difference, or mean squared error between the two. The preset error threshold is a pre-defined tolerance level used to determine whether the prediction error is large enough to require model adjustment; for example, it can be set as a percentage or a fixed value. When the prediction error exceeds this threshold, it indicates that the current demand forecasting model has shortcomings in some aspects and needs correction. Incremental training refers to further training and learning the model based on existing model parameters using newly collected feedback data, without having to retrain from scratch. This training method allows the demand forecasting model to gradually adapt to new data patterns and environmental changes, thereby updating the model parameters and making its prediction capabilities more accurate and real-time.

[0146] S63: Calculate the simulation deviation between the estimated delivery time and the actual delivery time in the virtual simulation model, and correct the operating state estimation parameters in the virtual simulation model based on the simulation deviation.

[0147] In this embodiment, simulation deviation refers to the difference between the delivery time estimated by the virtual simulation model when simulating the delivery process and the actual delivery time. It reflects the accuracy of the virtual simulation model in simulating the real physical world's operating state. For example, the virtual simulation model may estimate the average traffic speed of a certain road segment based on historical data, but in actual operation, unexpected events may cause changes in the traffic speed, resulting in deviation. Operating state estimation parameters are key parameters in the virtual simulation model used to describe the dynamic behavior of the physical world, such as the average vehicle speed, the parcel processing efficiency of the network points, and the travel time of road segments. Correcting the operating state estimation parameters based on simulation deviation involves adjusting the parameters within the virtual simulation model according to the observed differences, making it closer to the real situation, thereby improving the simulation accuracy and reliability of the virtual simulation model.

[0148] For example, as a specific implementation method, after optimizing the deployment of the execution plan, the vehicle-mounted GPS device can be used to record the driving trajectory and dwell time of the transport vehicle in real time. Combined with the parcel collection and delivery timestamps recorded by the parcel scanning system, the actual delivery time can be accurately calculated. Simultaneously, the actual transportation cost can be calculated by using the vehicle's fuel consumption sensor and driver working hours records. Furthermore, the parcel processing system within the network point can provide real-time parcel inbound and outbound volumes as feedback on the actual parcel inflow and the actual load of the network point. For example, if the predicted parcel inflow to a network point at 10:00 AM is 500 pieces, but the actual inflow recorded by the scanning system is 600 pieces, a prediction error of 100 pieces can be obtained. If the preset error threshold is 50 pieces, this error exceeds the threshold. At this point, the new actual inflow data can be used to incrementally train the demand forecasting model using an online learning algorithm. For example, the weight parameters related to features such as time, holidays, and geographical location in the demand forecasting model can be adjusted to better adapt to current market demand changes. Furthermore, if the virtual simulation model estimates the delivery time for a certain route as 30 minutes, but the actual vehicle travel record shows 40 minutes, there is a 10-minute simulation discrepancy. In this case, the average travel speed parameter for that road segment in the virtual simulation model can be adjusted based on this discrepancy, for example, by lowering it from the original 40 km / h to 30 km / h, or the toll cost parameter for that road segment can be adjusted to more accurately reflect the actual road conditions and traffic situation.

[0149] Through the above technical solution, this application can effectively solve the problem of diminishing optimization effects caused by inaccurate model predictions and the disconnect between simulation models and actual operating conditions during the dynamic optimization of postal logistics networks. By continuously collecting actual operational feedback data and incrementally training the demand forecasting model based on this data, the accuracy and real-time performance of parcel inflow prediction can be improved, enabling optimization decisions to be based on more reliable future demand information. Simultaneously, by calculating simulation deviations and correcting the operating state estimation parameters in the virtual simulation model, it can be ensured that the virtual simulation model always maintains a high degree of consistency with the operating conditions of the real physical world, thus providing a more realistic and reliable simulation environment for the optimization algorithm. Through this adaptive learning and correction mechanism, the entire postal logistics network dynamic optimization method can possess self-correction and continuous optimization capabilities, effectively coping with rapid changes and uncertainties in the external environment, continuously generating high-quality optimization execution plans, and improving the overall operational efficiency of the postal logistics network, reducing operating costs, and enhancing service quality.

[0150] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0151] In one embodiment, a dynamic optimization system for a postal logistics network based on digital twins and adaptive learning is provided. This system corresponds one-to-one with the dynamic optimization method for a postal logistics network based on digital twins and adaptive learning described in the previous embodiment. The dynamic optimization system for a postal logistics network based on digital twins and adaptive learning includes:

[0152] The information acquisition module is used to acquire the network information of the target network point and the vehicle information of the transport vehicles associated with the target network point, and to acquire the delivery demand information of the delivery packages associated with the transport vehicles.

[0153] The virtual simulation module is used to construct a virtual simulation model based on network information, vehicle information, and delivery demand information. The virtual simulation model maps the operating status of each target network point and transport vehicle in real time.

[0154] The data prediction module is used to obtain historical data of the target outlet. Based on the outlet information and historical data, the expected parcel inflow of the target outlet is obtained through the trained demand prediction model.

[0155] The traffic marking module is used to acquire real-time traffic data of the target postal logistics network associated area, identify road congestion events based on the real-time traffic data, and mark the corresponding area in the virtual simulation model based on the road congestion events.

[0156] The scheme generation module is used to determine the optimization objective, construct an objective function based on the optimization objective, and solve the objective function by running an optimization algorithm through a virtual simulation model in combination with the expected package inflow, thereby generating an optimized execution scheme;

[0157] The feedback update module is used to collect feedback data after the execution of the optimized execution plan, update the demand prediction model based on the feedback data, and correct the running state estimation parameters in the virtual simulation model.

[0158] Specific limitations regarding the dynamic optimization system for postal logistics networks based on digital twins and adaptive learning can be found in the limitations of the dynamic optimization method for postal logistics networks based on digital twins and adaptive learning described above, and will not be repeated here. Each module in the aforementioned dynamic optimization system for postal logistics networks based on digital twins and adaptive learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0159] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A dynamic optimization method for postal logistics networks based on digital twins and adaptive learning, characterized in that, Applied to a target postal logistics network, the target postal logistics network including several target outlets, wherein the method includes: Obtain the network information of the target network point and the vehicle information of the transport vehicles associated with the target network point, and obtain the delivery demand information of the delivery packages associated with the transport vehicles. A virtual simulation model is constructed based on network information, vehicle information, and delivery demand information. The virtual simulation model maps the operating status of each target network point and transport vehicle in real time. Obtain historical data of the target outlets; based on the outlet information and historical data, use the trained demand prediction model to obtain the expected parcel inflow of the target outlets. Acquire real-time traffic data for the target postal logistics network's associated area, identify road congestion events based on the real-time traffic data, and mark the corresponding areas in the virtual simulation model based on the road congestion events; The optimization objective is determined, an objective function is constructed based on the optimization objective, and the objective function is solved by running an optimization algorithm through a virtual simulation model in combination with the expected package inflow, thereby generating an optimized execution plan; Collect feedback data after the implementation of the optimized execution plan, update the demand prediction model based on the feedback data, and correct the running state estimation parameters in the virtual simulation model.

2. The method for dynamic optimization of postal logistics networks based on digital twins and adaptive learning according to claim 1, characterized in that: The step of constructing a virtual simulation model based on network point information, vehicle information, and delivery demand information, and having the virtual simulation model map the operational status of each target network point and transport vehicle in real time, includes: A network processing node is constructed based on the network information. The network processing node includes a parcel sorting rate parameter and a network capacity threshold parameter. A vehicle transportation unit is constructed based on vehicle information, and the vehicle transportation unit includes vehicle load parameters and vehicle speed parameters. A parcel flow object is constructed based on delivery demand information, and the parcel flow object includes parcel volume attributes and destination attributes; Establish the relationships between network processing nodes, vehicle transportation units, and parcel flow objects to form the topology of the virtual simulation model.

3. The method for dynamic optimization of postal logistics networks based on digital twins and adaptive learning according to claim 1, characterized in that: The demand prediction model includes a feature extraction layer, a time series modeling layer, and a prediction output layer. The step of obtaining historical branch data of the target branch, and based on the branch information and historical branch data, obtaining the expected parcel inflow of the target branch through the trained demand prediction model, includes: Obtain historical data and information of the target outlet. The historical outlet data includes historical parcel inflow sequence, historical timestamps, and historical holiday identifiers. The outlet information includes geographical location information, outlet type information, and surrounding outlet association information. The historical parcel inflow sequence is arranged chronologically according to historical timestamps to construct time-series input data; A static feature vector is generated by encoding geographic location information, branch type information, and historical holiday identifiers through a feature extraction layer. Temporal dependencies are extracted from the temporal input data through a temporal modeling layer, and temporal feature representations are generated by combining static feature vectors. The expected parcel inflow to the target network point is calculated by predicting the output layer based on time-series feature representation and surrounding network point association information.

4. The method for dynamic optimization of postal logistics networks based on digital twins and adaptive learning according to claim 3, characterized in that: The temporal modeling layer includes a first path unit, a second path unit, a feature fusion unit, and a fully connected layer arranged in parallel. The step of extracting temporal dependencies from the temporal input data through the temporal modeling layer and generating temporal feature representations by combining static feature vectors includes: Short-term fluctuation characteristics of historical parcel inflow sequence are extracted using the first path unit; The importance weight of each time step in the historical parcel inflow sequence is calculated using the second path unit, and key time node features are extracted based on the importance weight. The feature fusion unit fuses short-term fluctuation features and key time node features to generate time-dependent features; The static feature vector is concatenated with the temporally dependent features, and a temporal feature representation is generated through a fully connected layer.

5. The method for dynamic optimization of postal logistics networks based on digital twins and adaptive learning according to claim 1, characterized in that: The steps of acquiring real-time traffic data of the target postal logistics network associated area, identifying road congestion events based on the real-time traffic data, and marking corresponding areas in the virtual simulation model based on the road congestion events include: Acquire real-time and historical traffic data for each road segment within the target postal logistics network's associated area. The real-time traffic data includes the average vehicle speed and traffic density of the road segment, while the historical traffic data includes the historical average vehicle speed and historical average traffic density. Calculate the speed deviation between the average vehicle speed and the historical average vehicle speed for each road segment, and calculate the density deviation between the traffic density of each road segment and the historical average traffic density. When the speed deviation exceeds the preset first threshold and the density deviation exceeds the preset second threshold, the corresponding road segment is marked as a congested road segment, and the congestion level of the congested road segment is recorded. Based on the geographic coordinates of congested road sections, the corresponding road section nodes are located in the virtual simulation model; Congested road segments are marked in the virtual simulation model according to the congestion level, and the traffic speed parameters and traffic cost parameters of the road segment nodes are adjusted according to the congestion level.

6. The method for dynamic optimization of postal logistics networks based on digital twins and adaptive learning according to claim 5, characterized in that: The steps of marking congested road segments in the virtual simulation model according to congestion levels, and adjusting the traffic speed parameters and traffic cost parameters of road segment nodes according to congestion levels, include: Visual signs are assigned to congested road sections based on their congestion levels. These visual signs include color codes and congestion level labels. Congested road sections are visually marked in the virtual simulation model so that they are displayed with corresponding visual identifiers; Obtain the baseline traffic speed and baseline traffic cost of the road segment nodes corresponding to the congested road segment, and calculate the traffic speed adjustment coefficient and traffic cost adjustment coefficient according to the congestion level; The traffic speed parameters of road segment nodes are adjusted based on the traffic speed adjustment coefficient, and the traffic cost parameters of road segment nodes are adjusted based on the traffic cost adjustment coefficient.

7. The method for dynamic optimization of postal logistics networks based on digital twins and adaptive learning according to claim 1, characterized in that: The steps of determining the optimization objective, constructing an objective function based on the optimization objective, and solving the objective function by running an optimization algorithm through a virtual simulation model in conjunction with the expected package inflow to generate an optimized execution plan include: Determine the optimization objective, which includes at least one of minimizing total delivery time, minimizing total transportation cost, and maximizing network resource utilization. An objective function is constructed based on the optimization objective, which includes a delivery time term, a transportation cost term, and a resource utilization rate term. Based on the expected parcel inflow, the parcel processing requirements of each target network point are determined, and constraints are set in the virtual simulation model. These constraints include vehicle load constraints, network point capacity constraints, and time window constraints. The optimization algorithm is run in the virtual simulation model to solve the objective function and obtain the route planning scheme for each transport vehicle and the resource scheduling scheme for each target network point. An optimized execution plan is generated based on the path planning scheme and resource scheduling scheme.

8. The method for dynamic optimization of postal logistics networks based on digital twins and adaptive learning according to claim 7, characterized in that: Following the step of generating an optimized execution plan based on the path planning scheme and resource scheduling scheme, the method further includes: During the execution of the optimized implementation plan, the changes in the running status of the virtual simulation model are continuously monitored; When a triggering event is detected, the degree of impact of the triggering event on the optimized execution plan is calculated. The triggering events include road congestion events, sudden changes in parcel inflow events, and vehicle malfunction events. The optimization response strategy is determined based on the degree of impact, specifically including: when the degree of impact is lower than the preset first impact threshold, a local path adjustment strategy is adopted; when the degree of impact is higher than the preset first impact threshold but lower than the preset second impact threshold, a regional replanning strategy is adopted; and when the degree of impact is higher than the preset second impact threshold, a global re-optimization strategy is adopted. Based on the optimized response strategy, the optimization algorithm is rerun in the virtual simulation model to generate an updated optimized execution plan.

9. The method for dynamic optimization of postal logistics networks based on digital twins and adaptive learning according to claim 1, characterized in that: The steps of collecting feedback data after the execution of the optimized execution plan, updating the demand prediction model based on the feedback data, and correcting the operating state estimation parameters in the virtual simulation model include: Collect feedback data after the implementation of the optimized execution plan. The feedback data includes actual delivery time, actual transportation cost, actual parcel inflow, and actual load of the network points. Calculate the prediction error between the expected parcel inflow and the actual parcel inflow. When the prediction error exceeds the preset error threshold, incrementally train the demand prediction model based on the feedback data and update the model parameters of the demand prediction model. Calculate the simulation deviation between the estimated delivery time and the actual delivery time in the virtual simulation model, and correct the operating state estimation parameters in the virtual simulation model based on the simulation deviation.

10. A dynamic optimization system for postal logistics networks based on digital twins and adaptive learning, characterized in that, include: The information acquisition module is used to acquire the network information of the target network point and the vehicle information of the transport vehicles associated with the target network point, and to acquire the delivery demand information of the delivery packages associated with the transport vehicles. The virtual simulation module is used to construct a virtual simulation model based on network information, vehicle information, and delivery demand information. The virtual simulation model maps the operating status of each target network point and transport vehicle in real time. The data prediction module is used to obtain historical data of the target outlet. Based on the outlet information and historical data, the expected parcel inflow of the target outlet is obtained through the trained demand prediction model. The traffic marking module is used to acquire real-time traffic data of the target postal logistics network associated area, identify road congestion events based on the real-time traffic data, and mark the corresponding area in the virtual simulation model based on the road congestion events. The scheme generation module is used to determine the optimization objective, construct an objective function based on the optimization objective, and solve the objective function by running an optimization algorithm through a virtual simulation model in combination with the expected package inflow, thereby generating an optimized execution scheme; The feedback update module is used to collect feedback data after the execution of the optimized execution plan, update the demand prediction model based on the feedback data, and correct the running state estimation parameters in the virtual simulation model.

Images