An artificial intelligence-based logistics supply chain optimization system and method

Abstract

The application relates to the technical field of intelligent logistics, and discloses a logistics supply chain optimization system and method based on artificial intelligence, wherein the logistics supply chain optimization method based on artificial intelligence comprises the following steps: constructing a heterogeneous vehicle characteristic vector and establishing a digital twin model; calculating a matching relationship between a task and a vehicle based on a graph attention network; adopting a hierarchical reinforcement learning algorithm to decompose a scheduling problem into multiple hierarchical subproblems; constructing a decentralized collaborative decision-making system; and realizing a task exchange protocol based on game equilibrium; the application solves the problems of heterogeneous vehicle resource mismatch, centralized decision-making response lag, low collaborative efficiency of multi-level decision-making subjects, and insufficient large-scale task cooperation, and improves the efficiency and service quality of logistics distribution.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics technology, and more specifically, to an artificial intelligence-based logistics supply chain optimization system and method. Background Technology

[0002] With the rapid development of e-commerce and the fast growth of logistics demand, modern logistics and distribution systems are facing an increasingly complex operating environment. Logistics fleets are typically composed of heterogeneous vehicles with different load capacities, speed ranges, energy types, and route restrictions, and this diversity brings challenges to vehicle scheduling and resource allocation.

[0003] Currently, logistics and distribution management primarily employs centralized scheduling algorithms, such as integer programming, greedy algorithms, and heuristic algorithms. While these traditional methods perform well in handling static, small-scale problems, they exhibit significant limitations when facing complex scenarios involving dynamic, large-scale, and heterogeneous fleets. Traditional centralized scheduling algorithms struggle to simultaneously consider the diverse characteristics of various vehicles, leading to resource misallocation. For example, lightly loaded vehicles are frequently assigned heavy cargo, or refrigerated trucks are often used for routine delivery tasks. In situations where urban logistics environments change frequently and order demands are dynamically adjusted in real time, centralized decision-making lags behind, hindering flexible fleet collaboration. The supply chain involves multiple levels of decision-makers with inconsistent optimization goals, resulting in low overall collaborative efficiency and uneven resource utilization. Furthermore, under large-scale delivery tasks, the lack of effective collaboration mechanisms among vehicles makes dynamic task reallocation and load balancing difficult.

[0004] These problems severely restrict the efficiency improvement and service quality enhancement of logistics and distribution systems. A comprehensive technical solution is needed to address these issues, enabling intelligent and collaborative management of heterogeneous logistics fleets, improving resource utilization, reducing operating costs, and enhancing service response speed. Summary of the Invention

[0005] This invention provides an artificial intelligence-based logistics supply chain optimization system and method, which solves the technical problems existing in related technologies, such as mismatch of heterogeneous vehicle resources, lag in centralized decision-making response, low efficiency of multi-level decision-making entities, and insufficient large-scale task collaboration.

[0006] This invention provides an artificial intelligence-based logistics supply chain optimization method, comprising:

[0007] Construct a heterogeneous vehicle characteristic vector representation, which includes load capacity, speed range, energy type and path limitation parameters, and establish a digital twin model for each vehicle;

[0008] The vehicle characteristic vector and real-time status data of the digital twin model are input into the graph attention network to calculate the importance weight between the task and the vehicle, learn the optimal matching relationship between different tasks and vehicles, and generate a task-vehicle matching score matrix.

[0009] Based on the matching score matrix, a hierarchical reinforcement learning algorithm is used to decompose the global scheduling problem into multi-level sub-problems, establish a reward function, dynamically balance energy efficiency, time delay and operating costs, and generate a hierarchical scheduling strategy.

[0010] Based on the hierarchical scheduling strategy, a decentralized collaborative decision-making system is constructed, enabling each vehicle to make task allocation decisions based on local information and global objectives, and form a preliminary allocation plan.

[0011] For the initial allocation scheme, a task exchange protocol based on game equilibrium is constructed to resolve resource competition conflicts, support dynamic task redistribution and load balancing, and generate the final optimized task allocation scheme.

[0012] In a preferred embodiment, the digital twin model employs a multi-layer feedforward neural network structure, with the input layer receiving vehicle characteristic vectors and state data, and the output layer generating vehicle performance prediction values.

[0013] In a preferred embodiment, the graph attention network comprises multiple attention heads, each of which independently learns a different feature subspace, and the results are finally merged by weighted average.

[0014] In a preferred embodiment, the hierarchical reinforcement learning algorithm decomposes the scheduling problem into three levels: a regional resource allocation layer, a vehicle scheduling layer, and a path planning layer.

[0015] In a preferred embodiment, the hierarchical reinforcement learning algorithm constructs a state space, action space, and transition function for each level, and trains the decision model for each level using a deep Q-network.

[0016] In a preferred embodiment, the decentralized collaborative decision-making system includes edge computing nodes, a local communication network, and a consensus algorithm module.

[0017] In a preferred embodiment, the game equilibrium-based task exchange protocol comprises three components: utility function definition, exchange strategy generation, and revenue distribution rules.

[0018] In a preferred embodiment, the revenue distribution rule calculates the marginal contribution of each participant based on the Shapley value to achieve fair distribution.

[0019] In a preferred embodiment, the method further includes dynamically adjusting the weight coefficients of the reward function, increasing the weight of time delay during peak periods and increasing the weight of energy efficiency during off-peak periods.

[0020] In a preferred embodiment, an AI-based logistics supply chain optimization system is used to execute an AI-based logistics supply chain optimization method, including:

[0021] The heterogeneous vehicle digital twin model construction module is used to build vehicle characteristic vector representations and establish digital twin models. The characteristic vectors include load capacity, speed range, energy type and path limitation parameters.

[0022] The graph attention network module is used to calculate the importance weights between tasks and vehicles, learn the optimal matching relationship between different tasks and vehicles, and generate a task-vehicle matching score matrix.

[0023] The hierarchical reinforcement learning scheduling module is used to decompose the global scheduling problem into multi-level sub-problems, establish a reward function, dynamically balance energy efficiency, time delay and operating costs, and generate hierarchical scheduling strategies.

[0024] The decentralized collaborative decision-making module enables each vehicle to make task allocation decisions based on local information and global objectives, forming a preliminary allocation plan;

[0025] The game equilibrium task exchange module is used to resolve resource competition conflicts, support dynamic task redistribution and load balancing, and generate the final optimized task allocation scheme.

[0026] The beneficial effects of this invention are as follows:

[0027] This invention solves the problem of resource mismatch between heterogeneous vehicles. By constructing a digital twin model of vehicles and a matching algorithm based on graph attention networks, this invention can accurately characterize the differences in characteristics between different vehicles, achieve optimal matching between tasks and vehicles, avoid resource waste, and improve the overall efficiency of the fleet.

[0028] This invention overcomes the drawback of delayed response in centralized decision-making. It employs a decentralized collaborative decision-making system, enabling each vehicle to make rapid decisions based on local information, while maintaining global goal consistency through a consensus algorithm.

[0029] This invention improves the collaborative efficiency of multi-level decision-making entities. The hierarchical reinforcement learning algorithm decomposes the complex scheduling problem into multiple levels, each level optimizing for a specific decision problem, while maintaining information exchange between levels to ensure that local decisions are consistent with the global goal.

[0030] This invention enhances collaboration capabilities under large-scale tasks. By establishing a fair and efficient task redistribution mechanism through a game-theoretic equilibrium-based task exchange protocol, it resolves conflicts caused by resource competition and supports dynamic collaboration between vehicles.

[0031] This invention reduces logistics operating costs. The intelligent scheduling strategy optimizes energy consumption and time costs while ensuring service quality. Attached Figure Description

[0032] Figure 1 This is a flowchart of an artificial intelligence-based logistics supply chain optimization method according to the present invention;

[0033] Figure 2 This is a detailed flowchart of the present invention for creating a digital twin model for each vehicle;

[0034] Figure 3 This is a detailed flowchart of the process for generating the task vehicle matching score matrix according to the present invention;

[0035] Figure 4 This is a detailed flowchart of the generation of the hierarchical scheduling strategy of the present invention;

[0036] Figure 5 This is a detailed flowchart of the formation of the preliminary allocation scheme of the present invention;

[0037] Figure 6 This is a detailed flowchart of the process for generating the final optimized task allocation scheme according to the present invention. Detailed Implementation

[0038] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0039] At least one embodiment of the present invention discloses an artificial intelligence-based logistics supply chain optimization method, such as... Figures 1 to 6 As shown, it includes the following steps:

[0040] Step 1: Construct a heterogeneous vehicle characteristic vector representation. The characteristic vector includes load capacity, speed range, energy type and path limitation parameters, and establish a digital twin model for each vehicle.

[0041] Specifically, the following steps are included:

[0042] Step 1.1: Collect heterogeneous vehicle characteristic data;

[0043] Including load capacity (in kilograms), speed range (in kilometers per hour), energy type (e.g., classification values ​​for electric, gasoline, and hybrid vehicles) and route restrictions (For example, the set of road types that are allowed to pass).

[0044] Step 1.2: Standardize the various characteristic data;

[0045] Converting the properties of different dimensions into a unified numerical range This forms a standardized characteristic vector.

[0046] Step 1.3, Construct the vehicle characteristic vector representation:

[0047] ;

[0048] in Indicates the first Vehicle characteristic vectors This indicates the standardized load-bearing capacity. This indicates the standardized speed range. Indicates the type of energy after encoding. This indicates the path restrictions after encoding.

[0049] Step 1.4: Create a digital twin model for each vehicle;

[0050] Based on feature vectors, a digital twin model is built for each vehicle. This model includes the vehicle's current state (location, load, remaining energy, etc.) and historical performance indicators, which are used to reflect the vehicle's operating status and performance characteristics in real time.

[0051] The digital twin model employs a multi-layer feedforward neural network structure, with the input layer receiving vehicle characteristic vectors. and status data :

[0052] ;

[0053] in, This represents vehicle status data. Indicates the vehicle's current GPS location coordinates. Indicates the current load rate. Indicates the percentage of remaining energy;

[0054] The intermediate layer adopts a double hidden layer structure. The first hidden layer contains 64 neurons and uses the ReLU activation function for nonlinear transformation; the second hidden layer contains 32 neurons and also uses the ReLU activation function.

[0055] The output layer generates predicted vehicle performance values:

[0056] ;

[0057] in, This represents the predicted value of vehicle performance. This indicates the predicted remaining mileage. This indicates the projected energy consumption rate. This indicates the predicted service duration.

[0058] The model training uses the mean squared error loss function:

[0059] ;

[0060] in, Let the mean squared error loss function be . For the true value, For predicted values, The number of samples;

[0061] The parameters were updated using the Adam optimizer with a learning rate of 0.001 and a batch size of 64.

[0062] In practical applications, such as urban delivery scenarios, this model can predict the remaining mileage under different load and route conditions based on historical battery consumption data of electric logistics vehicles, thus avoiding delivery interruptions due to insufficient power.

[0063] Step 2: Input the vehicle characteristic vector and real-time state data of the digital twin model into the graph attention network, calculate the importance weight between the task and the vehicle, learn the optimal matching relationship between different tasks and vehicles, and generate the task-vehicle matching score matrix.

[0064] Specifically, the following steps are included:

[0065] Step 2.1: Construct the bipartite graph network of the mission vehicle;

[0066] In this context, task nodes represent delivery orders, vehicle nodes represent available logistics vehicles, and edges between nodes represent potential task allocation relationships.

[0067] Step 2.2: Extract feature vectors for task nodes;

[0068] Feature vector This includes information such as order location, time window, cargo volume and weight, and utilizes the vehicle characteristic vector generated in step 1. As the feature vector of vehicle nodes .

[0069] Step 2.3: Calculate the importance weights between the task and the vehicle;

[0070] A graph attention network is applied to calculate the importance weights between the task and the vehicles. The graph attention network is implemented using a multi-head attention structure, comprising eight attention heads, each learning a different feature subspace independently. The results are then combined through a weighted average. The specific implementation process is as follows:

[0071] Multi-head attention mechanism: For each attention head ( ), independently calculate attention weights:

[0072] ;

[0073] in, Indicates task With vehicles The importance weight between them For learnable attention vectors, The characteristic transformation matrix, and These are the feature vectors for the task and the vehicle, respectively. This represents a vector concatenation operation. Representation and Task The set of all connected vehicle nodes It is a modified linear unit activation function with a small slope. This represents an exponential function.

[0074] Feature transformation and aggregation: each attention head The generated output features are:

[0075] ;

[0076] in, Indicates the first A person's attention The generated output features For ELU activation function, Attention weights represent the node's attention weights. For nodes The degree of impact; The nodes after weight matrix transformation feature; Indicates the first Nodes generated by attention heads The new features are represented.

[0077] Multi-head result merging: The output features of the eight attention heads are concatenated and then merged through a linear transformation.

[0078] ;

[0079] in, It is the output transformation matrix. This refers to the final output feature dimension; This means concatenating the output features of the eight attention heads along their dimensions to form a... A dimensional vector; For nodes The final characteristic representation.

[0080] The input layer of this network receives the node feature matrix. ,in The number of nodes represents the total number of tasks and vehicles in the system. is the feature dimension, representing the length of the original feature vector for each node. The intermediate layer contains two graph convolutional layers; the first layer takes the input features... The first layer is converted to 64 dimensions, and the second layer retains the 64-dimensional features. Residual connections are used in each layer to alleviate the training problems of deep networks.

[0081] Specifically, the calculation formula for the convolutional layer in the first image is:

[0082] ;

[0083] in, This represents the node feature representation after the first layer of the graph convolutional network. This is the original node feature matrix; It is an adjacency matrix; The weight matrix for the jump connections; This represents a graph convolutional network. It is a non-linear activation function.

[0084] The output layer generates the attention weight matrix between nodes. The calculation formula is:

[0085] ;

[0086] in, Indicates task With vehicles Match scores between them The larger the value, the more important the task. With vehicles The higher the match rate, the more likely the system is to assign the task to that vehicle. It is a two-layer perceptron, with 64 neurons in the first layer activated using ReLU; and 1 neuron in the second layer, which outputs the raw matching score. Indicates task The final feature representation; Indicates vehicle The final feature representation; This indicates that the feature vectors of the task and the vehicle are concatenated; Indicates the total number of tasks. Indicates the total number of vehicles; This represents an exponential function.

[0087] In practical applications, such as fresh food e-commerce logistics, this network can learn the matching relationship between the timeliness of fresh produce and refrigerated trucks, as well as the matching patterns between large appliances and large trucks, thereby avoiding unreasonable resource allocation problems. For example, when the system receives a fresh produce order that needs to be delivered within 2 hours, the graph attention network will automatically increase the matching score between the order and the refrigerated truck, ensuring that time-sensitive goods are prioritized for allocation to specialized vehicles. At the same time, for large appliance orders with a volume exceeding 1 cubic meter, the network will identify that they have a higher matching degree with trucks with a load capacity of more than 1 ton, thereby avoiding situations where small vehicles cannot load the goods.

[0088] Step 2.4: Generate the task vehicle matching score matrix;

[0089] By aggregating the matching relationships of multiple subspaces through a multi-head attention module, a comprehensive task vehicle matching score matrix is ​​obtained, which is used for subsequent task allocation decisions.

[0090] Step 3: Based on the matching score matrix, a hierarchical reinforcement learning algorithm is used to decompose the global scheduling problem into multi-level sub-problems, establish a reward function, dynamically balance energy efficiency, time delay and operating costs, and generate a hierarchical scheduling strategy.

[0091] Specifically, the following steps are included:

[0092] Step 3.1: Decompose the global logistics scheduling problem into multiple hierarchical sub-problems;

[0093] It includes a regional resource allocation layer, a vehicle scheduling layer, and a route planning layer.

[0094] Regional resource allocation layer: responsible for dividing the entire delivery area into multiple sub-regions and allocating vehicle resources reasonably according to the order density, traffic conditions and timeliness requirements of each region.

[0095] In the specific implementation, the density-based spatial clustering algorithm DBSCAN is used to cluster order points according to their geographical location, forming several delivery sub-regions. The clustering parameter ε (neighborhood radius) is set to 1.5 kilometers, and MinPts (minimum number of points) is set to 10 order points. Then, based on the total number of orders, average delivery time requirement, and traffic congestion index of each region, the regional resource demand index is calculated.

[0096] ;

[0097] in, Indicates the first Resource demand index for each region; Indicates the first Total order volume for each region; This represents the average delivery time requirement (in hours) for orders in the k-th region. This indicates the urgency of the time requirement; the shorter the time requirement, the larger the value. This represents the traffic congestion index of the k-th region, with a value ranging from 0 to 1, where 0 indicates smooth traffic and 1 indicates severe congestion. , , These are weighting coefficients for order volume, timeliness, and traffic conditions, respectively.

[0098] Finally, vehicle resources are allocated to various regions according to the resource demand index to ensure that resource allocation matches actual demand.

[0099] Vehicle dispatching layer: Based on regional resource allocation, this layer is responsible for determining the task sequence and execution order for each vehicle. This layer employs a priority-based task allocation mechanism, first calculating the priority index for each order:

[0100] ;

[0101] in, Indicates order Priority index; Indicates order The remaining time window (hours); This indicates the urgency of the situation; the shorter the remaining time, the larger the value. Indicates order The value (in yuan); This indicates a special attribute indicator for order j; , , These are weighting coefficients for time urgency, order value, and special attributes, respectively. Based on order priority and vehicle characteristics, the Hungarian algorithm is used to solve the optimal matching problem, determine the task sequence for each vehicle, and maximize overall delivery efficiency.

[0102] Path planning layer: For each vehicle's assigned task sequence, optimize the specific driving path to minimize the total driving distance while satisfying the time window constraint.

[0103] This layer employs an improved variable neighborhood search algorithm. First, an initial solution is generated: an initial path is constructed by sorting orders by their time windows. Then, three neighborhood operations (2-pt swap, order insertion / movement, and order pair swap) are used alternately to search for a better solution. A 2-pt swap involves breaking two edges in the path and reconnecting them to form a new path; an order insertion / movement involves removing an order from its current position and inserting it into another position in the path; and an order pair swap involves exchanging the positions of two orders in the path. Each time a better solution is found, the algorithm accepts it and continues searching until no better solution is found after 50 consecutive iterations or the maximum number of iterations (200) is reached. The final output path scheme must simultaneously satisfy vehicle load constraints, time window constraints, and maximum travel distance constraints.

[0104] Step 3.2: Construct the state space, action space, and transition function for each level;

[0105] The state space contains the current state information of all orders and vehicles, the action space contains possible resource allocation decisions and scheduling schemes, and the transfer function describes the state change of the system after performing a specific action.

[0106] Step 3.3, Construct the hierarchical reward function:

[0107] ;

[0108] in, Indicates the overall reward. Indicates energy efficiency indicators. Indicates time delay metrics, Indicates operating cost metrics, , , These are the weighting coefficients for energy efficiency, time delay, and operating cost indicators, used to dynamically balance different optimization objectives.

[0109] Step 3.4: Generate a hierarchical scheduling strategy;

[0110] Deep Q-Network (DQN) is used to train decision models at each level. The stability and convergence efficiency of the algorithm are improved by experience replay and target network techniques, and a hierarchical scheduling strategy is output.

[0111] The hierarchical reinforcement learning algorithm is specifically implemented as a hierarchical deep Q-network structure, which includes the following components:

[0112] Top-level network: Receives the global state and outputs the regional resource allocation strategy. The network contains 3 fully connected layers (256-128-64 neurons).

[0113] The middle layer network receives the state within the region and outputs the vehicle scheduling strategy. The network contains two fully connected layers (128-64 neurons).

[0114] The underlying network receives the bicycle status and outputs a path planning strategy. The network contains two fully connected layers (64-32 neurons).

[0115] In emergency scenarios where priority is given to the delivery of medical supplies, the algorithm can dynamically adjust the weight of the reward function, increase the weight of the time delay indicator, ensure priority delivery of critical supplies, and reduce the complexity of the decision space by hierarchical division, thereby improving the response speed.

[0116] Step 4: Based on the hierarchical scheduling strategy, construct a decentralized collaborative decision-making system to enable each vehicle to make task allocation decisions based on local information and global objectives, and form a preliminary allocation plan;

[0117] Step 4.1: Construct a decentralized collaborative decision-making architecture;

[0118] Based on the results of hierarchical reinforcement learning, a decentralized collaborative decision-making architecture is constructed, enabling each vehicle to make autonomous decisions based on local environmental information.

[0119] The implementation of a decentralized collaborative decision-making system includes three core components:

[0120] Edge computing node construction: An embedded computing unit based on ARM architecture is deployed on each logistics vehicle, configured with a quad-core processor and 8GB of memory, and running a lightweight Linux operating system. This computing unit runs a compressed decision model (the model size is compressed from the original 350MB to 42MB), reducing computational complexity through model pruning and quantization techniques. Each node collects vehicle location, speed, load, and other status data every 3 seconds, and combines this data with pre-loaded local map information to generate preliminary decision suggestions, including the selection of the next delivery point, route planning, and task exchange proposals.

[0121] Local communication network implementation: A hybrid communication architecture of DSRC (Dedicated Short Range Communication) and 4G / 5G is adopted to achieve point-to-point communication between vehicles. DSRC is used for low-latency (<50ms) high-frequency communication within a 300-meter range, mainly transmitting location and intent data; the 4G / 5G network is used for communication over a wider range, transmitting detailed task information and decision data. The communication protocol adopts the lightweight MQTT protocol, with message size controlled within 2KB, including fields such as vehicle ID, location coordinates, current task list, remaining capacity, and decision intent. The message transmission frequency is dynamically adjusted according to the distance between vehicles, ranging from 0.2Hz to 2Hz.

[0122] Consensus Algorithm Module Design: Implements a distributed consensus mechanism based on improved federated averaging. Each vehicle first generates a decision based on local data, and then exchanges decision model parameters among vehicle groups (typically 3-8 vehicles) within the communication range. Upon receiving model parameters from other vehicles, weights are allocated according to each vehicle's workload and historical decision accuracy, and the model parameters are merged by weighted averaging to update the local decision model. The consensus process performs a global synchronization every 10 minutes, triggering additional local synchronization when vehicles encounter new situations (such as road congestion or new urgent orders).

[0123] In practical applications, when traffic congestion occurs in the city center during peak delivery hours, the system automatically detects that the average vehicle speed has dropped below 15 km / h, triggering a regional boundary negotiation mechanism. Vehicles located at the edge of the congested area exchange load information and road condition data through a local communication network to jointly redefine temporary delivery boundaries. Orders at the edge of the congested area are dynamically allocated to vehicles in the surrounding areas, while vehicles within the congested area focus on orders within their own area, reducing unnecessary travel across the congested area. The entire negotiation process is completed within 30 seconds, without the need for a central server, effectively avoiding the 3-5 minute response delay problem faced by traditional centralized dispatch systems during network congestion.

[0124] Step 4.2: Implement the local environment perception module;

[0125] The local environment perception module enables vehicles to obtain order information, traffic conditions, and the status of other vehicles within a certain range.

[0126] Step 4.3, Task allocation decision;

[0127] Based on local information and global objectives, the following optimization formula is applied to make task allocation decisions:

[0128] ;

[0129] in, Describes the optimal decision set. Represents all possible combinations of decisions. Indicates vehicle Decision execution The cost, Indicates vehicle Decision With vehicles Decision The cost of collaboration between them This is a weighting factor for collaboration costs. The total number of vehicles. This indicates the search for decision combinations that minimize the objective function.

[0130] Step 4.4: Formulate a preliminary allocation plan;

[0131] Information exchange and decision-making coordination among vehicles are achieved through local communication networks, forming a preliminary allocation plan.

[0132] Step 5: For the initial allocation scheme, construct a task exchange protocol based on game equilibrium to resolve resource competition conflicts, support dynamic task reallocation and load balancing, and generate the final optimized task allocation scheme.

[0133] Specifically, the following steps are included:

[0134] Step 5.1: Construct a vehicle-to-vehicle task exchange requirement assessment model;

[0135] Calculate the utility function value of each vehicle under the current task allocation scheme.

[0136] Step 5.2: Construct a task exchange protocol based on cooperative game theory;

[0137] Define the exchange conditions, profit distribution rules, and exchange process;

[0138] The specific implementation of the task exchange protocol includes:

[0139] Utility function definition:

[0140] ;

[0141] in, Indicates vehicle The overall utility value, For delivery distance, For time window margin, To improve vehicle capacity utilization, , , These are the weighting coefficients for delivery distance, time window, and vehicle capacity utilization, respectively.

[0142] Exchange strategy generation: An improved Monte Carlo tree search algorithm is used to explore possible task exchange combinations;

[0143] Profit distribution rules: The marginal contribution of each participant is calculated based on the Shapley value to achieve fair distribution.

[0144] In practical applications, when a sudden surge of orders occurs in a certain area, the protocol can automatically trigger task redistribution for surrounding vehicles, avoiding situations where a single vehicle is overloaded while other vehicles are idle. At the same time, it incentivizes vehicles to actively participate in collaboration through a fair revenue distribution mechanism.

[0145] Step 5.3: Calculate the optimal task swapping scheme;

[0146] By applying the Nash equilibrium solution algorithm, the conflict of resource competition is resolved, and the optimal task exchange scheme is calculated.

[0147] Step 5.4: Implement the dynamic task reallocation module;

[0148] When vehicle load is uneven or new orders appear, a task exchange process is triggered to achieve dynamic balance of system load.

[0149] Real-world application examples of this implementation method:

[0150] This implementation method is applied to the last-mile delivery scenario of e-commerce express delivery in City Y. This scenario has the following characteristics: the delivery area covers an urban area of ​​200 square kilometers, with an average daily order volume of about 20,000 orders. The peak order periods are concentrated in two time periods: 10:00 AM to 12:00 PM and 3:00 PM to 5:00 PM. The delivery fleet consists of 150 vehicles of different types, including 60 electric tricycles, 50 electric vans, 30 diesel trucks, and 10 refrigerated trucks.

[0151] The main problems faced in this scenario include: high dispatch pressure during peak delivery periods, uneven allocation of vehicle resources, low inter-regional coordination, and unbalanced vehicle utilization.

[0152] Implementation process example:

[0153] Real-world data examples of vehicle digital twin models:

[0154] This implementation method collected actual characteristic data of various types of vehicles from the logistics fleet in City Y, as shown in Table 1:

[0155] Table 1: Examples of vehicle characteristic data collection (partial);

[0156]

[0157] After standardization, vehicle characteristic vectors were constructed, and digital twin models were built based on these vectors. For example, for an electric van with ID V031, after three months of training with historical data, the digital twin model successfully learned the battery consumption pattern of this model under full load conditions, and the accuracy of predicting the remaining range reached 94.3%, significantly reducing delivery interruptions caused by insufficient battery power.

[0158] Examples of graph attention network applications:

[0159] In a real-world delivery scenario, the system collected order and vehicle matching data from the past 10 days as a training set, containing a total of 189,721 order records. Each record includes features such as order location, delivery time window, and goods type. The graph attention network was trained using the data from the first 5 days, and the data from the last 5 days was used as the test set. The graph attention network learns the vehicle matching relationships for the task, and the prediction results for the matching scores on the test set are shown in Table 2.

[0160] Table 2: Examples of vehicle matching scores for graph attention network tasks;

[0161]

[0162] The model achieved a matching accuracy of 91.2% on the test set, which is 23.6% higher than that of a human scheduler, and it is also able to take into account more dimensions of matching features.

[0163] Results of implementing the hierarchical reinforcement learning scheduling algorithm:

[0164] This implementation divides the delivery area of ​​City Y into 18 sub-regions. A hierarchical reinforcement learning algorithm runs independently within each region, while inter-region coordination is achieved through an upper-layer network. In actual operation, the algorithm dynamically adjusts the weight coefficients of the reward function, increasing the weight of time delay during peak periods and increasing the weight of energy efficiency during off-peak periods. The vehicle allocation in each region after algorithm convergence is shown in Table 3.

[0165] Table 3: Vehicle resource allocation results of the hierarchical reinforcement learning algorithm;

[0166]

[0167] Before the algorithm was implemented, the average vehicle load rate was 64.5%, which increased to 82.8% after implementation, an increase of 18.3 percentage points.

[0168] Implementation results of the decentralized collaborative decision-making system:

[0169] This implementation deploys an edge computing unit on each vehicle, enabling point-to-point communication between vehicles via 4G / 5G networks and vehicle-to-everything (V2X) technology. During system implementation, system response times under peak congestion conditions were recorded, as shown in Table 4.

[0170] Table 4: Comparison of System Response Times (Unit: seconds);

[0171]

[0172] The decentralized collaborative decision-making system reduces the average response time from 14.4 seconds to 3.2 seconds, an improvement rate of 77.8%, with particularly significant advantages in peak congestion scenarios.

[0173] Task exchange protocol running example:

[0174] In practical applications, the task exchange protocol is automatically triggered when unexpected situations such as vehicle malfunctions or a surge in orders occur. In one instance where electric tricycle V005 was unable to continue deliveries due to a battery failure, the system triggered the task exchange process, enabling rapid task redistribution. Surrounding vehicles V008, V003, and V012 jointly took over 6 unfinished orders, with an average delay increase of only 7 minutes, ensuring the continuity of delivery services.

[0175] Technical effectiveness verification:

[0176] After implementing this solution in the last-mile delivery scenario of e-commerce express delivery in City Y for 3 months, the effectiveness of two key technologies was verified:

[0177] Effect of improved transportation capacity utilization:

[0178] Table 5: Comparison of Transportation Capacity Utilization Rate;

[0179]

[0180] As shown in Table 5, this implementation method significantly improved various capacity resource utilization indicators, with an overall capacity utilization rate increase of 18.2%, verifying the effectiveness of the solution.

[0181] Improved system response speed:

[0182] Table 6: Comparison of System Response Speeds;

[0183]

[0184] As shown in Table 6, this implementation significantly reduces system response time in various scenarios, with an average response time reduction of 79.2%. In particular, in complex cross-regional collaboration scenarios, the response time is reduced from 32.5 seconds to 7.4 seconds, demonstrating the advantages of a decentralized collaborative decision-making system.

[0185] Overall system benefit assessment:

[0186] Table 7: Overall System Benefit Assessment (Comparison between before and after 3 months of implementation);

[0187]

[0188] As shown in Table 7, this implementation method significantly reduces operating costs while improving service quality, with an overall operating cost reduction of 27.9%. Meanwhile, service-related indicators such as on-time delivery rate and customer satisfaction increased by 14.3% and 11.5% respectively, achieving a dual improvement in service quality and operational efficiency.

[0189] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A logistics supply chain optimization method based on artificial intelligence, characterized in that, Includes the following steps: A heterogeneous vehicle characteristic vector representation is constructed, which includes load capacity, speed range, energy type, and route limitation parameters. A digital twin model is established for each vehicle. The digital twin model adopts a multi-layer feedforward neural network structure. The input layer receives the vehicle characteristic vector and state data, which are the vehicle's current GPS location coordinates, current load rate, and remaining energy percentage. The intermediate layer adopts a double hidden layer structure. The first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons, both of which use the ReLU activation function. The output layer generates vehicle performance prediction values ​​for predicting remaining mileage, predicted energy consumption rate, and predicted service duration. The vehicle characteristic vector and real-time state data of the digital twin model are input into the graph attention network to calculate the importance weights between tasks and vehicles, learn the optimal matching relationship between different tasks and vehicles, and generate a task-vehicle matching score matrix. The graph attention network contains 8 attention heads, each of which independently learns different feature subspaces. The results are merged by weighted average. The input layer receives the node feature matrix, and the intermediate layer contains two graph convolutional layers. The first layer converts the input features into 64 dimensions, and the second layer keeps the 64-dimensional features unchanged. Each layer uses residual connections. Based on the matching score matrix, a hierarchical reinforcement learning algorithm is used to decompose the global scheduling problem into multi-level sub-problems, establish a reward function, dynamically balance energy efficiency, time delay and operating costs, and generate a hierarchical scheduling strategy. The hierarchical reinforcement learning algorithm decomposes the scheduling problem into three levels: regional resource allocation layer, vehicle scheduling layer, and path planning layer. Based on a hierarchical scheduling strategy, a decentralized collaborative decision-making system is constructed, enabling each vehicle to make task allocation decisions based on local information and global objectives, forming a preliminary allocation scheme. The decentralized collaborative decision-making system includes edge computing nodes, local communication networks, and consensus algorithm modules. The edge computing nodes deploy ARM-based embedded computing units on each vehicle, configured with a 4-core processor and 8GB of memory. The local communication network adopts a DSRC and 4G / 5G hybrid communication architecture. The consensus algorithm module uses an improved federated averaging algorithm to achieve distributed consensus. For the initial allocation scheme, a task exchange protocol based on game equilibrium is constructed to resolve resource competition conflicts, support dynamic task reallocation and load balancing, and generate a final optimized task allocation scheme. The task exchange protocol based on game equilibrium includes three components: utility function definition, exchange strategy generation, and revenue distribution rules. The revenue distribution rules calculate the marginal contribution of each participant based on Shapley values ​​to achieve fair allocation.

2. The logistics supply chain optimization method based on artificial intelligence according to claim 1, characterized in that, In the hierarchical reinforcement learning algorithm, a state space, action space, and transition function are constructed for each level, and a deep Q-network is used to train the decision model for each level.

3. The logistics supply chain optimization method based on artificial intelligence according to claim 1, characterized in that, The method also includes dynamically adjusting the weight coefficients of the reward function, increasing the weight of time delay during peak periods and increasing the weight of energy efficiency during off-peak periods.

4. An artificial intelligence-based logistics supply chain optimization system, used to execute an artificial intelligence-based logistics supply chain optimization method as described in any one of claims 1-3, characterized in that, include: The heterogeneous vehicle digital twin model construction module is used to build vehicle characteristic vector representations and establish digital twin models. The characteristic vectors include load capacity, speed range, energy type and path limitation parameters. The graph attention network module is used to calculate the importance weights between tasks and vehicles, learn the optimal matching relationship between different tasks and vehicles, and generate a task-vehicle matching score matrix. The hierarchical reinforcement learning scheduling module is used to decompose the global scheduling problem into multi-level sub-problems, establish a reward function, dynamically balance energy efficiency, time delay and operating costs, and generate hierarchical scheduling strategies. The decentralized collaborative decision-making module enables each vehicle to make task allocation decisions based on local information and global objectives, forming a preliminary allocation plan; The game equilibrium task exchange module is used to resolve resource competition conflicts, support dynamic task redistribution and load balancing, and generate the final optimized task allocation scheme.

Images