Big data analysis method and system based on traffic transportation logistics

Abstract

The invention relates to the technical field of logistics distribution, and discloses a big data analysis method and system based on traffic transportation logistics, and the method comprises the steps: constructing a space-time diagram convolution network model, and predicting the traffic congestion propagation trend of an urban region; based on the congestion propagation trend, the order density and the distribution resource state, executing a space-time dynamic partitioning algorithm, and dividing the urban distribution region into a plurality of dynamic micro-regions; constructing a double-layer intelligent decision-making system, and generating a logistics fleet separation and combination decision and path planning; heterogeneous distribution resource dynamic marshalling is executed, and different types of distribution tools are combined into a miniature formation; using a multi-agent reinforcement learning algorithm to optimize the cooperation process in the micro formation; according to the invention, through fusion of congestion propagation prediction, dynamic micro-area division, double-layer intelligent decision and heterogeneous resource collaboration, key problems of traffic congestion, resource scheduling and'last kilometer 'distribution in urban logistics distribution are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of transportation and logistics, and more specifically, to a big data analysis method and system based on transportation and logistics. Background Art

[0002] With the acceleration of urbanization and the rapid development of e-commerce, the demand for urban logistics and distribution has exploded, generating massive amounts of logistics big data. However, the increasingly serious problem of urban road traffic congestion has brought huge challenges to logistics and distribution. The existing logistics and distribution systems mainly have the following technical problems: First, traditional logistics systems rely on independent route planning and lack a vehicle coordination mechanism based on big data analysis. This makes it impossible to fully leverage real-time data streams for intelligent decision-making, leading to traffic congestion and low delivery efficiency in high-density delivery areas. When multiple delivery vehicles simultaneously enter the same area, not only does this exacerbate existing traffic congestion, but it also causes delivery vehicles to compete for limited road resources, reducing overall delivery efficiency.

[0003] Second, existing logistics vehicle platooning technology lacks the ability to deeply mine massive amounts of historical traffic data, preventing it from predicting congestion trends and proactively adjusting platooning strategies. Most logistics systems rely solely on reactive adjustments based on real-time traffic conditions, failing to apply data mining and machine learning techniques to predict the development and spread of congestion. This often results in fleets reacting passively to existing congestion, missing the optimal opportunity for adjustment.

[0004] Furthermore, last-mile delivery areas face complex road conditions and a dense network of delivery points. Existing systems fail to fully leverage multi-source, heterogeneous data analysis to optimize the collaborative potential of diverse delivery resources (such as trucks, vans, electric tricycles, and drones), hindering efficient delivery. Traditional logistics and delivery systems often rely on a single type of delivery resource and lack intelligent scheduling mechanisms based on big data analytics. These systems struggle to adapt to the complex and ever-changing urban delivery environment, particularly in core urban areas where roads vary in width and access restrictions.

[0005] The existing technology lacks a method that can comprehensively solve the above problems. It is urgent to develop an intelligent logistics distribution solution based on big data analysis and artificial intelligence technology that can predict the spread trend of traffic congestion, realize dynamic spatiotemporal distribution area division, conduct intelligent fleet management, and optimize the collaboration of heterogeneous distribution resources. Summary of the Invention

[0006] The present invention provides a big data analysis method and system based on transportation logistics, which solves the technical problems in related technologies such as the inability of logistics distribution systems to effectively cope with urban traffic congestion, the lack of vehicle coordination mechanisms, and the irrational use of distribution resources.

[0007] The present invention provides a big data analysis method based on transportation logistics, comprising: Construct a congestion propagation prediction model based on spatiotemporal graph convolutional networks, analyze historical traffic data, and predict future congestion trends in urban areas; Implement a spatiotemporal dynamic partitioning algorithm to divide the urban delivery area into multiple dynamic micro-zones based on congestion prediction results and delivery order distribution; Build a two-layer intelligent decision-making system to generate logistics fleet splitting and joining decisions and route planning based on congestion prediction and micro-area division; Dynamically group heterogeneous distribution resources, combining different types of distribution tools to form an optimal micro-fleet based on micro-area characteristics and order characteristics; Based on the formation results, a multi-agent reinforcement learning algorithm is applied to optimize the formation collaboration and realize adaptive formation and separation of vehicles.

[0008] In a preferred embodiment, the spatiotemporal dynamic partitioning algorithm includes: Collect and process delivery order data and available logistics resource data to form order density matrix and resource availability matrix; Construct a spatiotemporal partitioning evaluation function that integrates congestion prediction results, order density, and resource availability; The improved spatiotemporal DBSCAN algorithm is applied to cluster the evaluation functions and the spatiotemporal distance metric is defined; Based on the clustering results, the city's dynamic micro-division is determined and the final dynamic micro-division results are output.

[0009] In a preferred embodiment, the two-layer intelligent decision-making system includes: A high-level decision-making system is used to determine the set of key nodes suitable for fleet decomposition based on the congestion prediction heat map and micro-division results, and generate a fleet decomposition strategy; A low-level decision-making system is used to generate detailed delivery routes for each sub-fleet and corresponding delivery micro-area; The decomposition decision results of the high-level system are passed to the low-level system as constraints for path planning, while the path execution effect and real-time feedback of the low-level system are passed back to the high-level system for dynamic adjustment of subsequent decomposition decisions.

[0010] In a preferred embodiment, the dynamic grouping of heterogeneous distribution resources includes: Establish a distribution resource capability model to describe the performance characteristics of different types of distribution tools; Construct a micro-area order feature analysis model to extract the key features of the order set; Construct resource micro-area matching algorithm and calculate matching scores; Combinatorial optimization algorithms are applied to determine the optimal heterogeneous resource combination for each micro-area.

[0011] In a preferred embodiment, the multi-agent reinforcement learning algorithm includes: Build a vehicle agent model and define the state space, action space, and observation space; Build an agent collaboration model based on maximizing mutual information to enhance information sharing and decision-making collaboration among agents; Construct a reward function based on expected time benefits to guide the agent to learn the optimal strategy; A hierarchical multi-agent reinforcement learning algorithm is applied to train the team collaboration strategy, adopting the Actor-Critic architecture.

[0012] In a preferred embodiment, the mutual information optimization objective in the agent collaboration model based on mutual information maximization is defined as: ; in, Indicates that in a known agent Status Under these conditions, the agent Status With the agent Action The conditional mutual information between and Represents the intelligent agent and agents status, Representing an agent action, p represents the probability, Represents a logarithmic function, which measures the agent The state of the agent The degree of influence on action decision.

[0013] In a preferred embodiment, the reward function based on expected time benefit is defined as: ; in, Indicates that the status Next action The reward value obtained, and Represent the joint action and state at time t, For the The weight of an order, is the baseline delivery time, For the current action and state The expected completion time of each order, Indicates the total number of orders waiting to be delivered in the current system.

[0014] In a preferred embodiment, the spatiotemporal partitioning evaluation function is defined as: ; in, Indicates time and location The time and space partition evaluation function value of Represents the position coordinates, Indicates a point in time, Indicates location In time The order density function of Indicates location In time Traffic condition function, Indicates time Resource availability function.

[0015] In a preferred embodiment, the decision function in the high-level decision system is defined as: ; in, Indicates that the node and time The fleet decomposition decision function, Represents a road network node, Indicates a point in time, Representation node In time The congestion probability, Representation node In time The micro-region to which it belongs, Indicates the current fleet resource status, Represents the decision mapping function.

[0016] In a preferred embodiment, a transportation logistics-based big data analysis system is used to perform a transportation logistics-based big data analysis method, including: The congestion propagation prediction module is used to build a congestion propagation prediction model based on spatiotemporal graph convolutional networks, analyze historical traffic data, and predict future congestion trends in urban areas; The spatiotemporal dynamic partitioning module is used to implement the spatiotemporal dynamic partitioning algorithm, which divides the urban delivery area into multiple dynamic micro-zones based on congestion prediction results and delivery order distribution; A two-layer intelligent decision-making module is used to generate logistics fleet splitting and joining decisions and route planning based on congestion prediction and micro-area division; Heterogeneous resource grouping module, used to dynamically group heterogeneous distribution resources. Based on micro-area characteristics and order characteristics, different types of distribution tools are combined to form an optimal micro-formation. The multi-agent collaboration module is used to apply multi-agent reinforcement learning algorithms to optimize platoon collaboration and achieve adaptive vehicle platooning and separation.

[0017] The beneficial effects of the present invention are: The spatiotemporal graph convolutional network model of the present invention can accurately predict the congestion development trend of urban areas in the next 1-3 hours, improving the prediction accuracy, providing a reliable basis for logistics fleets to avoid congested sections in advance, reducing the overall driving distance and shortening the waiting time in congested areas.

[0018] The spatiotemporal dynamic partitioning algorithm of the present invention can divide the urban distribution area into micro-regions with similar characteristics according to real-time traffic conditions and order distribution, thereby improving the consistency of order processing within the region, making route planning more accurate, distribution resource allocation more reasonable, and improving vehicle utilization.

[0019] The dual-layer intelligent decision-making system of the present invention realizes the dynamic adjustment of the logistics fleet by combining the fleet splitting and joining decision-making with path planning. The system can decompose a large logistics fleet into multiple micro heterogeneous formations in advance before predicting the congested area, thus avoiding the congested area and improving the overall traffic efficiency of the fleet.

[0020] The dynamic grouping technology of heterogeneous distribution resources in the present invention solves the problems of complex road conditions and dense distribution points in last-mile distribution. By combining different distribution tools such as trucks, vans, electric tricycles, and drones into an optimal formation, the most appropriate distribution method is selected for different types of distribution areas, thereby improving the delivery success rate and reducing energy consumption per unit order.

[0021] The multi-agent reinforcement learning algorithm of the present invention optimizes the collaborative behavior within the formation and realizes the adaptive formation and separation of vehicles. The system can automatically adjust the delivery strategy according to real-time traffic conditions and order changes. It has strong adaptability and robustness, and reduces delivery delay time when responding to sudden traffic conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of a big data analysis method based on transportation logistics of the present invention. DETAILED DESCRIPTION

[0023] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0024] At least one embodiment of the present invention discloses a big data analysis method based on transportation logistics, such as Figure 1 As shown, the following steps are included: Step 1: Build a congestion propagation prediction model based on spatiotemporal graph convolutional networks, analyze historical traffic data, and predict future congestion trends in urban areas. This step uses the Spatiotemporal Graph Convolutional Network (ST-GCN) to process historical traffic data and generate a forecast of urban area congestion trends within the next 1-3 hours. It includes the following sub-steps: 1.1. Collect urban traffic network data and historical traffic flow data, including road topology, historical traffic flow, vehicle speed, traffic incident records, etc., to form a spatiotemporal traffic dataset ,in, represents a spatiotemporal traffic dataset, represents the traffic network diagram at time t, represents the traffic flow characteristic matrix at time t, Represents a collection of time series.

[0025] 1.2. Construct a traffic network graph structure and represent the urban traffic network as a weighted directed graph ,in, Represents the graph structure model of the entire traffic network, Represents a set of road network nodes, Represents a set of road segments, It represents the adjacency matrix, characterizing the connection relationship and weight between nodes. The node characteristics include time-varying characteristics such as road traffic volume, average speed, and road grade.

[0026] 1.3. Apply a spatiotemporal graph convolutional network model to process traffic data. This model combines a graph convolutional network (GCN) and a temporal convolutional network (TCN) to capture the spatial dependency and temporal evolution characteristics of traffic data. The graph convolution operation is defined as: ; in, Represents the node feature matrix output after the graph convolution operation, To add a self-connected adjacency matrix, is the original adjacency matrix representing the road network connection relationship, is the identity matrix, is the corresponding degree matrix, and are the input and output features, is the learnable weight matrix, is the sigmoid activation function.

[0027] The spatiotemporal graph convolutional network model in this implementation adopts a multi-layered design, consisting of an architecture that alternates spatial graph convolutional layers and temporal convolutional layers. Specifically, the model contains four spatiotemporal convolutional blocks, each consisting of a graph convolutional layer and a temporal convolutional layer.

[0028] The spatial graph convolution layer is responsible for capturing the spatial dependencies between different nodes in the road network, taking into account factors such as road connectivity and traffic flow transfer patterns; The temporal convolution layer captures the temporal patterns of traffic data through one-dimensional convolution operations, such as the characteristics of peak hours in the morning and evening, and the differences between weekdays and weekends.

[0029] The model also introduces an attention mechanism to weight important spatiotemporal features and enhance the ability to predict congested propagation paths.

[0030] In practical application scenarios, the model receives historical traffic flow and speed data from the city's main road network as input, processes traffic state sequences (sampled every 5 minutes) at approximately 500 key road nodes within Beijing's Fifth Ring Road, and outputs predictions of congestion probabilities for each node within the next 1-3 hours. For example, when predicting the spread of traffic congestion on roads surrounding several commercial areas in Chaoyang District, Beijing, the model was able to identify a pattern in which congestion first begins at shopping mall exits and then spreads along main roads to adjacent areas. Furthermore, based on traffic conditions at 3 p.m. on weekdays, the model accurately predicted the dynamic expansion of congested areas during the evening rush hour from 5 p.m. to 7 p.m., providing precise decision-making basis for logistics fleet planning.

[0031] 1.4. Based on the processed data, identify the congestion diffusion pattern and calculate the congestion probability of each road section in the future time period through the model , representing a node In the future The congestion probability at time hours, a congestion propagation prediction heat map is generated to provide a basis for subsequent decision-making.

[0032] Step 2: Implement a spatiotemporal dynamic partitioning algorithm to divide the urban delivery area into multiple dynamic micro-zones based on congestion prediction results and delivery order distribution; This step uses the spatiotemporal dynamic partitioning algorithm to generate dynamic micro-division results for the urban delivery area based on the congestion prediction results, delivery order density, and available resource data from step 1. Specifically, it includes the following sub-steps: 2.1. Collect and process delivery order data and available logistics resource data, including order location, time window, delivery type, and other information, as well as the location, type, and load of various delivery vehicles, to form an order density matrix. and resource availability matrix .

[0033] 2.2, construct a spatiotemporal partition evaluation function, integrating congestion prediction results, order density, and resource availability, and define it as: ; in, represents the spatiotemporal partition evaluation function, Represents the position coordinates, Indicates a point in time, Indicates location In time The order density function of Indicates location In time Traffic condition function, Indicates time Resource availability function.

[0034] 2.3, Apply clustering algorithm to the evaluation function Cluster analysis is performed using the improved spatiotemporal DBSCAN (density clustering) algorithm, and the spatiotemporal distance metric is defined as: ; in, represents the spatiotemporal distance metric function, represents the first spacetime point, is the spatial position coordinate, For time point, represents the second spacetime point, is the spatial position coordinate, For time point, is the spatial distance function, is the time distance function, and is the weight parameter for balancing the spatial and temporal dimensions.

[0035] The spatiotemporal dynamic partitioning algorithm in this implementation makes several improvements to the traditional DBSCAN algorithm, making it more suitable for dynamic partitioning requirements in urban logistics scenarios. Specific improvements include: Introducing adaptive density thresholding , so that it changes dynamically with location and time, automatically lowering the threshold in areas with high order density to form a finer-grained micro-region division; Design a gradual boundary processing mechanism to avoid the problem of discontinuous order processing between adjacent micro-zones caused by traditional hard boundary division; A dynamic adjustment strategy is adopted to increase the weight of the time dimension, increase the weight of the time dimension during traffic peak hours, and pay more attention to the time continuity of micro-areas.

[0036] In practical applications, the algorithm can flexibly adjust micro-district divisions based on order distribution and traffic conditions at different times. For example, during the morning rush hour in Xuhui District, Shanghai, the algorithm split the business district, which would have been classified as a single large area, into three micro-districts based on the traffic congestion predicted in the previous step. Each micro-district was then assigned a different type of delivery resource: micro-districts adjacent to main roads were primarily delivered by medium-sized trucks, while micro-districts with narrower internal roads were primarily delivered by electric tricycles. Furthermore, over time, when it detects that traffic congestion in a particular area has eased, the algorithm dynamically merges adjacent micro-districts to optimize resource allocation, significantly improving logistics and delivery efficiency.

[0037] 2.4, Based on the clustering results, determine the urban dynamic micro-district division, each micro-district Contains points with similar spatiotemporal characteristics, where Indicates the micro-districts, Represents a point in spacetime, Represents the first clusters, taking into account the smoothness and connectivity of the micro-interval boundaries, and outputting the final dynamic micro-area partitioning results ,in, 、 、 Respectively represent 、 、 micro-districts, is the number of micro-areas, which is determined by the algorithm adaptively.

[0038] Step 3: Build a two-layer intelligent decision-making system to generate logistics fleet splitting and joining decisions and route planning based on congestion prediction and micro-area division; This step, based on the congestion prediction results of step 1 and the micro-division of step 2, generates logistics fleet splitting and joining decisions and detailed route planning through a two-layer decision-making system. It specifically includes the following sub-steps: 3.1. Build a high-level decision model that integrates the congestion prediction heat map and micro-district segmentation results to determine the key node set suitable for fleet decomposition. The decision function is: ; in, Indicates that the node and time The decision function value for fleet decomposition at Represents a road network node, Indicates a point in time, represents the decision function, Representation node In time The congestion probability, Representation node In time The micro-region to which it belongs, Indicates the current fleet resource status.

[0039] 3.2. Based on the high-level decision-making results, a fleet decomposition strategy is generated for each decomposition point, including the decomposition timing, decomposition structure, and resource allocation after decomposition, and a decomposition plan is generated: ; in, represents the set of fleet decomposition solutions, represents the decomposition nodes and time, Indicates the composition of the fleet before decomposition, Indicates the multiple sub-fleet components formed after decomposition.

[0040] 3.3. Build a low-level decision model and use an adaptive path planning algorithm to generate detailed delivery routes for each sub-fleet and corresponding delivery micro-area: ; in, represents the path planning function, represents a sub-team, Indicates the corresponding delivery micro-area, Represents the generated path set, 、 、 Respectively represent 、 、 Planning paths, Indicates the total number of paths.

[0041] The two-layer intelligent decision-making system in this embodiment adopts a layered architecture design, where the high-level decision-making system and the low-level decision-making system respectively assume different decision-making responsibilities, and collaborative decision-making is achieved between the two-layer systems through a two-way information flow.

[0042] The high-level decision-making system handles global fleet splitting and joining decisions from a macro perspective. Its core consists of a decision-making model based on the Graph Attention Network (GAT). This model treats road network nodes as nodes in the graph and road connections as edges. It uses the attention mechanism to capture the influence relationship between different nodes and determine the key fleet decomposition nodes.

[0043] The low-level decision-making system focuses on the micro level and is responsible for specific path planning. It uses an improved A* algorithm combined with real-time traffic information to search for paths, and introduces the concept of "safety time margin" to reserve buffer time for each path to deal with sudden traffic conditions.

[0044] Information exchange between the two systems is achieved through a two-way feedback mechanism: the decomposition and decision-making results of the higher-level system are transmitted to the lower-level system as constraints for route planning. The route execution results and real-time feedback from the lower-level system are then transmitted back to the higher-level system to dynamically adjust subsequent decomposition decisions. The system also features a conflict detection and resolution module. When potential conflicts are detected in the route plans of different sub-fleet fleets, a coordination mechanism is automatically triggered to reallocate routes or adjust schedules.

[0045] In practical application scenarios, this two-tier system has performed exceptionally well in Guangzhou's logistics and distribution operations. For example, when handling deliveries from the Baiyun Logistics Center to various commercial locations in the city center, the system, based on congestion predictions, preemptively split a fleet of five large trucks into three sub-fleets at the intersection of Tianhe Road and Zhujiang New Town: two medium-sized trucks for the northern Tianhe District, a mixed fleet consisting of two small trucks and several electric tricycles for the high-density commercial areas of Zhujiang New Town, and one large truck for the more distant Haizhu District. Guided by the lower-tier system, each sub-fleet completed its delivery mission within its respective micro-district along the optimal route, reducing overall delivery time by 28% compared to traditional methods. In particular, the system can quickly re-route delivery routes in response to unexpected traffic control situations to ensure timely delivery.

[0046] 3.4. Integrate high-level and low-level decision-making results to form a complete distribution execution plan, including fleet decomposition nodes, sub-fleet composition and distribution paths of each sub-fleet, and generate an executable logistics scheduling plan.

[0047] Step 4: Dynamically group heterogeneous delivery resources. Based on micro-area characteristics and order characteristics, different types of delivery tools are combined to form an optimal micro-fleet. Based on the micro-division in step 2 and the decomposition decision in step 3, this step combines different types of delivery resources (such as trucks, vans, electric tricycles, drones, etc.) into optimal micro-fleet to match the corresponding order clusters. It includes the following sub-steps: 4.1. Establish a distribution resource capability model to describe the performance characteristics of different types of distribution tools, including parameters such as load capacity, driving speed, access restrictions, and energy consumption, to form a resource capability description matrix. ,in, represents the distribution resource capability description matrix, Indicates the Class resources are in Ability value on each performance dimension.

[0048] 4.2, build a micro-area order feature analysis model for each micro-area , extract the key features of its order set, including order density, time urgency, package size distribution, etc., and generate order feature vector ,in, Represents micro-area The order feature vector of 、 、 They represent the order feature vector 、 、 dimensional components, Represents the total number of dimensions of the order feature vector.

[0049] 4.3. Build a resource micro-area matching algorithm and calculate the matching score based on resource capabilities and micro-area order characteristics: ; in Indicates the Class distribution resources and The matching score of each micro-area, Indicates the Class resources, Indicates the micro-districts, is the weight of different feature dimensions, is the matching function on a specific dimension, Indicates the Class resources are in The capability value of each performance dimension, Indicates the The micro-area The value of the order feature dimension, Indicates the total number of feature dimensions.

[0050] 4.4, Apply combinatorial optimization algorithms to solve the resource allocation problem and determine the optimal heterogeneous resource combination for each micro-zone: ; in, Represents micro-area The optimal resource combination solution, Show all possible resource combinations In the above example, we look for the combination that maximizes the objective function. Represents a resource combination, Indicates the Class distribution resources, Indicates the Class distribution resources and The matching score of each micro-area, Represent the cost function of the combination, including resource usage cost and coordination cost. Output the optimal formation configuration plan for each micro-area.

[0051] The dynamic grouping algorithm for heterogeneous distribution resources in this implementation utilizes a method that combines mixed integer programming with heuristic search to solve combinatorial optimization problems in high-dimensional resource allocation spaces. The core of the algorithm consists of two components: a resource-task matching evaluation module and a resource combination generation module. The matching evaluation module employs a multi-dimensional scoring mechanism to calculate a matching score for each pair of resource micro-area combinations. The scoring dimensions include load matching, timeliness, energy efficiency, and traffic adaptability. The resource combination generation module, based on a genetic algorithm framework, employs a specialized encoding scheme to represent heterogeneous resource combinations. It then searches for the global optimal solution through iterative optimization, while also incorporating a simulated annealing strategy to avoid falling into local optimality.

[0052] A key innovation of the algorithm is the introduction of a resource complementarity evaluation mechanism. When different types of distribution resources are combined, their complementary gains are calculated. For example, when large trucks are combined with drones, both trunk transportation efficiency and terminal distribution flexibility can be taken into account, giving the combination a higher complementarity score. Another innovation is the dynamic adjustment strategy, which periodically re-evaluates resource allocation plans based on real-time order changes and traffic conditions, and triggers resource reorganization when necessary.

[0053] In practical logistics and delivery practices in Shenzhen, this algorithm has demonstrated exceptional adaptability. For example, in the high-density commercial areas of Futian District, the system divides the area into distinct micro-zones and configures each with a customized heterogeneous resource combination: Within busy business districts, an "electric truck + walking delivery driver" combination is employed to improve delivery efficiency in the last 100 meters; near residential areas, a "van + electric tricycle" combination is employed to balance delivery efficiency and noise control; and in corporate parks, a "truck + drone" combination is employed to address the coexistence of large express deliveries and urgent small items. This customized heterogeneous grouping strategy not only improves delivery success rates but also significantly enhances customer satisfaction and energy efficiency.

[0054] Step 5: Based on the formation results, a multi-agent reinforcement learning algorithm is applied to optimize the formation collaboration and realize the adaptive formation and separation of vehicles; This step is based on the heterogeneous formation formed in step 4 and applies a multi-agent reinforcement learning algorithm to further optimize the collaborative behavior within the formation to achieve adaptive formation and separation of vehicles. It specifically includes the following sub-steps: 5.1, build a vehicle agent model, treat each delivery tool in the formation as an agent, and define its state space (including position, speed, remaining load, remaining energy, etc.), action space (including movement, loading and unloading, formation, separation, etc.) and observation space (Contains local environment perception information).

[0055] 5.2. Construct an agent collaboration model based on mutual information maximization to enhance information sharing and decision-making coordination among agents and define the mutual information optimization objective: ; in, Indicates that in a known agent Status Under these conditions, the agent Status With the agent Action The conditional mutual information between and Represents the intelligent agent and agents status, Representing an agent action, represents the probability, Represents a logarithmic function, which measures the agent The state of the agent The degree of influence on action decision.

[0056] 5.3. Construct a reward function based on expected time benefits to guide the agent to learn the optimal strategy: ; in, Indicates at time Status Perform joint actions The reward value obtained, and Represent the joint action and state at time t, For the The weight of an order, is the baseline delivery time, For the current action and state The expected completion time of each order, Indicates the total number of orders waiting to be delivered in the current system.

[0057] 5.4. We apply a hierarchical multi-agent reinforcement learning algorithm to train the team collaboration strategy, using an actor-critic architecture where the actor network outputs the action probability distribution of each agent: ; in, Indicated by the parameter Determine the policy function, Representing an agent action, Representing an agent Local observation, Indicates that the parameter is Actor network; Critic network evaluation joint state action value function: ; in, Indicated by the parameter Determined state-action-value function, represents the global state of the environment, represents the joint action of all agents, Indicates that the parameter is Critic network; Optimize network parameters through policy gradient method and , generating the optimal formation cooperation strategy.

[0058] The multi-agent reinforcement learning algorithm in this implementation is based on a hierarchical actor-critic architecture and incorporates several innovations tailored to the collaborative decision-making characteristics of heterogeneous distribution resources in logistics and distribution scenarios. The algorithm incorporates an attention communication mechanism to establish a dynamic connection structure between agents, fostering stronger information exchange channels between highly correlated agents. It also employs a hierarchical policy network structure, comprising two layers: formation-level policies responsible for macro-coordination and individual-level policies responsible for micro-execution. It employs a centralized training-distributed execution framework, utilizing global information optimization strategies during training while relying solely on local observations for decision-making during execution.

[0059] A core innovation of the algorithm is the mutual information maximization training mechanism. By maximizing the mutual information between an agent's state and the behavior of other agents, it encourages agents to learn strategies that consider each other's behavior and enhances their collaborative capabilities. Another innovation is a tiered reward design based on expected time returns. This approach combines global and local rewards. Global rewards focus on the efficiency of the entire delivery task, while local rewards focus on the execution quality of individual agents. This multi-layered reward mechanism guides agent behavior towards overall optimization.

[0060] In delivery applications around the scenic area of ​​West Lake District in Hangzhou, the algorithm demonstrated excellent collaborative capabilities. On weekends when tourists flocked to the area, the system organized a mixed fleet of trucks, electric tricycles, and drones for delivery tasks around the West Lake area. When encountering temporary traffic control on the north line of West Lake, the system automatically adjusted its strategy: the area originally handled by trucks was taken over by multiple electric tricycles; drones were reassigned to the most timely delivery points; and the remaining trucks automatically rerouted to avoid the controlled areas. The entire adjustment process required no human intervention. Each agent autonomously completed role switching and task reallocation through the learned collaborative strategy, ensuring the continuity and timeliness of delivery tasks and reducing delays by 42% compared to traditional methods.

[0061] Real-world application examples of this implementation: This implementation has been implemented in Chengdu's logistics and distribution system. The system serves a large e-commerce logistics company covering Chengdu's central urban area (including Jinjiang, Qingyang, Jinniu, Wuhou, and Chenghua districts), handling an average daily delivery volume of approximately 50,000 orders. The company has three major logistics centers located outside Chengdu's central urban area and 12 distribution sites within the city. The system primarily optimizes trunk line transportation from logistics centers to distribution sites and the "last mile" of delivery from distribution sites to customers.

[0062] As a transportation hub and commercial center in southwest China, Chengdu faces multiple challenges in logistics and distribution: Traffic congestion in the main urban area is severe, especially during peak hours in the morning and evening, and traditional delivery methods are inefficient; Chengdu is densely populated with commercial areas, residential areas, and schools, resulting in significant differences in delivery demand and traffic conditions across different areas. Emerging areas such as Tianfu New Area and High-tech Zone are developing rapidly, the volume of delivery orders is growing fast, and the distribution of delivery resources is uneven.

[0063] Before implementing this technology, the company's logistics and distribution primarily relied on a traditional model with fixed vehicle types and routes, resulting in low delivery efficiency and low customer satisfaction. Average delivery time was approximately 4.2 hours, average vehicle utilization was only 65%, and order delays in congested areas during peak hours reached as high as 38%. After implementing this technology, the system integrated historical traffic data, real-time road conditions, order data, and delivery resource information, achieving intelligent delivery optimization throughout the entire process.

[0064] Implementation process of the congestion propagation prediction model: The system first collected traffic data from Chengdu's main urban area over the past three years, including information on traffic volume, speeds, and traffic incidents on major roads. Data sources included traffic surveillance cameras, vehicle GPS, and roadside sensors. Through data cleaning and preprocessing, a traffic network diagram was constructed, encompassing 385 key road nodes in Chengdu's main urban area. Data was sampled every five minutes.

[0065] In terms of model architecture, the system implements a four-layer spatiotemporal graph convolutional network model, with each layer comprising graph convolution and temporal convolution components. The model uses 32 feature channels, a graph convolution kernel size of 3, a temporal convolution kernel size of 3, and a stride of 1. To improve model performance, the system also introduces skip connections and batch normalization layers.

[0066] Table 1 shows the performance comparison of the model under different time prediction windows: Table 1: Performance comparison of congestion propagation prediction models under different prediction windows; The system accurately predicts the congestion propagation path and time nodes, providing an important basis for subsequent logistics and distribution decisions.

[0067] In actual applications, the model can accurately predict severe congestion in the Chunxi Road business district from 10:00 to 11:00 based on the current traffic conditions at 8:00, providing an early warning to the logistics distribution system three hours in advance, allowing the fleet to adjust routes or schedules in time to avoid potential congested areas.

[0068] Implementation process of spatiotemporal dynamic partitioning algorithm: The system implements dynamic spatiotemporal zoning based on congestion propagation predictions and delivery order distribution data. It processes approximately 50,000 delivery orders daily, including recipient address, time window, package type, and weight. It also maintains real-time status information for various delivery resources, including 38 large trucks at three logistics centers, 86 small and medium-sized delivery vehicles at 12 delivery stations, 45 electric tricycles, and 12 drones.

[0069] The system's spatiotemporal dynamic partitioning algorithm uses a modified DBSCAN clustering method. Key parameters include: minimum sample size (MinPts) = 8, spatial distance threshold (εspace) = 1.2 km, temporal distance threshold (εtime) = 30 minutes, and spatial and temporal weighting parameters (α) = 0.7 and (β) = 0.3. The algorithm runs every 30 minutes, dynamically adjusting micro-zoning based on the latest congestion forecasts and order distribution.

[0070] Table 2 shows some micro-area feature data generated by the system at 9:00 am on a weekday: Table 2: Characteristic data of some micro-districts in Chengdu’s main urban area (9:00 on weekdays); A key optimization of the algorithm is the introduction of a gradual boundary processing mechanism, which creates a 0.3-0.5 km transition zone between adjacent micro-zones. This avoids the discontinuous order processing issues that can occur with traditional hard boundary partitioning. This dynamic partitioning method improves delivery efficiency by 31.5% compared to traditional static zone partitioning, and the effect is particularly significant in areas with rapidly changing traffic conditions.

[0071] Implementation process of the two-layer intelligent decision-making system: The system's two-tiered intelligent decision-making system, based on congestion predictions and micro-zoning, enables fleet splitting and consolidation decisions and route planning. The system comprises high-level and low-level decision-making modules, which work together to combine global optimization with localized, refined scheduling.

[0072] In the high-level decision-making module, the system employs a decision-making model based on a graph attention network (GAT). This model uses key road network nodes in Chengdu's main urban area as graph nodes and uses an attention mechanism to capture the influence relationships between different nodes. The model's inputs include congestion prediction heat maps and micro-district segmentation results, and its outputs are key fleet decomposition nodes and decomposition strategies. The model gradually optimizes its decision-making strategy by learning from decomposition decision-making experience in historical transportation data.

[0073] Table 3 shows an example of a fleet split decision generated by the system on a weekday afternoon: Table 3: Example of logistics fleet decomposition decision (weekday 14:30); The low-level decision-making module generates detailed delivery routes for each sub-fleet based on the decomposed decisions made by the high-level modules. The system uses a modified A* algorithm, combined with real-time traffic information, for route search. It also incorporates a "safety time margin" concept, allowing a 5-15 minute buffer for each route to accommodate unexpected traffic conditions. The system plans different routes for different delivery vehicles, such as selecting routes that allow for non-motorized vehicle lanes for electric tricycles and planning the shortest straight-line flight paths for drones.

[0074] The two-tier systems maintain information exchange through a two-way feedback mechanism: When the lower-tier system detects that the actual travel time on a particular road section exceeds the expected time by more than 20%, it immediately provides feedback to the higher-tier system, triggering it to reassess its fleet split decision. Similarly, when the higher-tier system makes a new split decision, it immediately notifies the lower-tier system to adjust the route plan. This two-way feedback mechanism enables the system to quickly respond to changing traffic conditions and maintain efficient delivery.

[0075] In practical applications, this two-tiered decision-making system has demonstrated strong adaptability. For example, during a temporary urban traffic control incident in June 2023, the system detected the closure of several roads in Jinjiang District. The high-level decision-making module immediately adjusted the fleet split plan, moving the fleet, originally planned for the city center, to the Second Ring Road and reallocating delivery resources. The low-level decision-making module then rerouted each fleet, ultimately keeping delivery delays to under 15 minutes. Meanwhile, other logistics companies not using the system experienced delays exceeding 90 minutes.

[0076] Implementation process of dynamic grouping of heterogeneous distribution resources: Based on micro-district division and logistics fleet decomposition decisions, the system enables dynamic grouping of heterogeneous delivery resources. This function aims to combine different types of delivery vehicles (such as trucks, vans, electric tricycles, drones, etc.) into optimal micro-fleet to match the characteristics of different micro-districts and order requirements.

[0077] The system first establishes a detailed distribution resource capability model, describing the performance characteristics of various distribution tools owned by the enterprise. Table 4 shows the capability parameters of some distribution resources in the system: Table 4: Distribution resource capacity parameter table; Next, the system constructed an order feature analysis model for each micro-zone, extracting key characteristics of the order set, including order density, time urgency, and package size distribution. Based on these characteristics, the system calculated matching scores for different resource micro-zone combinations using a multi-dimensional scoring mechanism encompassing load matching, timeliness, energy efficiency, and traffic adaptability.

[0078] The system's resource combination generation module, based on a genetic algorithm framework, employs a specialized encoding scheme to represent heterogeneous resource combinations, searching for the global optimal solution through iterative optimization. A key innovation of the algorithm is the introduction of a resource complementarity assessment mechanism. When combining different types of distribution resources, the complementary gains are calculated. For example, Table 5 shows the complementarity scores for some resource combinations calculated by the system: Table 5: Distribution resource combination complementarity scoring table; The system applies a combinatorial optimization algorithm to determine the optimal heterogeneous resource combination for each micro-zone based on its characteristics and order requirements.

[0079] A notable feature of the system is its dynamic adjustment capability. It reassesses resource allocation plans hourly, triggering resource reorganization when necessary based on real-time order changes and traffic conditions. For example, upon detecting a sudden surge in urgent orders in a certain business district in Wuhou District, the system immediately dispatched two drones from a nearby micro-district, shortening delivery times.

[0080] Actual application data shows that compared with the fixed vehicle type delivery model, this dynamic grouping solution for heterogeneous resources has increased the delivery success rate by 15.3%, reduced energy consumption per unit order by 28.6%, and increased customer satisfaction by 23.1%, especially in the core areas of cities where traffic is complex and changeable.

[0081] Implementation process of multi-agent reinforcement learning algorithm: Based on the results of heterogeneous delivery resource grouping, the system applies a multi-agent reinforcement learning algorithm to further optimize the collaborative behavior within the formation, achieving adaptive formation and separation of vehicles. The system treats each delivery tool as an agent and optimizes the overall delivery task through collaborative decision-making.

[0082] The system builds a multi-agent reinforcement learning model with a hierarchical actor-critic architecture. The model consists of two layers: formation-level strategies and individual-level strategies. The formation-level strategy is responsible for macro-coordination, while the individual-level strategy is responsible for micro-execution. The system defines the state space, action space, and observation space for each type of delivery vehicle. Table 6 shows some parameter definitions for the agents in the system: Table 6: Parameter definition of delivery agent; The system designs an agent collaboration model based on maximizing mutual information. By maximizing the mutual information between an agent's state and the behavior of other agents, it encourages agents to learn strategies that consider each other's behavior and enhances their collaborative capabilities. The model also incorporates an attention communication mechanism to establish a dynamic connection structure between agents, fostering stronger information exchange channels between highly related agents.

[0083] The system constructs a reward function based on expected time benefits and introduces a combination of global rewards and local rewards. The global reward focuses on the completion efficiency of the entire delivery task, while the local reward focuses on the execution quality of a single agent.

[0084] In practical applications, the algorithm has demonstrated remarkable adaptability. For example, when handling a sudden surge in orders in Chengdu's Shuangliu District (due to a promotion on an e-commerce platform), the system automatically identified changes in order patterns and dynamically adjusted delivery strategies. It reorganized previously dispersed small vehicles into a more efficient mixed fleet, increased vehicle handover points, and reduced duplicate routes. Ultimately, despite a 35% increase in order volume, delivery time increased by only 15%, while other logistics companies not using the system saw delivery times increase by over 60% during the same period.

[0085] Technical effect verification: To validate the effectiveness of this implementation, the company conducted a six-month system application test in Chengdu. The test was divided into two phases: the first three months used traditional logistics delivery methods (fixed vehicle types and fixed routes), and the second three months fully implemented the technical system proposed in this solution. The order volume, delivery areas, and staffing structure were essentially the same across the two phases, making them comparable. Table 7 shows a comparison of key performance indicators across the two phases: Table 7: Comparison of performance indicators before and after technology implementation; Table 7 shows that this solution achieved significant improvements across all key metrics. In particular, order delays in peak-hour congested areas decreased by 67.0%, primarily due to the synergy between the congestion propagation prediction model and the two-layer intelligent decision-making system, which enabled the fleet to avoid congested areas in advance. Average vehicle utilization increased by 35.8%, primarily due to the dynamic grouping of heterogeneous delivery resources and the optimization effects of the multi-agent reinforcement learning algorithm, which enabled more rational allocation and utilization of different types of delivery resources.

[0086] The system also demonstrates excellent adaptability and robustness. It exhibits strong anti-interference capabilities in response to emergencies such as traffic control, severe weather, sudden surges in order volume, road construction, and major events. Delivery delays are significantly reduced compared to traditional methods, by 60%-75%. This is primarily due to the system's predictive capabilities and adaptive adjustment mechanisms.

[0087] In summary, this implementation method has built an efficient, flexible and intelligent urban logistics distribution system through the organic combination of core technologies such as congestion propagation prediction, spatiotemporal dynamic partitioning, two-layer intelligent decision-making, heterogeneous resource grouping and multi-agent reinforcement learning. It has achieved remarkable technical results in practical applications and provided a new solution for urban logistics distribution.

[0088] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A big data analysis method based on transportation logistics, characterized in that: The following steps are involved: Construct a congestion propagation prediction model based on spatiotemporal graph convolutional networks, analyze historical traffic data, and predict future congestion trends in urban areas; Implement a spatiotemporal dynamic partitioning algorithm to divide the urban delivery area into multiple dynamic micro-zones based on congestion prediction results and delivery order distribution; Build a two-layer intelligent decision-making system to generate logistics fleet splitting and joining decisions and route planning based on congestion prediction and micro-area division; Dynamically group heterogeneous distribution resources, combining different types of distribution tools to form an optimal micro-fleet based on micro-area characteristics and order characteristics; Based on the formation results, a multi-agent reinforcement learning algorithm is applied to optimize the formation collaboration and realize adaptive formation and separation of vehicles.

2. A big data analysis method based on transportation logistics according to claim 1, characterized in that: The spatiotemporal dynamic partitioning algorithm includes: Collect and process delivery order data and available logistics resource data to form order density matrix and resource availability matrix; Construct a spatiotemporal partitioning evaluation function that integrates congestion prediction results, order density, and resource availability; The improved spatiotemporal DBSCAN algorithm is applied to cluster the evaluation functions and the spatiotemporal distance metric is defined; Based on the clustering results, the city's dynamic micro-division is determined and the final dynamic micro-division results are output.

3. The big data analysis method based on transportation logistics according to claim 1 is characterized in that: The two-layer intelligent decision-making system includes: A high-level decision-making system is used to determine the set of key nodes suitable for fleet decomposition based on the congestion prediction heat map and micro-division results, and generate a fleet decomposition strategy; A low-level decision-making system is used to generate detailed delivery routes for each sub-fleet and corresponding delivery micro-area; The decomposition decision results of the high-level system are passed to the low-level system as constraints for path planning, while the path execution effect and real-time feedback of the low-level system are passed back to the high-level system for dynamic adjustment of subsequent decomposition decisions.

4. The method for analyzing big data based on transportation logistics according to claim 1, characterized in that: The dynamic grouping of heterogeneous distribution resources includes: Establish a distribution resource capability model to describe the performance characteristics of different types of distribution tools; Construct a micro-area order feature analysis model to extract the key features of the order set; Construct resource micro-area matching algorithm and calculate matching scores; Combinatorial optimization algorithms are applied to determine the optimal heterogeneous resource combination for each micro-area.

5. The big data analysis method based on transportation logistics according to claim 1 is characterized in that: The multi-agent reinforcement learning algorithm includes: Build a vehicle agent model and define the state space, action space, and observation space; Build an agent collaboration model based on maximizing mutual information to enhance information sharing and decision-making collaboration among agents; Construct a reward function based on expected time benefits to guide the agent to learn the optimal strategy; A hierarchical multi-agent reinforcement learning algorithm is applied to train the team collaboration strategy, adopting the Actor-Critic architecture.

6. A method for analyzing big data based on transportation logistics according to claim 5, characterized in that: The mutual information optimization objective in the agent collaboration model based on mutual information maximization is defined as: ； in, Indicates that in a known agent Status Under these conditions, the agent Status With the agent Action The conditional mutual information between and Represents the intelligent agent and agents status, Representing an agent action, p represents the probability, Represents a logarithmic function, which measures the agent The state of the agent The degree of influence on action decision.

7. The method for analyzing big data based on transportation logistics according to claim 5, characterized in that: The reward function based on expected time benefit is defined as: ； in, Indicates that the status Next action The reward value obtained, and Represent the joint action and state at time t, For the The weight of an order, is the baseline delivery time, For the current action and state The expected completion time of each order, Indicates the total number of orders waiting to be delivered in the current system.

8. The method for analyzing big data based on transportation logistics according to claim 2, characterized in that: The spatiotemporal partition evaluation function is defined as: ； in, Indicates time and location The time and space partition evaluation function value of Represents the position coordinates, Indicates a point in time, Indicates location In time The order density function of Indicates location In time Traffic condition function, Indicates time Resource availability function.

9. The method for analyzing big data based on transportation logistics according to claim 3, characterized in that: The decision function in the high-level decision system is defined as: ； in, Indicates that the node and time The fleet decomposition decision function, Represents a road network node, Indicates a point in time, Representation node In time The congestion probability, Representation node In time The micro-region to which it belongs, Indicates the current fleet resource status, Represents the decision mapping function.

10. A big data analysis system based on transportation logistics, used to execute a big data analysis method based on transportation logistics according to any one of claims 1 to 9, characterized in that: include: The congestion propagation prediction module is used to build a congestion propagation prediction model based on spatiotemporal graph convolutional networks, analyze historical traffic data, and predict future congestion trends in urban areas; The spatiotemporal dynamic partitioning module is used to implement the spatiotemporal dynamic partitioning algorithm, which divides the urban delivery area into multiple dynamic micro-zones based on congestion prediction results and delivery order distribution; A two-layer intelligent decision-making module is used to generate logistics fleet splitting and joining decisions and route planning based on congestion prediction and micro-area division; Heterogeneous resource grouping module, used to dynamically group heterogeneous distribution resources. Based on micro-area characteristics and order characteristics, different types of distribution tools are combined to form an optimal micro-formation. The multi-agent collaboration module is used to apply multi-agent reinforcement learning algorithms to optimize platoon collaboration and achieve adaptive vehicle platooning and separation.

Images