Artificial intelligence-based logistics transportation dynamic scheduling method and system

Abstract

This application discloses an artificial intelligence-based dynamic scheduling method and system for logistics transportation, relating to the field of artificial intelligence technology. It addresses the technical problem that existing technologies often rely on fixed rules based on experience or simple optimization algorithms, which become rigid in complex and dynamic real-world environments, unable to make forward-looking and global trade-offs, leading to low efficiency in transportation scheduling methods. The method obtains a dynamic scheduling knowledge graph by processing and integrating multimodal transportation data; optimizes the dynamic scheduling knowledge graph to obtain an optimized scheduling instruction set, and constructs a dynamic scheduling knowledge graph that transforms fragmented static information into a semantically related global dynamic relationship network, providing a panoramic and structured situational awareness foundation for decision-making. Based on this knowledge graph, rolling optimization scheduling is performed using an artificial intelligence model, enabling the learning of complex relationships, prediction of dynamic changes, and global trade-offs and adaptive decision-making, thereby improving the efficiency of transportation scheduling methods.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically an artificial intelligence-based dynamic scheduling method and system for logistics transportation. Background Technology

[0002] Dynamic scheduling of logistics transportation is the core decision-making mechanism of modern intelligent logistics systems in response to complex and ever-changing environments. Its essence lies in breaking through the rigid constraints of traditional static planning and utilizing IoT, big data, and AI technologies to instantly perceive and respond to uncertainties arising in real-time during transportation, such as traffic congestion, emergency order insertions, vehicle breakdowns, and demand fluctuations. It is no longer limited to pre-planned routes but dynamically reconstructs vehicle routes, loading schemes, and delivery sequences through millisecond-level data calculations, achieving globally optimal resource allocation.

[0003] Existing technologies often rely on fixed rules based on experience or simple optimization algorithms, which are rigid in complex and dynamic real-world environments and cannot make forward-looking and global trade-offs, resulting in low efficiency of transportation scheduling methods. Therefore, further improvements are still needed for dynamic scheduling methods of logistics transportation. Summary of the Invention

[0004] This application aims to solve at least one of the technical problems existing in the prior art; to this end, this application proposes a dynamic scheduling method and system for logistics transportation based on artificial intelligence, which is used to solve the technical problem that the prior art often relies on fixed rules based on experience or simple optimization algorithms, which are rigid in complex and dynamic real-world environments and cannot make forward-looking and global trade-offs, resulting in low efficiency of transportation scheduling methods.

[0005] To achieve the above objectives, the first aspect of this application provides a dynamic scheduling method for logistics transportation based on artificial intelligence, comprising: Acquire multimodal transportation data; the multimodal transportation data includes transportation vehicle data, environmental data, order data, and infrastructure data; Data processing and integration of multimodal transportation data yield a dynamic scheduling knowledge graph; The dynamic scheduling knowledge graph is optimized to obtain an optimized scheduling instruction set.

[0006] This application, through the aforementioned steps, deeply integrates multimodal real-time data from transportation vehicles, the environment, orders, and infrastructure to construct a unified dynamic scheduling knowledge graph. This reshapes previously fragmented static information into a globally dynamic relationship network with semantic connections, thus laying a panoramic and structured foundation for decision-making. Based on this, relying on artificial intelligence models to implement rolling optimization scheduling, it can gain real-time insight into complex relationships, accurately predict dynamic evolution, and execute global trade-offs and adaptive decisions. It abandons the rigid mode of relying on fixed rules or simple algorithms, and promotes a fundamental leap in scheduling decision-making from experience-driven and locally static to data-intelligent driven and globally dynamic, improving the overall efficiency, response agility, and resource utilization of transportation scheduling in complex and ever-changing environments.

[0007] Furthermore, the process of processing and integrating multimodal transportation data to obtain a dynamic scheduling knowledge graph includes: Acquire multimodal transportation data; the multimodal transportation data includes vehicle data, environmental data, order data, and infrastructure data; the vehicle data includes vehicle ID, real-time GPS location, speed, direction of travel, remaining mileage, planned stop, loaded cargo list, and destination; the environmental data includes road segment ID, real-time traffic conditions, and short-term weather forecasts; the order data includes information on newly added, modified, canceled, and expedited orders, and the order information includes order ID, pickup point, delivery point, time window, cargo attributes, and priority; the infrastructure data includes warehouse ID and its corresponding real-time operational load, number of available loading / unloading platforms, and estimated waiting time; Multimodal transportation data is cleaned and standardized to obtain several clean data streams; the data cleaning and standardization operations include noise reduction and repair, formatting and structuring operations; A set of association relationships is obtained by performing spatiotemporal alignment and entity association based on several clean data streams; Define an initial knowledge graph; the initial knowledge graph is created using a graph database approach and includes several entity nodes and relation edges; The set of relationships is converted into MERGE and SET operations on the initial knowledge graph to obtain a dynamically scheduled knowledge graph; the MERGE operation refers to merging nodes and relationships, and the SET operation refers to updating attributes.

[0008] Furthermore, the set of association relationships obtained by performing spatiotemporal alignment and entity association based on several clean data streams includes: Define a sliding time window; the size of the sliding time window is set based on experience. An entity association and binding calculation is performed on several clean data streams within a sliding time window to obtain an association set; the association set includes several entities and the relationships between entities; all data in the association set has a unified timestamp; The entity association and binding calculation includes vehicle-road segment binding, vehicle-order binding, road segment-environment binding, and order-warehouse binding; The vehicle-to-road-segment binding refers to binding a vehicle ID to a road segment ID and calculating the estimated progress of the vehicle ID on the road segment ID. The vehicle-order binding refers to binding the vehicle ID to the order ID; The road segment and environment binding refers to binding the road segment ID with its corresponding real-time traffic conditions and future short-term weather forecasts; The order-warehouse binding refers to parsing the pickup and delivery points corresponding to each order ID and binding them to the warehouse ID.

[0009] This application provides high-quality input for subsequent processing by meticulously defining the composition of multi-source data and cleaning and standardizing it. Then, it utilizes a sliding time window to perform precise spatiotemporal alignment and dynamic binding calculations on various entities such as vehicles, road segments, orders, and the environment, establishing real-time and accurate relationships between them. Finally, by continuously updating these relationships into a knowledge graph based on a graph database, it constructs and maintains a dynamic scheduling knowledge graph that comprehensively and structurally reflects the real-time status and evolution of logistics. This breaks down barriers between various types of data, forming a unified, computable, and millisecond-updated digital twin ecosystem. It provides a precise, reliable, and semantically rich global information foundation for upper-level intelligent decision-making, fundamentally solving the problem of low-quality scheduling decisions caused by information fragmentation and lag.

[0010] Furthermore, the optimized scheduling instruction set obtained by optimizing the dynamic scheduling knowledge graph includes: Extract dynamic scheduling knowledge graphs corresponding to several historical timestamps; A predicted graph is obtained by inputting a dynamic scheduling knowledge graph corresponding to several historical timestamps into a graph prediction model; the graph prediction model is constructed using a spatiotemporal graph neural network model; the spatiotemporal graph neural network model includes a spatiotemporal embedding module, a temporal convolution module, and a spatiotemporal fusion prediction module; The dynamic scheduling knowledge graph and prediction graph are input into the instruction optimization model to obtain the optimized scheduling instruction set; the instruction optimization model is constructed through a hierarchical attention decision network; the hierarchical attention decision network includes a scheduling policy generator and a distributed action executor.

[0011] Furthermore, the spatiotemporal embedding module includes: The spatiotemporal embedding module obtains the spatial feature vector of each entity node in the dynamically scheduled knowledge graph at several historical timestamps by using a spatial attention encoder. The spatial attention encoder is used to calculate the attention coefficients corresponding to several entity nodes and aggregate their information to obtain the spatial feature vectors of the entity nodes on the timestamp. The attention coefficient satisfies: Where v represents the entity node whose attention coefficient needs to be calculated, u represents the neighboring entity nodes of the entity node; R represents the edge relationship between entity nodes; and h represents the feature vector corresponding to the entity node. This is represented as concatenating the feature vectors of two entity nodes; It is represented as a learnable attention vector; T represents the matrix transpose operation; Used to map the concatenated high-dimensional feature vector to a scalar attention coefficient; It is represented as a learnable weight matrix related to edge relationships. Represented as attention coefficient; This is represented as an activation function, which is an improved activation function based on the standard ReLU function.

[0012] Furthermore, the information aggregation satisfies: ;in, Represented as attention weights, derived from attention coefficients. Obtained by normalization using the softmax function; This is represented as the set of neighboring entity nodes corresponding to entity node v; Represented as the spatial feature vector corresponding to entity node v; It is represented as a non-linear activation function.

[0013] Furthermore, the temporal convolution module and the spatiotemporal fusion prediction module include: The input data of the temporal convolution module is the spatial feature vector corresponding to several historical timestamps of the entity node; the output data is the spatiotemporal fusion feature vector of the entity node at the current timestamp; the spatiotemporal fusion feature vector fuses the spatial feature vectors of several historical timestamps before the current timestamp; the temporal convolution module uses a one-dimensional causal convolutional network to perform temporal feature fusion. The spatiotemporal fusion prediction module consists of several prediction heads. Its input data is the spatiotemporal fusion feature vector output by several entity nodes through the temporal convolution module, and its output data is the future prediction result corresponding to several entity nodes. The prediction head includes a core fully connected layer and a task-specific output layer. The core fully connected layer is used to map a general spatiotemporal fusion feature vector to a feature space that matches the task-specific output layer. The task-specific output layer is used to predict the future prediction results from the output data of the core fully connected layer. The predicted future results corresponding to several entity nodes are integrated to obtain a prediction map.

[0014] Furthermore, the scheduling policy generator includes: The input data of the scheduling policy generator are a dynamic scheduling knowledge graph and a prediction graph, and the output data is a meta-policy vector; the scheduling policy generator includes global graph encoding operation and global pooling and policy generation operation. The global graph encoding operation includes: The fused graph is input into the graph attention encoder to obtain global context feature vectors corresponding to several entity nodes; the fused graph is obtained by adding the future prediction results corresponding to several entity nodes in the prediction graph as attributes of the entity nodes to the dynamic scheduling knowledge graph. The global pooling and policy generation operations include: Attention pooling is performed on the global context feature vectors corresponding to all entity nodes to obtain a global state vector. The global state vector is then mapped to a meta-policy vector through a policy generation network.

[0015] Furthermore, the distributed action executor includes: The input data of the distributed action executor includes vehicle ID and its corresponding attributes, local observation graph, and meta-policy vector; the local observation graph refers to the subgraph structure collected within the K-hop neighbor entity nodes centered on the vehicle ID as the local observation graph corresponding to the vehicle ID; the output data is an optimized scheduling instruction set; where K is an integer, K>0; The working steps of the distributed action executor include: The local observation map is input into the graph neural network to obtain the local structured feature vector; The decision feature vector corresponding to the vehicle ID is obtained by concatenating the attributes, local structured feature vector, and meta-policy vector corresponding to the vehicle ID. The decision feature vector corresponding to the vehicle ID is input into the action evaluation network to obtain the optimized scheduling instruction corresponding to the vehicle ID. The core of the action evaluation network employs an attention mechanism; the workflow of the attention mechanism includes: A set of candidate actions is generated for the vehicle ID, and the set of candidate actions consists of several candidate actions; Each candidate action is encoded as a feature vector as an action embedding; the attention score corresponding to the candidate action is calculated using the decision feature vector as the query and the action embedding as the key. The attention scores of several candidate actions are normalized using a softmax function to obtain the selection probability of each candidate action. Select the candidate action with the highest probability from the vehicle IDs and use it as the optimized scheduling instruction for the vehicle ID. The optimized scheduling instruction set is determined based on the optimized scheduling instructions corresponding to several vehicle IDs.

[0016] This application first integrates real-time and predictive knowledge graphs through an upper-layer scheduling policy generator, encodes the global situation using a graph attention network, and extracts the core state through attention pooling, ultimately generating a meta-policy vector to guide the global scheduling direction. At the lower layer, each distributed action executor focuses on a single vehicle, analyzes its local environment graph through a graph neural network, and combines vehicle attributes, local features, and the meta-policy vector issued from the upper layer. Through an attention-based action evaluation network, the optimal localized scheduling instruction is precisely selected from the candidate action set. This cleverly combines global optimization with rapid local response, ensuring the consistency of the overall goal through meta-policy while giving each executor the flexibility to make refined decisions based on real-time local information. Thus, highly collaborative, adaptive, and scalable efficient scheduling is achieved in complex and dynamic logistics scenarios.

[0017] Another aspect of the present invention provides a dynamic scheduling system for logistics transportation based on artificial intelligence, comprising: a data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected together; The data acquisition module acquires multimodal transportation data through data acquisition equipment; the multimodal transportation data includes transportation vehicle data, environmental data, order data, and infrastructure data. The data analysis module includes a map construction unit and a result generation unit; The graph construction unit processes and integrates multimodal transportation data to obtain a dynamic scheduling knowledge graph. The result generation unit performs scheduling optimization on the dynamic scheduling knowledge graph to obtain an optimized scheduling instruction set.

[0018] Compared with the prior art, the beneficial effects of this application are: 1. This application obtains a dynamic scheduling knowledge graph by processing and integrating multimodal transportation data; optimizes the dynamic scheduling knowledge graph to obtain an optimized scheduling instruction set; acquires and integrates real-time multimodal data such as transportation vehicles, environment, orders, and infrastructure; constructs a unified dynamic scheduling knowledge graph; transforms fragmented static information into a global dynamic relationship network with semantic associations; provides a panoramic and structured situational awareness foundation for decision-making; based on this knowledge graph, rolling optimization scheduling is performed using artificial intelligence models; it can learn complex relationships in real time, predict dynamic changes, and make global trade-offs and adaptive decisions; replacing the rigid mode that relies on fixed rules or simple algorithms; realizing a fundamental transformation of scheduling decision-making from experience-driven and locally static to data-intelligent driven and globally dynamic; and improving the overall efficiency, response speed, and resource utilization of transportation scheduling in complex and ever-changing environments.

[0019] 2. This application utilizes a spatiotemporal graph neural network to perform in-depth analysis of historical knowledge graphs. Its spatiotemporal embedding module employs an improved attention mechanism to accurately model the dynamic spatial dependencies between entities, while the temporal convolution module captures the temporal evolution patterns of each entity's state, thereby accurately predicting key situations such as future road conditions and node loads. Furthermore, the real-time and predicted graphs are input into a hierarchical attention decision network. The upper-layer policy generator of this network integrates global information to formulate macro-level scheduling guidelines, while the lower-layer distributed executor generates specific instructions based on these guidelines and local observations, achieving a closed loop from prediction to decision-making. This enables scheduling to possess forward-looking cognition and adaptive collaborative decision-making capabilities for the first time. Based on the prediction of future states, it can perform dynamic, global, and multi-objective optimization, fundamentally overcoming the inherent defects of traditional rule-based systems, such as delayed response and limited optimization perspective in complex dynamic environments, and greatly improving scheduling efficiency and intelligence.

[0020] 3. This application first integrates real-time and predictive knowledge graphs through an upper-layer scheduling strategy generator, encodes the global situation using a graph attention network, and extracts the core state through attention pooling, ultimately generating a meta-policy vector that guides the global scheduling direction. Each distributed action executor at the lower layer takes a single vehicle as its center, analyzes its local environment graph through a graph neural network, and combines vehicle attributes, local features, and the meta-policy vector issued by the upper layer. Through an action evaluation network based on an attention mechanism, the optimal localized scheduling instruction is accurately selected from the candidate action set. This cleverly combines global optimization with rapid local response, ensuring the consistency of the overall goal through meta-policy while giving each executor the flexibility to make fine decisions based on real-time local information. Thus, highly collaborative, adaptive, and scalable efficient scheduling is achieved in complex and dynamic logistics scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the AI-based dynamic scheduling method for logistics transportation according to this application. Figure 2 This is a schematic diagram illustrating the principle of the AI-based dynamic scheduling system for logistics transportation in this application. Detailed Implementation

[0023] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0024] Please see Figure 1 The first aspect of this application provides a dynamic scheduling method for logistics transportation based on artificial intelligence, including: Acquire multimodal transportation data; multimodal transportation data includes vehicle data, environmental data, order data, and infrastructure data; Data processing and integration of multimodal transportation data yield a dynamic scheduling knowledge graph; The dynamic scheduling knowledge graph is optimized to obtain an optimized scheduling instruction set.

[0025] In this embodiment, data processing and integration of multimodal transportation data to obtain a dynamic scheduling knowledge graph includes: Acquire multimodal transportation data; multimodal transportation data includes vehicle data, environmental data, order data, and infrastructure data; vehicle data includes vehicle ID, real-time GPS location, speed, direction of travel, remaining mileage, planned stop, loaded cargo list, and destination; environmental data includes road segment ID, real-time traffic conditions, and short-term weather forecasts; in this embodiment, real-time traffic conditions include smooth traffic, congestion, accidents, and road closures, and short-term weather forecasts include sunny, rainy, snowy, and foggy conditions; order data includes information on newly added, modified, canceled, and expedited orders, and order information includes order ID, pickup point, delivery point, time window, cargo attributes, and priority; infrastructure data includes warehouse ID and its corresponding real-time operational load, number of available loading / unloading platforms, and estimated waiting time; Data cleaning and standardization operations are performed on multimodal transportation data to obtain several clean data streams; data cleaning and standardization operations include noise reduction and repair, formatting and structuring operations; A set of association relationships is obtained by performing spatiotemporal alignment and entity association based on several clean data streams; An initial knowledge graph is defined. This initial knowledge graph is created using a graph database approach and includes several entity nodes and relational edges. In this embodiment, entity nodes include vehicles, orders, road segments, and warehouses. Relationship edges represent the associations between these entity nodes. The graph database used in this embodiment is Neo4j. Both entity nodes and relational edges in this embodiment have several attributes. For example, vehicle entity nodes have attributes such as vehicle ID, vehicle status, and vehicle capacity; order entity nodes have attributes such as order ID and priority; road segment entity nodes have attributes such as road segment ID, road segment length, and road segment speed limit; there is a relationship between vehicle entity nodes and road segment entity nodes regarding the location of a vehicle on a road segment, with attributes such as estimated progress and speed; and there is a relationship between vehicle entity nodes and order entity nodes regarding vehicle transport orders, with attributes such as departure time. The set of relationships is transformed into a dynamically scheduled knowledge graph by performing MERGE and SET operations on the initial knowledge graph; the MERGE operation refers to merging nodes and relationships, and the SET operation refers to updating attributes.

[0026] In this embodiment, a set of association relationships is obtained by performing spatiotemporal alignment and entity association based on several clean data streams, including: Define a sliding time window; the size of the sliding time window is set based on experience, and in this embodiment, the size of the sliding time window is set to 10 seconds. Entity association and binding calculations are performed on several clean data streams within a sliding time window to obtain an association set; the association set includes several entities and the relationships between them; all data within the association set has a unified timestamp; Entity association and binding calculations include vehicle-road segment binding, vehicle-order binding, road segment-environment binding, and order-warehouse binding; Vehicle-to-road segment binding refers to binding a vehicle ID to a road segment ID and calculating the estimated progress of the vehicle ID on the road segment ID. In this embodiment, the real-time GPS location corresponding to the vehicle ID is used to determine which road segment ID the vehicle ID is matched with. Vehicle-order binding refers to binding a vehicle ID to an order ID. In this embodiment, when an order ID is in transit status, it is strongly associated with a vehicle ID. Linking road segments to the environment refers to binding road segment IDs with their corresponding real-time traffic conditions and short-term weather forecasts. Binding orders to warehouses means parsing the pickup and delivery points corresponding to each order ID and binding them to the warehouse ID.

[0027] This embodiment first lays a high-quality data input foundation by meticulously defining the composition of multi-source data and implementing rigorous cleaning and standardization. Then, a sliding time window mechanism is introduced to accurately align and dynamically bind multi-dimensional entities such as vehicles, road segments, orders, and the environment in time and space, constructing a real-time and precise relational network. Finally, these dynamic relationships are continuously injected into a knowledge graph based on a graph database, successfully constructing and maintaining a dynamic scheduling knowledge graph that can panoramically and structurally map the real-time status and evolution of logistics. This graph completely breaks down data silos, forming a unified, computable, and millisecond-level updated digital twin ecosystem. It provides a precise, reliable, and semantically rich global information foundation for upper-level intelligent decision-making, fundamentally overcoming the problem of low scheduling decision-making efficiency caused by information fragmentation and lag.

[0028] In this embodiment, the optimized scheduling instruction set is obtained by optimizing the dynamic scheduling knowledge graph, including: Extract dynamic scheduling knowledge graphs corresponding to several historical timestamps; The predicted graph is obtained by inputting a dynamic scheduling knowledge graph corresponding to several historical timestamps into the graph prediction model; the graph prediction model is constructed using a spatiotemporal graph neural network model; the spatiotemporal graph neural network model includes a spatiotemporal embedding module, a temporal convolution module, and a spatiotemporal fusion prediction module; The dynamic scheduling knowledge graph and prediction graph are input into the instruction optimization model to obtain the optimized scheduling instruction set; the instruction optimization model is constructed through a hierarchical attention decision network; the hierarchical attention decision network includes a scheduling policy generator and a distributed action executor.

[0029] The spatiotemporal embedding module in this embodiment includes: The spatiotemporal embedding module obtains the spatial feature vector of each entity node in the dynamically scheduled knowledge graph at several historical timestamps by using a spatial attention encoder. The spatial attention encoder is used to calculate the attention coefficients corresponding to several entity nodes and aggregate their information to obtain the spatial feature vectors of the entity nodes on the timestamp. Attention coefficient satisfies: Where v represents the entity node whose attention coefficient needs to be calculated, u represents the neighboring entity nodes of the entity node; R represents the edge relationship between entity nodes; and h represents the feature vector corresponding to the entity node. This is represented as concatenating the feature vectors of two entity nodes; It is represented as a learnable attention vector; T represents the matrix transpose operation; Used to map the concatenated high-dimensional feature vector to a scalar attention coefficient; It is represented as a learnable weight matrix related to edge relationships. Represented as attention coefficient; This is represented as an activation function, which is an improved activation function based on the standard ReLU function. In this embodiment, when calculating the attention coefficients corresponding to entity nodes, the attention coefficients corresponding to several entity nodes are calculated by introducing the relationship edges between entity nodes.

[0030] The information aggregation in this embodiment satisfies: ;in, Represented as attention weights, derived from attention coefficients. Obtained by normalization using the softmax function; This is represented as the set of neighboring entity nodes corresponding to entity node v; Represented as the spatial feature vector corresponding to entity node v; It is represented as a non-linear activation function.

[0031] The temporal convolution module and spatiotemporal fusion prediction module in this embodiment include: The input data of the temporal convolution module is the spatial feature vector corresponding to several historical timestamps of the entity node; the output data is the spatiotemporal fusion feature vector of the entity node at the current timestamp; the spatiotemporal fusion feature vector fuses the spatial feature vectors of several historical timestamps before the current timestamp; the temporal convolution module uses a one-dimensional causal convolutional network to fuse temporal features. The spatiotemporal fusion prediction module consists of several prediction heads. Its input data is the spatiotemporal fusion feature vectors output by several entity nodes through the temporal convolution module, and its output data is the future prediction results corresponding to several entity nodes. The prediction head includes a core fully connected layer and a task-specific output layer. The core fully connected layer is used to map a general spatiotemporal fusion feature vector to a feature space that matches the task-specific output layer. The task-specific output layer is used to predict the future prediction results from the output data of the core fully connected layer. In this embodiment, the prediction head includes a road segment speed prediction head and a node delay risk prediction head, etc. The prediction results corresponding to several entity nodes are integrated to obtain a prediction map.

[0032] This embodiment utilizes a spatiotemporal graph neural network to deeply mine historical knowledge graphs. Its spatiotemporal embedding module, with the help of an improved attention mechanism, accurately depicts the dynamic spatial dependencies between entities. The temporal convolution module keenly captures the temporal evolution of the states of each entity, thereby achieving accurate prediction of key situations such as future road conditions and node loads. Based on this, the real-time graph and the predicted graph are jointly input into a hierarchical attention decision network. The upper-layer policy generator coordinates global information to formulate macro-level scheduling guidelines, while the lower-layer distributed executor generates specific instructions based on these guidelines and local observations. This constructs a complete closed loop from situation prediction to intelligent decision-making, enabling scheduling to have forward-looking cognition and adaptive collaborative capabilities. By performing dynamic, global, and multi-objective trade-off optimization based on future state prediction, it fundamentally overcomes the inherent defects of traditional rules in complex dynamic environments, such as delayed response and limited vision, and significantly improves scheduling efficiency and intelligence.

[0033] The scheduling policy generator in this embodiment includes: The input data for the scheduling policy generator are a dynamic scheduling knowledge graph and a prediction graph, and the output data is a meta-policy vector. The scheduling policy generator includes global graph encoding operations and global pooling and policy generation operations. Global graph encoding operations include: The fused graph is input into the graph attention encoder to obtain global context feature vectors corresponding to several entity nodes; the fused graph is obtained by adding the future prediction results corresponding to several entity nodes in the prediction graph as attributes of the entity nodes to the dynamically scheduled knowledge graph. Global pooling and policy generation operations include: The global context feature vectors corresponding to all entity nodes are attention-pooled to obtain the global state vector. The global state vector is then mapped to a meta-policy vector through a policy generation network. In this embodiment, the policy generation network adopts a multilayer perceptron (MLP). In this embodiment, the dimension of the meta-policy vector is pre-defined, and each dimension corresponds to an interpretable macroscopic instruction direction. The meta-policy vector indicates the key direction for global optimization in the current stage.

[0034] The distributed action executor in this embodiment includes: The input data of the distributed action executor includes vehicle ID and its corresponding attributes, local observation graph, and meta-policy vector; the local observation graph refers to the subgraph structure collected within the K-hop neighbor entity nodes centered on the vehicle ID as the local observation graph corresponding to the vehicle ID; the output data is an optimized scheduling instruction set; where K is an integer, K>0; the specific value is set according to experience, and in this embodiment it is set to 2; The working steps of a distributed action executor include: The local observation map is input into the graph neural network to obtain the local structured feature vector; The decision feature vector corresponding to the vehicle ID is obtained by concatenating the attributes, local structured feature vector, and meta-policy vector corresponding to the vehicle ID. The decision feature vector corresponding to the vehicle ID is input into the action evaluation network to obtain the optimized scheduling instruction corresponding to the vehicle ID; in this embodiment, the action evaluation network adopts MLP. The core of the action evaluation network employs an attention mechanism; the workflow of the attention mechanism includes: Generate a set of candidate actions for the vehicle ID. The set of candidate actions consists of several candidate actions. Each candidate action is encoded as a feature vector as an action embedding; the attention score corresponding to the candidate action is calculated using the decision feature vector as the query and the action embedding as the key. The attention scores of several candidate actions are normalized using a softmax function to obtain the selection probability of each candidate action. The candidate action with the highest probability is selected from the vehicle IDs and used as the optimized scheduling instruction for the vehicle IDs. In this embodiment, the higher the attention score, the higher the fit between the candidate action and the attributes, local environment, and global meta-policy corresponding to the vehicle ID. The optimized scheduling instruction set is determined based on the optimized scheduling instructions corresponding to several vehicle IDs.

[0035] Please see Figure 2 Another embodiment of this application provides an artificial intelligence-based dynamic scheduling system for logistics transportation, including: a data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected. Data acquisition module: Acquires multimodal transportation data through data acquisition equipment; multimodal transportation data includes transportation vehicle data, environmental data, order data, and infrastructure data; the data acquisition equipment in this embodiment includes several sensors, etc. The data analysis module includes a map construction unit and a result generation unit; Knowledge graph construction unit: Data processing and integration of multimodal transportation data to obtain a dynamic scheduling knowledge graph; Result generation unit: Perform scheduling optimization on the dynamic scheduling knowledge graph to obtain an optimized scheduling instruction set.

[0036] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0037] The working principle of this application is as follows: Multimodal transportation data is acquired; data processing and integration of the multimodal transportation data yield a dynamic scheduling knowledge graph; scheduling optimization of the dynamic scheduling knowledge graph yields an optimized scheduling instruction set; real-time multimodal data from transportation vehicles, environment, orders, and infrastructure are acquired and integrated to construct a unified dynamic scheduling knowledge graph, transforming fragmented static information into a globally dynamic relationship network with semantic connections, providing a panoramic and structured situational awareness foundation for decision-making; based on this knowledge graph, rolling optimization scheduling is performed using an artificial intelligence model, enabling real-time learning of complex relationships, prediction of dynamic changes, and global trade-offs and adaptive decisions. This replaces the rigid mode relying on fixed rules or simple algorithms, achieving a fundamental shift in scheduling decision-making from experience-driven and locally static to data-intelligent driven and globally dynamic. This improves the overall efficiency, response speed, and resource utilization of transportation scheduling in complex and ever-changing environments, avoiding the problem that existing technologies often rely on experience-based fixed rules or simple optimization algorithms, which are rigid in complex and dynamic real-world environments and unable to make forward-looking and global trade-offs, resulting in low efficiency in transportation scheduling methods.

[0038] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A dynamic scheduling method for logistics transportation based on artificial intelligence, characterized in that, include: Acquire multimodal transportation data; The multimodal transportation data includes transportation vehicle data, environmental data, order data, and infrastructure data; Data processing and integration of multimodal transportation data yield a dynamic scheduling knowledge graph; The dynamic scheduling knowledge graph is optimized to obtain an optimized scheduling instruction set.

2. The dynamic scheduling method for logistics transportation based on artificial intelligence according to claim 1, characterized in that, The process of processing and integrating multimodal transportation data to obtain a dynamic scheduling knowledge graph includes: Acquire multimodal transportation data; the multimodal transportation data includes vehicle data, environmental data, order data, and infrastructure data; the vehicle data includes vehicle ID, real-time GPS location, speed, direction of travel, remaining mileage, planned stop, loaded cargo list, and destination; the environmental data includes road segment ID, real-time traffic conditions, and short-term weather forecasts; the order data includes information on newly added, modified, canceled, and expedited orders, and the order information includes order ID, pickup point, delivery point, time window, cargo attributes, and priority; the infrastructure data includes warehouse ID and its corresponding real-time operational load, number of available loading / unloading platforms, and estimated waiting time; Multimodal transportation data is cleaned and standardized to obtain several clean data streams; the data cleaning and standardization operations include noise reduction and repair, formatting and structuring operations; A set of association relationships is obtained by performing spatiotemporal alignment and entity association based on several clean data streams; Define an initial knowledge graph; the initial knowledge graph is created using a graph database approach and includes several entity nodes and relation edges; The set of relationships is converted into MERGE and SET operations on the initial knowledge graph to obtain a dynamically scheduled knowledge graph; the MERGE operation refers to merging nodes and relationships, and the SET operation refers to updating attributes.

3. The AI-based dynamic scheduling method for logistics transportation according to claim 2, characterized in that, The set of association relationships obtained by performing spatiotemporal alignment and entity association based on several clean data streams includes: Define a sliding time window; An entity association and binding calculation is performed on several clean data streams within a sliding time window to obtain an association set; the association set includes several entities and the relationships between entities; all data in the association set has a unified timestamp; The entity association and binding calculation includes vehicle-road segment binding, vehicle-order binding, road segment-environment binding, and order-warehouse binding; The vehicle-to-road-segment binding refers to binding the vehicle ID to the road segment ID; The vehicle-order binding refers to binding the vehicle ID to the order ID; The road segment and environment binding refers to binding the road segment ID with its corresponding real-time traffic conditions and future short-term weather forecasts; The order-warehouse binding refers to parsing the pickup and delivery points corresponding to each order ID and binding them to the warehouse ID.

4. The dynamic scheduling method for logistics transportation based on artificial intelligence according to claim 1, characterized in that, The optimized scheduling instruction set obtained by optimizing the dynamic scheduling knowledge graph includes: Extract dynamic scheduling knowledge graphs corresponding to several historical timestamps; A predicted graph is obtained by inputting a dynamic scheduling knowledge graph corresponding to several historical timestamps into a graph prediction model; the graph prediction model is constructed using a spatiotemporal graph neural network model; the spatiotemporal graph neural network model includes a spatiotemporal embedding module, a temporal convolution module, and a spatiotemporal fusion prediction module; The dynamic scheduling knowledge graph and prediction graph are input into the instruction optimization model to obtain the optimized scheduling instruction set; the instruction optimization model is constructed through a hierarchical attention decision network; the hierarchical attention decision network includes a scheduling policy generator and a distributed action executor.

5. The AI-based dynamic scheduling method for logistics transportation according to claim 4, characterized in that, The spatiotemporal embedding module includes: The spatiotemporal embedding module obtains the spatial feature vector of each entity node in the dynamically scheduled knowledge graph at several historical timestamps by using a spatial attention encoder. The spatial attention encoder is used to calculate the attention coefficients corresponding to several entity nodes and aggregate their information to obtain the spatial feature vectors of the entity nodes on the timestamp. The attention coefficient satisfies: Where v represents the entity node whose attention coefficient needs to be calculated, u represents the neighboring entity nodes of the entity node; R represents the edge relationship between entity nodes; and h represents the feature vector corresponding to the entity node. This is represented as concatenating the feature vectors of two entity nodes; It is represented as a learnable attention vector; T represents the matrix transpose operation; It is represented as a learnable weight matrix related to edge relationships. Represented as attention coefficient; This is represented as an activation function.

6. The dynamic scheduling method for logistics transportation based on artificial intelligence according to claim 5, characterized in that, The information aggregation satisfies: ;in, Represented as attention weights, derived from attention coefficients. Obtained by normalization using the softmax function; This is represented as the set of neighboring entity nodes corresponding to entity node v; Represented as the spatial feature vector corresponding to entity node v; It is represented as a non-linear activation function.

7. The AI-based dynamic scheduling method for logistics transportation according to claim 4, characterized in that, The temporal convolution module and the spatiotemporal fusion prediction module include: The input data of the temporal convolution module is the spatial feature vector corresponding to several historical timestamps of the entity node; the output data is the spatiotemporal fusion feature vector of the entity node at the current timestamp; the spatiotemporal fusion feature vector fuses the spatial feature vectors of several historical timestamps before the current timestamp; the temporal convolution module uses a one-dimensional causal convolutional network to perform temporal feature fusion. The spatiotemporal fusion prediction module consists of several prediction heads. Its input data is the spatiotemporal fusion feature vector output by several entity nodes through the temporal convolution module, and its output data is the future prediction result corresponding to several entity nodes. The prediction head includes a core fully connected layer and a task-specific output layer; The core fully connected layer is used to map a general spatiotemporal fusion feature vector to a feature space that matches the task-specific output layer; The task-specific output layer is used to predict the output data of the core fully connected layer to obtain future prediction results; the predicted future results corresponding to several entity nodes are integrated to obtain a prediction graph.

8. The dynamic scheduling method for logistics transportation based on artificial intelligence according to claim 4, characterized in that, The scheduling policy generator includes: The input data of the scheduling policy generator are a dynamic scheduling knowledge graph and a prediction graph, and the output data is a meta-policy vector; the scheduling policy generator includes global graph encoding operation and global pooling and policy generation operation. The global graph encoding operation includes: The fused graph is input into the graph attention encoder to obtain global context feature vectors corresponding to several entity nodes; The fusion graph is obtained by adding the future prediction results corresponding to several entity nodes in the prediction graph as attributes of the entity nodes to the dynamic scheduling knowledge graph. The global pooling and policy generation operations include: Attention pooling is performed on the global context feature vectors corresponding to all entity nodes to obtain a global state vector. The global state vector is then mapped to a meta-policy vector through a policy generation network.

9. The AI-based dynamic scheduling method for logistics transportation according to claim 4, characterized in that, The distributed action executor includes: The input data of the distributed action executor includes vehicle ID and its corresponding attributes, local observation graph, and meta-policy vector; the local observation graph refers to the subgraph structure collected within the K-hop neighbor entity nodes centered on the vehicle ID as the local observation graph corresponding to the vehicle ID; the output data is an optimized scheduling instruction set; where K is an integer, K>0; The working steps of the distributed action executor include: The local observation map is input into the graph neural network to obtain the local structured feature vector; The decision feature vector corresponding to the vehicle ID is obtained by concatenating the attributes, local structured feature vector, and meta-policy vector corresponding to the vehicle ID. The decision feature vector corresponding to the vehicle ID is input into the action evaluation network to obtain the optimized scheduling instruction corresponding to the vehicle ID. The core of the action evaluation network employs an attention mechanism; the workflow of the attention mechanism includes: A set of candidate actions is generated for the vehicle ID, and the set of candidate actions consists of several candidate actions; Each candidate action is encoded as a feature vector as an action embedding; the attention score corresponding to the candidate action is calculated using the decision feature vector as the query and the action embedding as the key. The attention scores of several candidate actions are normalized using a softmax function to obtain the selection probability of each candidate action. Select the candidate action with the highest probability from the vehicle IDs and use it as the optimized scheduling instruction for the vehicle ID. The optimized scheduling instruction set is determined based on the optimized scheduling instructions corresponding to several vehicle IDs.

10. A dynamic scheduling system for logistics transportation based on artificial intelligence, characterized in that: include: A data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected together. The data acquisition module acquires multimodal transportation data through data acquisition equipment; the multimodal transportation data includes transportation vehicle data, environmental data, order data, and infrastructure data. The data analysis module includes a map construction unit and a result generation unit; The graph construction unit processes and integrates multimodal transportation data to obtain a dynamic scheduling knowledge graph. The result generation unit performs scheduling optimization on the dynamic scheduling knowledge graph to obtain an optimized scheduling instruction set.

Images