Hierarchical Graph-Based Emergency Supply Prediction System and Method for Warehouse and Distribution Integration Logistics

Abstract

The present invention discloses a hierarchical graph-based integrated warehouse and distribution logistics emergency supply prediction system and method, which at least includes a micrograph learning module, a macro-graph learning module, and a spatio-temporal joint prediction module. The micrograph module introduces meta-path aggregation of warehouse and distribution network routing features and adopts multi-view learning to obtain the spatial features of warehouse and distribution sites from the routing view and the event view respectively. The macro-graph learning module uses a lightweight graph convolution method to depict the relationship between warehouse clusters and distribution stations in the form of cities in emergency scenarios and learn the spatial correlation features of warehouse and distribution nodes in the macro dimension. The spatio-temporal joint prediction module captures the temporal correlation of supply data and fuses it with spatial features to accurately predict the supply capacity of future logistics sites. This method analyzes the spatio-temporal correlation features of warehouses and distribution stations in the emergency logistics network to achieve the overall goal of improving the prediction accuracy and enhancing the supply capacity at both ends of the warehouse and distribution in emergency scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer intelligent computing and application, and particularly relates to a hierarchical spatio-temporal graph learning technology for emergency supply capacity prediction, and specifically relates to a warehouse-distribution integrated logistics emergency supply prediction system and method based on a hierarchical graph. Background Art

[0002] The integrated warehouse and distribution logistics network (IWDLN) refers to integrating the two links of warehousing and distribution. By integrating warehousing and distribution resources, the efficient operation of logistics nodes is achieved, and logistics costs are optimized. The integrated warehouse and distribution network generally includes three types of nodes: warehouses, sorting centers, and distribution stations. The workflow is as follows:

[0003] 1) The platform distributes users' orders to warehouses across the country;

[0004] 2) The warehouse packs the goods and sends them to the next-level logistics node, usually a warehouse or a sorting center;

[0005] 3) After the goods in the warehouse arrive at the sorting center, further refined sorting will be carried out;

[0006] 4) The goods sent from the sorting center will be delivered to the distribution station near the user;

[0007] 5) Finally, the courier at the distribution station delivers the package to the user's hands.

[0008] The above completes the process of one order delivery. However, in the face of disruptions caused by unforeseen events such as heavy rain, typhoons, public health emergencies, and security incidents, the IWDSN may face difficulties. The vulnerabilities caused by these emergencies make it a challenging task to predict the logistics supply capacity in actual scenarios.

[0009] Due to the large scale and long cycle of routes within the entire country or even the whole world in the IWDSN, the impact of emergency events on each node on the route will be significantly different. Therefore, although researchers in academia and industry have made certain progress, predicting the logistics supply capacity is still challenging in the face of complex actual situations. Therefore, to solve the problem of emergency supply prediction, it is necessary to predict the supply capacity of both warehouses and distribution stations simultaneously.

[0010] Existing research on the warehousing, distribution, and supply capabilities in emergency situations mainly focuses on the design and optimization of temporary emergency logistics networks, the optimal allocation and scheduling of inventory, and the assessment of emergency logistics capabilities. These studies mainly use mathematical models and optimization algorithms to analyze resource planning and network design in emergency logistics scenarios. These efforts regard emergency logistics networks as temporary solutions to disasters. Additionally, due to the inherent complexity and uncertainty of emergency logistics scenarios, these techniques often struggle to handle large-scale global data, reveal potential correlations within heterogeneous data, and thus overlook the effective representation of nodes in the emergency network and the prediction of their future states.

[0011] Furthermore, although graph neural networks have made progress in spatio-temporal prediction tasks and are widely used in accident prediction, traffic demand prediction, and mobility prediction, they typically focus on the autoregressive evolution of spatio-temporal data and do not consider the different impacts of emergency events on human activities under different time and space conditions. Summary of the Invention

[0012] The present invention precisely addresses the problem in the prior art of lacking consideration for the impact of emergency events in the warehousing and distribution integrated logistics supply scenario. It provides a warehousing and distribution integrated logistics emergency supply prediction system and method based on a hierarchical graph, including a micro-graph learning module, a macro-graph learning module, and a spatio-temporal joint prediction module. The micro-graph module introduces meta-paths to aggregate large-scale warehousing and distribution network routing features and adopts multi-view learning to obtain the spatial features of warehousing and distribution sites from the routing view and the event view respectively. The macro-graph learning module uses a lightweight graph convolution method to depict the relationship between warehouse clusters and distribution stations at the city level in the emergency scenario and learn the spatial correlation features of warehousing and distribution nodes at the macro dimension. The spatio-temporal joint prediction module captures the temporal correlation of supply data and fuses it with the spatial features to accurately predict the supply capabilities of future logistics sites. This method aims to improve the prediction accuracy and enhance the overall supply capabilities at both the warehousing and distribution ends in the emergency scenario by analyzing the spatio-temporal correlation features of warehouses and distribution stations in the emergency logistics network.

[0013] To achieve the above objective, the technical solution adopted by the present invention is: a warehousing and distribution integrated logistics emergency supply prediction system based on a hierarchical graph, including a micro-graph learning module, a macro-graph learning module, and a spatio-temporal joint prediction module,

[0014] The micro-graph learning module: performs routing aggregation operations at the micro level using meta-paths and enhances event information using the method of multi-view learning to obtain the spatial features of warehousing and distribution sites from the routing view and the event view respectively;

[0015] The macroscopic graph learning module: According to the macroscopic logistics graph information, based on the lightweight graph convolution method of LightGCN, capture the relationship between the warehouse cluster and the distribution station, and learn the spatial association features between the warehouse and the distribution station at the macroscopic dimension; the warehouse cluster is based on cities;

[0016] The spatio-temporal joint prediction module: Align the spatial features output by the microscopic graph learning module and the macroscopic graph learning module, capture the correlation of supply data in the time dimension, and fuse it with the learned multi-layer spatial features to achieve the prediction of the future supply capacity of logistics sites.

[0017] As an improvement of the present invention, the microscopic graph learning module includes routing aggregation and event enhancement,

[0018] In the routing aggregation, various meta-paths are used to manage the microscopic topology graph data Extract the routing spatial features of the warehouse and the distribution station Identify the key meta-paths based on the weights of all nodes to obtain the spatial representation of the routing view

[0019] In the event enhancement, multi-view learning is used to introduce event-related information and learn in parallel with the model to respectively obtain the spatial representation of node v from the routing view and the event view and

[0020] As another improvement of the present invention, the spatio-temporal joint prediction module includes hierarchical spatial feature alignment and spatio-temporal fusion,

[0021] In the hierarchical spatial feature alignment, fuse the spatial representations of node types between the microscopic graph and the macroscopic graph Introduce a hierarchical gating mechanism to control the amount of spatial information flowing from the macroscopic level to the microscopic level, so as to respectively obtain the final spatial representations g w and g d ;

[0022] In the spatio-temporal fusion, the time series of each type of node is input into a set of standard dilated 1D convolutional filters to extract high-level time features The time features are respectively passed to two TCN blocks to obtain the spatial representations of the warehouse node and the distribution station node.

[0023] To achieve the above object, the technical solution adopted by the present invention is also: a hierarchical graph-based integrated warehouse and distribution logistics emergency supply prediction method, which is characterized in that it specifically includes the following steps:

[0024] S1: Based on the collected microscopic graph data Consider a form like Warehouse - Sorting Center - Delivery Station W i →S j →…→D m of the meta - path By aggregating the node P(u, v), learn the spatial representation of the warehouse and the delivery station in the micro - space

[0025] S2: Identify the key meta - path based on the weights of all nodes, and obtain the spatial representation of the routing view

[0026] S3: Deeply study the micro - graph from the perspective of events, consider the routing information and event information, and use graph deep learning to realize from different perspectives and to observe and learn the logistics graph structure, model the complex relationships in the logistics scenario under events, and understand the spatial correlation of the micro - graph ;

[0027] S4: Based on the lightweight graph convolution method of LightGCN, learn the spatial association features between the warehouse in the macro - urban dimension and the delivery station

[0028] S5: According to the geographical location correspondence between the city and the warehouse, map the city nodes in the macro - graph to the semantic space of the warehouse nodes in the micro - graph, use the gating mechanism to control the amount of information flowing from the macro - level to the micro - level, and obtain the final embedded representation of the warehouse that integrates hierarchical spatial information

[0029] S6: Repeat step S5, use the gating mechanism to control the amount of information flowing from the macro - level to the micro - level, and obtain the final embedded representation of the delivery station that integrates hierarchical spatial information

[0030] S7: Input the time series of each type of node into a set of standard dilated 1D convolutional filters to extract high - level time features The time features are respectively passed to two TCN blocks to obtain the supply - capacity embedded representations of the warehouse nodes and the delivery - station nodes, and output the predicted future supply order volume of the model through a layer of MLP

[0031] As an improvement of the present invention, in step S1, design a meta - path encoder to aggregate the neighbor nodes of the target node v:

[0032]

[0033] wherein, is the embedded - layer representation vector of the node, and W P is the trainable meta - path - specific weight matrix

[0034] As another improvement of the present invention, the modeling in step S3 is as follows:

[0035]

[0036] wherein, is the routing view of node v of type τ, is to obtain the spatial representation of node v in the event view.

[0037] As another improvement of the present invention, in step S4, based on the LightGCN method, the spatial feature information representation of cities and distribution stations in the macro graph is extracted and learned at the macro level: and :

[0038]

[0039] wherein, contains the information passed to cities and distribution stations at the k-th layer; the information at the 0-th layer will be randomly initialized; and represent the neighbor nodes of cities and distribution stations in the macro graph;

[0040]

[0041] wherein, and respectively represent the spatial feature information of cities and distribution stations at the macro level, and K is the number of layers of the model.

[0042] As yet another improvement of the present invention, in step S5, the city nodes in the macro graph are mapped to the semantic space of the warehouse nodes in the micro graph

[0043]

[0044] wherein, is the spatial feature information representation of the city at the macro level, is the spatial representation of the warehouse node; M cw is the 0-1 conversion matrix for recording node information, 1 represents that warehouse w is located in city i, and 0 vice versa;

[0045] The gating mechanism is used to control the amount of information flowing from the macro level to the micro level:

[0046]

[0047] wherein, represents the gating matrix of the warehouse node; is the learnable network parameter;

[0048]

[0049] where h w is the final embedded representation of the warehouse that fuses hierarchical spatial information.

[0050] As a further improvement of the present invention, in step S7, the gated TCN block includes a 1-D causal convolution and also includes a gated linear unit (GLU) as a non-linear activation function:

[0051]

[0052] where is the high-level temporal feature and serves as the input to the TCN; Θ1, Θ2, b, c are learnable network parameters, * represents a one-dimensional convolution operation, and ⊙ represents an element-wise Hadamard product; φ(·) represents the hyperbolic tangent activation function, and σ(·) represents the sigmoid function.

[0053] Compared with the prior art, the technical advantages and effects of the present invention are as follows: The present invention provides a method for predicting the integrated warehouse and distribution emergency supply based on a hierarchical graph. First, a micrograph and a macro-graph are depicted to describe the relationships among warehouses, sorting centers, and distribution stations. In micrograph learning, event information is injected into the micrograph through node matching, and an event-enhanced node representation is obtained based on multi-view learning. A meta-path aggregation strategy is designed to dynamically select the key routes in the IWDSN and update the route view node representation through the graph learning layer. In macro-graph learning, a bipartite graph learning method based on LightGCN is adopted to obtain the spatial representations of warehouses and distribution stations. Finally, hierarchical embedding fusion is performed on the nodes from the micrograph and the macro-graph, and the macro-node information is passed to the micro-warehouse nodes and distribution stations. The output layer is responsible for accurately predicting the future supply capacity of the IWDSN. The method of the present invention is based on a mature integrated warehouse and distribution network, and can efficiently predict the daily supply capacity of warehouses and distribution stations under the influence of emergencies, ensuring the efficient allocation and stable operation among logistics nodes in emergency logistics, and has a wide range of applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a structural diagram of the integrated warehouse and distribution emergency supply prediction system based on a hierarchical graph according to the present invention;

[0055] Figure 2 is a schematic diagram of the step sequence of the method for predicting the integrated warehouse and distribution emergency supply based on a hierarchical graph according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] The present invention will be further illustrated below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

[0057] Embodiment 1

[0058] The specific definitions of the relevant symbols involved in this embodiment are as follows:

[0059]

[0060]

[0061] The integrated warehouse and distribution logistics emergency supply prediction system based on the hierarchical graph, as Figure 1 shown, at least includes a microscopic graph learning module, a macroscopic graph learning module, and a spatio-temporal joint prediction module.

[0062] Microscopic graph learning module: including routing aggregation and event enhancement; first, perform routing aggregation operations, effectively learn large-scale warehouse and distribution network graph data through the aggregation of meta-paths, and use the method of multi-view learning to enhance event information, and obtain the spatial characteristics of warehouse and distribution sites from the routing view and the event view respectively.

[0063] Since IWDSN will reconstruct its routing strategy in case of emergency, for example, temporarily allocate new warehouse supplies or use interchangeable sorting centers for transportation, the routes are often variable and dynamic. Therefore, first use meta-paths to perform routing aggregation operations at the microscopic level to effectively learn large-scale warehouse and distribution network graph data and obtain dynamic topological information; and use the method of multi-view learning to enhance event information, thereby obtaining the comprehensive spatial characteristics of warehouse and distribution sites from the routing view and the event view respectively. The input view information includes the characteristics of various sites (such as warehouse level, area where the distribution station is located), and event information (such as severity of sudden public events, number of sealed areas, rainfall, etc.). The learned characteristics include context information at the microscopic level, heterogeneous site information, and dynamic event information.

[0064] In the routing aggregation operation, first use various meta-paths to efficiently manage large-scale topological graph data and efficiently extract the spatial representations of warehouses and distribution stations in the microscopic space Secondly, identify the key meta-paths based on the weights of all nodes to obtain the spatial representation of the routing view

[0065] The event enhancement operation first uses multi-view learning, introduces event-related information, and follows the parallel learning of the model, and then realizes obtaining the spatial representations of node v from the routing view and the event view respectively and And comprehensively understand the microscopic graph 's spatial correlation, including routing association, event impact, and context semantic information.

[0066] Macroscopic graph learning module: Designed a learning method for urban-distribution station features. According to the macroscopic logistics graph information, a lightweight graph convolution method based on LightGCN was used to effectively capture the relationship between the warehouse clusters and distribution stations at the city level during emergency logistics, and then learn the spatial association features between the warehouses and distribution stations in the macroscopic urban dimension;

[0067] Due to the regional characteristics of the nodes in IWDSN (including population, economy, and climate), the responses of nodes in different cities to supply vary greatly. It is very important to capture the diverse features of nodes according to their regional subordination relationships to enhance the response modeling of different nodes. Therefore, according to the geographical feature that there are multiple warehouses with high similarity under a city, multiple warehouse sites in the same city are abstracted as city sites to construct a macroscopic graph; designed a learning method for urban-distribution station features. According to the macroscopic logistics graph information, the LightGCN method was introduced to effectively capture the relationship between the city and the distribution station during emergency logistics, lightly capture the relationship between the warehouse clusters and the distribution station at the city level during emergency logistics, and then learn the spatial association features between the warehouse and the distribution station in the macroscopic urban dimension, while reducing the computational complexity of the model.

[0068] The spatio-temporal joint prediction module: This module is a spatio-temporal feature fusion block based on the Gated Temporal Convolutional Network (Gated TCN), which uses multi-layer temporal convolution to capture temporal correlation and performs fusion between spatial and temporal features, including a spatio-temporal fusion block and an output module. In the spatio-temporal module, information at different levels is fused to align the spatial features of the microscopic graph and the macroscopic graph on various types of sites, and a gating mechanism is used to control the amount of information flow; secondly, the historical data of warehouse distribution is encoded to capture the correlation of supply sequence data in the time dimension and fuse it with the multi-layer spatial features. Finally, an accurate prediction of the future supply capacity of logistics sites is achieved in the output module.

[0069] This method models the dynamic spatio-temporal impact of IWDSN to achieve accurate prediction of the future supply capacity of logistics sites and reduce the negative impact caused by emergencies to the logistics supply network.

[0070] Example 2

[0071] Using the method described in Example 1, a method for predicting emergency supply of integrated warehouse and distribution based on a hierarchical graph, as Figure 2 shown, includes the following steps:

[0072] In the micrograph learning module, first, a meta-path is used to perform routing aggregation operations at the micro level to effectively learn large-scale warehouse distribution network graph data and obtain dynamic topological information. Then, a multi-view learning method is adopted to enhance event information, and thus the comprehensive spatial features of warehouse distribution sites are obtained from the routing view and the event view respectively.

[0073] Step S1: Based on the collected micrograph data Taking into account meta-paths in the form of warehouse - sorting center - distribution station W i →S j →…→D m such as By aggregating nodes P(u, v), the spatial representations of the warehouse and the distribution station in the micro space are learned Effectively express the spatial routing diversity.

[0074] Step S2: Identify the key meta-paths based on the weights of all nodes to obtain the spatial representation of the routing view

[0075] Design a meta-path encoder to aggregate the neighbor nodes of the target node v:

[0076]

[0077] where is the embedding layer representation vector of the node, and W P is the trainable meta-path specific weight matrix.

[0078] After encoding the meta-path node representations, the multi-head attention mechanism is used to obtain the importance of each meta-path, denoted as where:

[0079]

[0080] where represents the spatial information of the node on the meta-path P i and σ(·) represents the activation function.

[0081] Step S3: Deeply study the micrograph from the perspective of events, considering routing information and event information. The mapping of external heterogeneous event information is defined as:

[0082]

[0083] where is the original feature vector, is the projected latent vector of node i. is the parameter weight matrix of nodes of type τ; Using graph deep learning to achieve from different perspectives and To observe and learn the logistics graph structure, and accurately model the complex relationships in logistics scenarios under events:

[0084]

[0085] in, The routing view for a node v of type τ (e.g., a warehouse or distribution station), It is to obtain the spatial representation of node v in the event view. Through the event injection enhancement method, the emergency prediction accuracy of the model under sudden events can be obtained.

[0086] Step S4: Based on the lightweight graph convolution method of LightGCN, learn the spatial correlation characteristics between warehouses and distribution stations in the macro city dimension.

[0087] In the macro graph learning module, the city-business station feature learning operation uses the LightGCN method to learn the macro graph. Spatial characteristic information representation of medium-sized cities and distribution stations at the macro level and Extract and learn, and ignore unnecessary self-connection features to achieve lightweight model calculation:

[0088]

[0089] in, Contains the information passed to cities and distribution stations at the kth layer, and the information at the 0th layer will be randomly initialized; and Represents the neighbor nodes of cities and delivery stations in the macro graph.

[0090]

[0091] in, and They represent the spatial characteristic information of the city and the distribution station at the macro level respectively, and K is the number of layers of the model.

[0092] Step S5: According to the geographical location correspondence between cities and warehouses, the city nodes in the macro graph are mapped to the semantic space of the warehouse nodes in the micro graph. The gating mechanism is used to control the amount of information flowing from the macro level to the micro level, and the final embedded representation of the warehouse that integrates the hierarchical spatial information is obtained.

[0093] The spatiotemporal joint prediction module includes hierarchical spatial feature alignment and spatiotemporal fusion. The hierarchical spatial feature alignment processes the micro-graphs. and macro picture The semantic difference between and macro picture The spatial representation of the node types in between is presented, and a hierarchical gating mechanism is introduced to control the amount of spatial information flowing from the macroscopic level to the microscopic level, so as to obtain the final spatial representations g w and g d . The hierarchical spatial feature alignment maps the city nodes in the macroscopic graph to the semantic space of the warehouse nodes in the microscopic graph according to the geographical location correspondence between the city and the warehouse:

[0094]

[0095] Among them, is the representation of the spatial feature information of the city at the macroscopic level, is the spatial representation of the warehouse node; M cw is the 0-1 conversion matrix recording node information, where 1 means the warehouse w is located in the city i, and 0 means otherwise. Then, the gating mechanism is used to control the amount of information flowing from the macroscopic level to the microscopic level:

[0096]

[0097] Among them, represents the gating matrix of the warehouse node; is the learnable network parameter.

[0098]

[0099] The obtained h w is the final embedded representation of the warehouse integrating hierarchical spatial information.

[0100] Step S6: Since the distribution station nodes in the macroscopic graph and the distribution station nodes in the microscopic graph exist in the same semantic space, there is no need for mapping. The gating mechanism is used to control the amount of information flowing from the macroscopic level to the microscopic level, and the final embedded representation h d of the distribution station integrating hierarchical spatial information is obtained.

[0101] Step S7: The spatio-temporal fusion inputs the time series of each type of node into a set of standard dilated 1D convolutional filters to extract high-level time features Then, these time features are respectively passed to two TCN blocks. Finally, the spatial representations of the warehouse nodes and the distribution station nodes are obtained, which capture the node routing features, event impacts, and context semantic information across different levels and perspectives.

[0102] To learn complex time dependencies, the spatio-temporal fusion introduces gated TCN blocks, which contain a 1-D causal convolution followed by a gated linear unit (GLU) as the non-linear activation function.

[0103]

[0104] Among them, is a high-level temporal feature and serves as the input to the TCN. Θ1, Θ2, b, c are learnable network parameters, * represents a one-dimensional convolution operation, and ⊙ represents an element-wise Hadamard product; φ(·) represents the hyperbolic tangent activation function, and σ(·) represents the sigmoid function.

[0105] To maintain the high purity of temporal features during the spatio-temporal feature fusion process, the spatial feature is only passed to one of the two gated TCN blocks, while the other gated TCN block focuses on the temporal features of deep learning. Residual connections and skip connections are introduced to avoid the vanishing gradient problem and maintain the consistency of the sequence length. The output block consists of two 1×1 standard convolutional layers, which convert the channel dimension to the required output dimension. When the goal is to predict a sequence of continuous Y time steps, the required output dimension is Y.

[0106] In summary, the present invention discloses the problem of predicting the supply capacity of integrated warehouse and distribution under emergency conditions and proposes a hierarchical spatio-temporal graph learning including three modules. 1) At the micro level, a novel meta-path-based dynamic graph learning method is designed to accurately represent the complex spatial features of nodes affected by events. 2) At the macro level, a lightweight heterogeneous graph learning method is designed to effectively capture different patterns of nodes and hierarchically extract spatial features. 3) In the prediction, a spatio-temporal feature fusion method is designed to accurately predict the future supply capacity of warehouses and delivery stations. The framework of the method of the present invention can be extended to handle multiple emergency or periodic events, such as shopping festivals and social news events, contribute to the development of emergency logistics management, and enhance stability and performance in different types of emergency situations.

[0107] It should be noted that the above content only illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, several improvements and refinements can still be made, and these improvements and refinements all fall within the protection scope of the claims of the present invention.

Claims

1. A method for predicting emergency supply of integrated warehouse and distribution logistics based on a hierarchical graph, characterized in that Specifically, it includes the following steps: S1: Based on the collected microscopic image data Consider the meta-path in the form of warehouse - sorting center - distribution station W i →S j →···→D m of the meta-path By aggregating the nodes P(u, v), learn the spatial representation of the warehouse and the distribution station in the microscopic space S2: Identify key meta-paths based on the weights of all nodes to obtain the spatial representation of the routing view S3: Deeply study the microscopic graph from the perspective of events, consider routing information and event information, and use graph deep learning to observe and learn the logistics graph structure from different perspectives and to model the complex relationships in the logistics scenario under events, understand the spatial correlation of the microscopic graph ; S4: Based on the lightweight graph convolution method of LightGCN, learn the spatial correlation features between the warehouse and the distribution station at the macroscopic urban dimension; S5: According to the geographical location correspondence between the city and the warehouse, map the city nodes in the macroscopic graph to the semantic space of the warehouse nodes in the microscopic graph, and use the gating mechanism to control the amount of information flowing from the macroscopic level to the microscopic level, so as to obtain the final embedded representation of the warehouse that integrates hierarchical spatial information; S6: Repeat step S5, use the gating mechanism to control the amount of information flowing from the macroscopic level to the microscopic level, and obtain the final embedded representation of the distribution station that integrates hierarchical spatial information; S7: The time series of each type of node are input into a set of standard dilated 1D convolutional filters to extract high-level temporal features. The temporal features are respectively passed to two TCN blocks to obtain the embedding representations of the supply capabilities of the warehouse nodes and the distribution station nodes, and the future supply order quantities predicted by the model are output through a layer of MLP.

2. The method for predicting emergency supply of integrated warehouse and distribution logistics based on a hierarchical graph according to claim 1, wherein: In step S1, design a meta-path encoder to aggregate the neighbor nodes of the target node v: Among them, is the embedding layer representation vector of the node, and W P is a trainable meta-path specific weight matrix.

3. The method for predicting emergency supply of integrated warehouse and distribution logistics based on a hierarchical graph according to claim 1, wherein: Modeled in step S3 as: Among them, is the routing view of node v of type τ, is to obtain the spatial representation of node v in the event view.

4. The method for predicting emergency supply of warehouse distribution integration logistics based on a hierarchical graph according to claim 1, characterized in that: In step S4, based on the LightGCN method, for the macro graph the spatial feature information of cities and distribution stations at the macro level and is extracted and learned: Among them, it contains the information passed from the k-th layer to the cities and distribution stations; the information of the 0-th layer will be randomly initialized; and represent the neighbor nodes of the cities and distribution stations in the macroscopic graph; Among them, and respectively represent the spatial characteristic information of the city and the distribution station at the macro level, and K is the number of layers of the model.

5. The method for predicting emergency supply of integrated warehouse and distribution logistics based on a hierarchical graph according to claim 1, wherein: In step S5, map the city nodes in the macroscopic graph to the semantic space of the warehouse nodes in the microscopic graph Among them, is the spatial feature information representation of the city at the macroscopic level, is the spatial representation of the warehouse node; M cw is the 0-1 conversion matrix for recording node information, where 1 represents that warehouse w is located in city i, and 0 vice versa; Use the gating mechanism to control the amount of information flowing from the macroscopic level to the microscopic level: Among them, represents the gating matrix of the warehouse node; is a learnable network parameter; Among them, h w is the final embedded representation of the warehouse that integrates hierarchical spatial information.

6. The method for predicting emergency supply of warehouse distribution integration logistics based on a hierarchical graph according to claim 1, wherein: In step S7, the gated TCN block includes a 1-D causal convolution and also includes a gated linear unit (GLU) as a non-linear activation function: Among them, is a high-level time feature and serves as the input to the TCN; Θ1, Θ2, b, c are learnable network parameters, ★ represents a one-dimensional convolution operation, and ⊙ represents an element-specific Hadamard product; φ(·) represents the hyperbolic tangent activation function, and σ(·) represents the sigmoid function.

7. The integrated warehouse and distribution logistics emergency supply prediction system based on a hierarchical graph, using the method described in claim 1, is characterized in that: It includes a microscopic graph learning module, a macroscopic graph learning module, and a spatio-temporal joint prediction module, The microscopic graph learning module: uses meta-paths to perform routing aggregation operations at the microscopic level, and uses the method of multi-view learning to enhance event information, and obtains the spatial features of the warehouse and distribution sites from the routing view and the event view respectively; The macroscopic graph learning module: According to the macroscopic logistics graph information, based on the lightweight graph convolution method of LightGCN, capture the relationship between the warehouse cluster and the distribution station, and learn the spatial correlation features between the warehouse and the distribution station at the macroscopic dimension; the warehouse cluster is based on cities; The spatio-temporal joint prediction module: align the spatial features output by the microscopic graph learning module and the macroscopic graph learning module, capture the correlation of supply data in the time dimension, and fuse it with the learned multi-layer spatial features to realize the prediction of the future supply capacity of the logistics site.

8. The integrated warehouse and distribution logistics emergency supply prediction system based on a hierarchical graph according to claim 7, characterized in that: The microscopic graph learning module includes routing aggregation and event enhancement, In the routing aggregation, various meta-paths are used to manage the micro-topology graph data Extract the spatial representations of the warehouse and distribution stations in the micro space Identify the key meta-paths based on the weights of all nodes to obtain the spatial representation of the routing view In the event enhancement, multi-view learning is utilized to introduce event-related information and learn in parallel with the model, so as to obtain the spatial representation of node v from the routing view and the event view respectively. and 9. The integrated warehouse and distribution logistics emergency supply prediction system based on a hierarchical graph according to claim 8, wherein: The spatio-temporal joint prediction module includes hierarchical spatial feature alignment and spatio-temporal fusion, In the hierarchical spatial feature alignment, the spatial representations of node types between the microscopic graph and the macroscopic graph are fused, and a hierarchical gating mechanism is introduced to control the amount of spatial information flowing from the macroscopic level to the microscopic level, so as to obtain the final spatial representations g w and g d ; In the spatio-temporal fusion, the time series of each type of node are input into a set of standard dilated 1D convolutional filters to extract high-level temporal features. The temporal features are respectively passed to two TCN blocks to obtain the spatial representations of the warehouse nodes and the distribution station nodes.

Images