Method and device for traffic prediction of logistics nodes
By mapping logistics nodes to implicit representation nodes and combining temporal convolution and graph convolutional networks, the problems of prediction accuracy and computational complexity in dynamic logistics networks are solved, achieving efficient and accurate logistics flow prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2026-05-29
- Publication Date
- 2026-07-24
AI Technical Summary
Existing urban logistics spatiotemporal forecasting schemes are ill-suited to dynamically generated and disappearing active flow units. They are prone to mean regression bias when dealing with long-tailed distributed order demand data, have high computational complexity, and are difficult to meet the requirements of real-time scheduling and online deployment.
Multiple real logistics nodes are mapped to a small number of preset implicit representation nodes. The node features are processed by temporal convolution and graph convolutional networks to generate an allocation matrix. The features of the implicit representation nodes are used for traffic prediction, and the model is optimized by combining an adaptive weight function.
It improves the prediction accuracy of high-demand nodes under long-tail distribution, reduces computational complexity, and meets the online deployment needs of instant delivery and urban logistics scheduling.
Smart Images
Figure CN122453292A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of urban logistics forecasting technology, and in particular to a method and apparatus for forecasting the flow of logistics nodes. Background Technology
[0002] Urban logistics networks are characterized by dynamic topology changes, long-tail distribution of order demand, and massive network scale. Accurate and efficient spatiotemporal forecasting of urban logistics is the core foundation for the implementation of on-demand delivery, shared mobility, and intelligent urban logistics scheduling, which can effectively improve logistics efficiency and reduce delivery scheduling costs.
[0003] Current solutions for spatiotemporal forecasting of urban logistics have significant technical shortcomings in practical applications, making them difficult to adapt to complex and dynamic logistics scenarios. Firstly, existing solutions, relying on fixed nodes or fixed graph structures, struggle to accommodate dynamically generated and disappearing active flow units. Secondly, when dealing with long-tailed order demand data, the common use of a single forecasting mechanism for unified modeling easily leads to mean regression bias, resulting in insufficient forecasting of high-demand nodes. Thirdly, performing global modeling on large-scale logistics networks incurs high computational complexity, making it difficult to meet the requirements of real-time scheduling and online deployment. Summary of the Invention
[0004] This application provides a method and apparatus for predicting the flow of logistics nodes, which addresses the shortcomings of existing urban logistics spatiotemporal prediction schemes in practical applications.
[0005] This application provides a method for predicting the flow of logistics nodes, applied to electronic devices, including the following steps: Obtain historical features corresponding to multiple real logistics nodes; based on the historical features, map each logistics node to multiple preset implicit representation nodes, where the number of implicit representation nodes is less than the number of logistics nodes; based on the features of each implicit representation node, determine the change range of each implicit representation node in the future period; perform reverse mapping on the change range of each implicit representation node in the future period to determine the traffic prediction result of each logistics node in the future period.
[0006] According to the flow prediction method for logistics nodes provided in this application, based on historical characteristics, each logistics node is mapped to a plurality of preset implicit representation nodes, including: The historical features of each logistics node are processed by a temporal convolutional layer to obtain the temporal features of the logistics node; the temporal features of each logistics node and the features of the adjacent nodes of the logistics node are fused by a graph convolutional network to obtain the node spatial fusion features; the allocation matrix corresponding to each logistics node is determined; the allocation matrix is used to represent the probability that each logistics node belongs to different implicit representation nodes; based on the allocation matrix, the node spatial fusion features are weighted and aggregated to obtain the features of each implicit representation node.
[0007] According to the flow prediction method for logistics nodes provided in this application, the allocation matrix corresponding to each logistics node is determined, including: Based on the spatial fusion features and volume features of nodes, an allocation matrix is generated. The volume features are used to characterize the volume level of the historical flow of each logistics node.
[0008] According to the flow prediction method for logistics nodes provided in this application, an allocation matrix is generated based on node spatial fusion features and volume features, including: The node space fusion features and volume features are concatenated; the concatenated result is subjected to a nonlinear transformation to obtain the matching degree between each logistics node and each implicit representation node; based on the matching degree between each logistics node and each implicit representation node, an allocation matrix is generated.
[0009] According to the flow prediction method for logistics nodes provided in this application, an allocation matrix is generated based on the matching degree between each logistics node and each implicit representation node, including: The matching degree between each logistics node and each implicit representation node is selected using Top-k filtering. The results of Top-k filtering are then normalized to generate an allocation matrix.
[0010] According to the flow prediction method for logistics nodes provided in this application, based on the characteristics of each implicit representation node, the change range of each implicit representation node in future time periods is determined, including: The features of each implicit representation node are input into the encoder to extract the temporal features of each implicit representation node; the temporal features of each implicit representation node and the mask label embedding vector are concatenated; the mask label embedding vector carries temporal information; the decoder performs multi-head self-attention calculation on the concatenated result to obtain the change magnitude of each implicit representation node in the future time period.
[0011] According to the flow prediction method for logistics nodes provided in this application, the change magnitude of each implicitly represented node in future time periods is back-mapped to determine the flow prediction result of each logistics node in future time periods, including: Based on the allocation matrix corresponding to each logistics node, the change range of each implicit representation node in the future time period is reverse-mapped to obtain the change range of each logistics node in the future time period; the allocation matrix is used to represent the probability that each logistics node belongs to different implicit representation nodes; based on the change range of each logistics node in the future time period and the node space fusion characteristics of the logistics node, the traffic prediction result of each logistics node in the future time period is determined.
[0012] According to the flow prediction method for logistics nodes provided in this application, based on the change range of each logistics node in the future time period and the node spatial fusion characteristics of the logistics nodes, the flow prediction result of each logistics node in the future time period is determined, including: The variation range of each logistics node in the future time period and the node spatial fusion characteristics of the logistics node are residually fused to obtain the fusion characteristics; based on the volume characteristics and fusion characteristics, corresponding prediction experts are assigned to logistics nodes with different traffic volumes from the expert hybrid prediction network; the expert hybrid prediction network contains multiple prediction experts used to fit the characteristics of logistics nodes with different traffic volumes; through the prediction experts corresponding to each logistics node, the traffic prediction results of each logistics node in the future time period are determined.
[0013] According to the flow prediction method for logistics nodes provided in this application, an electronic device is equipped with a gating network, which is used to assign corresponding prediction experts to logistics nodes with different flow rates; the method further includes: Obtain the actual traffic flow of each logistics node in the future time period; optimize the gating network based on the prediction experts corresponding to the actual traffic flow and the corresponding prediction experts assigned to the logistics nodes with different traffic flows.
[0014] According to the flow prediction method for logistics nodes provided in this application, a spatiotemporal prediction model is deployed in the electronic device. The spatiotemporal prediction model is used to determine the flow prediction results of each logistics node in future time periods. The method further includes: Obtain the actual traffic flow of each logistics node in the future time period; construct an adaptive weight function; the adaptive weight function is positively correlated with the actual traffic flow; based on the actual traffic flow and traffic prediction results, optimize the spatiotemporal prediction model through the adaptive weight function.
[0015] This application also provides a flow prediction device for a logistics node, comprising the following modules: The acquisition module is used to acquire historical features corresponding to multiple real logistics nodes.
[0016] The mapping module is used to map each logistics node to a number of preset implicit representation nodes based on historical features; the number of implicit representation nodes is less than the number of logistics nodes.
[0017] The determination module is used to determine the magnitude of change of each implicit representation node in future time periods based on the characteristics of each implicit representation node.
[0018] The reverse mapping module is used to reverse map the change magnitude of each implicit representation node in the future period to determine the flow prediction results of each logistics node in the future period.
[0019] According to the flow prediction device for a logistics node provided in this application, the mapping module is specifically used for: The historical features of each logistics node are processed by a temporal convolutional layer to obtain the temporal features of the logistics node; the temporal features of each logistics node and the features of the adjacent nodes of the logistics node are fused by a graph convolutional network to obtain the node spatial fusion features; the allocation matrix corresponding to each logistics node is determined; the allocation matrix is used to represent the probability that each logistics node belongs to different implicit representation nodes; based on the allocation matrix, the node spatial fusion features are weighted and aggregated to obtain the features of each implicit representation node.
[0020] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the flow prediction method for any of the logistics nodes described above.
[0021] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the flow prediction method for logistics nodes as described above.
[0022] This application also provides a computer program product, including a computer program that, when executed by a processor, implements a flow prediction method for any of the logistics nodes described above.
[0023] This solution acquires historical features of multiple real logistics nodes and maps the dynamically changing and variable number of physical logistics nodes to a smaller and fixed number of pre-defined implicit representation nodes. This completely decouples the modeling from the scale and topology of the original logistics graph, effectively adapting to the dynamic generation and disappearance of active flow units in the logistics network. This enables inductive generalization of unseen subgraphs in dynamic logistics networks. Furthermore, this latent space modeling approach abandons the traditional model of using a uniform prediction mechanism for all long-tail demand nodes. It can accurately infer the future change magnitude of each implicit representation node based on its characteristics, avoiding mean regression bias and significantly improving the prediction accuracy of high-demand nodes under long-tail distributions. Furthermore, by compressing high-dimensional, large-scale physical node operations into a low-dimensional, fixed-number implicit node space for time-series extrapolation, the computational volume and resource consumption are significantly reduced, effectively lowering the computational complexity in large-scale graph prediction tasks. Finally, by mapping the evolutionary changes of implicit nodes back to the physical node space, the traffic prediction results for each logistics node in the future are accurately output, comprehensively improving the generalization, accuracy, and efficiency of dynamic logistics network prediction, thereby better meeting the online deployment needs in scenarios such as instant delivery, shared mobility, and urban logistics scheduling. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is one of the flowcharts illustrating a flow prediction method for a logistics node provided in this application.
[0026] Figure 2 This is the second flowchart of a logistics node flow prediction method provided in this application.
[0027] Figure 3 This is the third flowchart of a logistics node flow prediction method provided in this application.
[0028] Figure 4 This is the fourth flowchart of a logistics node flow prediction method provided in this application.
[0029] Figure 5 This is a schematic diagram of a flow prediction method for a logistics node provided in this application.
[0030] Figure 6 This is a schematic diagram of the distribution of urban characteristics provided in this application.
[0031] Figure 7 This is a schematic diagram of an experimental result provided in this application.
[0032] Figure 8 This is a schematic diagram of the structure of a flow prediction device for a logistics node provided in this application.
[0033] Figure 9 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation
[0034] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0035] Urban logistics networks are characterized by dynamic topology changes, long-tail distribution of order demand, and massive network scale. Accurate and efficient spatiotemporal forecasting of urban logistics is the core foundation for the implementation of on-demand delivery, shared mobility, and intelligent urban logistics scheduling, which can effectively improve logistics efficiency and reduce delivery scheduling costs.
[0036] Current solutions for urban logistics spatiotemporal prediction include: 1. By acquiring spatiotemporal map data and combining it with historical logistics time information, a prediction model is used to predict the probability distribution of package arrival times at target nodes, and the number of packages is obtained by sampling based on the probability distribution, achieving high-precision logistics prediction. 2. Combining order and non-order package data, graph neural networks are used to extract package correlation features, predict package distribution, and optimize logistics plans, improving prediction accuracy and scheduling efficiency. 3. Based on the capacity spatiotemporal map, candidate routes and multi-dimensional routing features for express shipments are calculated, and these are input into a prediction model to obtain prediction results of the express shipment arrival status, which are then used to optimize urban logistics scheduling.
[0037] While the aforementioned existing technologies can all perform basic logistics forecasting, they still have significant technical shortcomings in practical applications, making them difficult to adapt to complex and dynamic logistics scenarios. Firstly, existing solutions, relying on fixed nodes or fixed graph structures for modeling, struggle to adapt to dynamically generated and disappearing active flow units. Secondly, when dealing with long-tailed order demand data, the common use of a single forecasting mechanism for unified modeling easily leads to mean regression bias, resulting in insufficient forecasting of high-demand nodes. Thirdly, performing global modeling on large-scale logistics networks incurs high computational complexity, making it difficult to meet the requirements of real-time scheduling and online deployment.
[0038] Based on this, this application provides a method for predicting the flow of logistics nodes. It can obtain the historical features corresponding to multiple real logistics nodes, map each logistics node to a preset set of multiple implicit representation nodes based on the historical features, where the number of implicit representation nodes is less than the number of logistics nodes, determine the change range of each implicit representation node in the future based on the features of each implicit representation node, and perform reverse mapping on the change range of each implicit representation node in the future to determine the flow prediction result of each logistics node in the future.
[0039] This solution acquires historical features of multiple real logistics nodes and maps the dynamically changing and variable number of physical logistics nodes to a smaller and fixed number of pre-defined implicit representation nodes. This completely decouples the modeling from the scale and topology of the original logistics graph, effectively adapting to the dynamic generation and disappearance of active flow units in the logistics network. This enables inductive generalization of unseen subgraphs in dynamic logistics networks. Furthermore, this latent space modeling approach abandons the traditional model of using a uniform prediction mechanism for all long-tail demand nodes. It can accurately infer the future change magnitude of each implicit representation node based on its characteristics, avoiding mean regression bias and significantly improving the prediction accuracy of high-demand nodes under long-tail distributions. Furthermore, by compressing high-dimensional, large-scale physical node operations into a low-dimensional, fixed-number implicit node space for time-series extrapolation, the computational volume and resource consumption are significantly reduced, effectively lowering the computational complexity in large-scale graph prediction tasks. Finally, by mapping the evolutionary changes of implicit nodes back to the physical node space, the traffic prediction results for each logistics node in the future are accurately output, comprehensively improving the generalization, accuracy, and efficiency of dynamic logistics network prediction, thereby better meeting the online deployment needs in scenarios such as instant delivery, shared mobility, and urban logistics scheduling.
[0040] The following is combined with Figures 1 to 9 This application describes the flow prediction method for logistics nodes.
[0041] Figure 1 This is a flowchart illustrating the logistics node traffic prediction method provided in this application, applied to an electronic device, which can be a prediction server for an urban instant delivery platform. Figure 1 As shown, the method includes the following steps S101-S104.
[0042] S101, obtain historical features corresponding to multiple real logistics nodes.
[0043] S102, based on historical characteristics, maps each logistics node to multiple preset implicit representation nodes.
[0044] The number of implicit representation nodes is less than the number of logistics nodes.
[0045] S103, based on the characteristics of each implicit representation node, determine the magnitude of change of each implicit representation node in future time periods.
[0046] S104, reverse map the change magnitude of each implicit representation node in the future period to determine the flow prediction result of each logistics node in the future period.
[0047] Multiple real logistics nodes can be N active logistics nodes dynamically sampled in the current window. N is an integer greater than 1.
[0048] Historical characteristics refer to the past The historical feature tensor of each time step.
[0049] In one example, each real logistics node in the past The historical feature tensor at each time step can be described as... Among them, feature channel C can include normalized historical throughput and multi-dimensional time identifiers (such as time attribute information such as morning peak, evening peak, weekend, weekday, and holiday).
[0050] In one example, to uniformly represent the origin and destination flow attributes of real logistics nodes, each logistics node can be represented by an Active Flow Unit (AFU). Specifically, an AFU can abstract a logistics node as a spatiotemporal interaction vector with a clear direction, connecting a specific starting task region and an ending task region, and recording a non-zero effective interaction frequency within a given discrete time window. Unlike traditional traffic sensors with fixed physical locations, the identification of an AFU is entirely determined by the origin and destination flow attributes it carries. In one example, the set of all potential AFUs in a city system can be denoted as the global flow set. .
[0051] In one example, a spatiotemporal prediction model (also known as a logistics prediction model) can be deployed in an electronic device. This model can be used to execute the scheme to determine the traffic forecast results for each logistics node in future time periods.
[0052] Specifically, the optimization objective of the spatiotemporal prediction model is to learn a parameterized prediction function. Output the future Determining the flow prediction results at each time step The model expression for the spatiotemporal prediction model can be: in, Indicates the logistics node in the past Historical feature tensor of each time step This is a spatiotemporal subgraph (also known as an induced spatiotemporal subgraph) of N active logistics nodes obtained by dynamic sampling in the current window. It represents the general spatiotemporal evolution law learned by the spatiotemporal prediction model, which is unaffected by individual nodes.
[0053] It should be noted that, due to the high degree of openness of the physical system, the spatiotemporal prediction model can be directly applied to sets containing entirely new nodes during the testing phase. Unseen subgraph This enables high-precision zero-shot inference.
[0054] Spacetime Subgraph . This refers to the dynamically sampled subset of logistics nodes at the current time step (referred to as the node subset). . For edge set, This is a weight matrix used to represent the strength of the relationship between various logistics nodes. The value can be determined based on the overlap of logistics nodes in the spatial task area and the alignment of the time window. Edge set and weight matrix A multi-relationship affinity matrix can be constructed. Although each sampling yields... The size and topology of the system exhibit significant randomness, but its internal structural dependencies are deterministic within a single inference instance.
[0055] In one example, the above method of mapping each logistics node to multiple preset implicit representation nodes (also known as latent variable nodes) based on historical characteristics can be implemented as follows: Figure 2 As shown in S201-S204, this method can be understood to be implemented using the aforementioned spatiotemporal prediction model.
[0056] S201 processes the historical features of each logistics node through a temporal convolutional layer to obtain the temporal features of the logistics node.
[0057] S202 uses a graph convolutional network to fuse the temporal features of each logistics node and the features of its neighboring nodes to obtain the node spatial fusion features.
[0058] Temporal convolutional layers can be one-dimensional, and the kernel size and stride of the temporal convolutional layer can be set to the temporal block size. .
[0059] Graph Convolutional Networks (GCNs) can be lightweight.
[0060] Specifically, since the historical feature tensor of each logistics node may contain high-frequency observation noise, a preliminary aggregation of short sequences of the historical feature tensors of each logistics node can be performed using a one-dimensional temporal convolutional layer. This operation not only smooths out temporal fluctuations but also effectively shortens the sequence length. Subsequently, a lightweight GCN can be used to aggregate the spatiotemporal subgraphs. The local spatial affinity feature within the node is obtained by fusing the temporal features of each logistics node and the features of the adjacent nodes of the logistics node through a lightweight GCN.
[0061] For example, S201-S202 can be implemented by the following formula: in, For node space fusion features, For the hidden layer feature dimension, , where is the length of the partitioned sequence. This encoding process characterizes the spatiotemporal dynamics of each logistics node within its micro-topological neighborhood.
[0062] S203, determine the allocation matrix corresponding to each logistics node.
[0063] The allocation matrix (also known as the sparse allocation matrix) is used to characterize the probability that each logistics node belongs to a different implicit representation node.
[0064] The implicit representation nodes are predefined, and the number of implicit representation nodes is fixed. The number of implicit representation nodes, K... The number of logistics nodes, N. In one example, the number of implicitly represented nodes, K, can be 100.
[0065] In one example, an allocation matrix can be generated based on node space fusion features and volume features.
[0066] Among them, the volume characteristic (also known as the static scale characteristic) is used to characterize the volume level of the historical flow of each logistics node.
[0067] Based on the node spatial fusion features and volume features, the generation of the allocation matrix can be specifically implemented as follows: the node spatial fusion features and volume features are concatenated, and the concatenated result is subjected to a nonlinear transformation to obtain the matching degree between each logistics node and each implicit representation node. Based on the matching degree between each logistics node and each implicit representation node, the allocation matrix is generated.
[0068] In one example, to ensure clear semantic discernibility and filter out low-relevance noise interference during the determination of the allocation matrix, this scheme can introduce a sparse masking strategy based on Top-k truncation. Specifically, after obtaining the matching degree between each logistics node and each implicit representation node, Top-k filtering can be performed on the matching degree between each logistics node and each implicit representation node. The results after Top-k filtering are then normalized to generate the allocation matrix.
[0069] For example, the matching degree between each logistics node and each implicit representation node can be obtained first using the following formula.
[0070] in, The relevance score can be used to characterize the matching degree between logistics nodes and implicit representation nodes. The higher the relevance score, the higher the matching degree between logistics nodes and implicit representation nodes. As a physical characteristic, This refers to the spatial fusion features of nodes (also known as local coding results). This indicates that the spatial fusion features of nodes are combined with the volume features. This represents the number of implicit representation nodes. This indicates the number of logistics nodes.
[0071] After obtaining the matching degree between each logistics node and each implicit representation node using the above method, the matching degree between each logistics node and each implicit representation node can be Top-k filtered using the following formula.
[0072] in, This represents the relevance score (or matching degree) of the i-th logistics node belonging to the k-th implicit representation node.
[0073] After Top-k filtering is completed, the results of Top-k filtering can be processed using the following formula. Perform a normalization operation to obtain the allocation weights of the i-th logistics node to the k-th implicit representation node. By combining the allocation weights corresponding to all logistics nodes, the final allocation matrix S can be formed, and .
[0074] S204, based on the allocation matrix, performs weighted aggregation of the node space fusion features to obtain the features of each implicit representation node.
[0075] After obtaining the assignment matrix S, the features of each implicit representation node, i.e., the latent space representation, can be obtained through the weighted projection operation of matrix multiplication using the following formula. ,and .
[0076] The above technical solution differs from traditional static graph clustering algorithms (such as METIS based on node connectivity). This solution introduces an allocation matrix representing the probability that each logistics node belongs to a different implicit representation node. Multiple logistics nodes are assigned to a K-dimensional latent space, i.e., K implicit representation nodes. Since the number of implicit representation nodes is pre-limited, the computational scale of the subsequent modeling process no longer increases linearly or quadratically with the number of physical nodes, thus eliminating the model architecture's dependence on the number N of input logistics nodes. Furthermore, the allocation matrix is dynamically generated entirely from the features of the logistics nodes. Simultaneously, a Top-k truncation-based sparse masking strategy is introduced to ensure that each logistics node contributes only to a small number of implicit representation nodes that are semantically closest to it. Compared to non-differentiable hard clustering algorithms, this sparse attention mechanism maintains the sparsity constraint of the mapping while supporting end-to-end gradient backpropagation. In addition, volume features are introduced. It can guide the model to clearly distinguish between the dominant head nodes and the low-frequency tail nodes when calculating and allocating weights, thus avoiding the dilution of latent space features by a large number of trivial tail features.
[0077] Moreover, the aforementioned weighted projection operation based on matrix multiplication mathematically constitutes a topological bottleneck. This bottleneck mechanism forces the system to filter out the specific random noise of local nodes, extracting a flow prototype with macroscopic universality. More importantly, the number of nodes K is implicitly represented by setting parameters. The number of logistics nodes N and the subsequent temporal evolution calculation dimensions can be strictly defined, giving the model the ability to perform elastic inference on networks with extremely high volatility.
[0078] Standard transductive graph neural networks heavily rely on a fixed graph Laplacian matrix, making knowledge transfer impossible across sampled subgraphs of varying sizes. If the graph structure is directly abandoned in favor of a global attention mechanism to handle dynamically appearing N physical nodes, the problem arises that the number of nodes increases quadratically. The computational complexity of physical graphs greatly limits their potential application in large-scale online systems. The inductive graph quantization mechanism proposed in this scheme maps a variable-length physical graph into a fixed number of implicit representation nodes using a differentiable allocation matrix, thus achieving complete decoupling between computational overhead and the size of the physical graph.
[0079] In one example, the above method, based on the characteristics of each implicit representation node, determines the magnitude of change of each implicit representation node in future time periods. Specifically, this can be implemented as follows: Figure 3 S301-S303 are shown.
[0080] S301: Input the features of each implicit representation node into the encoder to extract the temporal features of each implicit representation node.
[0081] The encoder can be a Transformer encoder.
[0082] In one example, the temporal features of each implicit representation node can be obtained using the following formula. .
[0083] S302 concatenates the temporal features and mask label embedding vectors of each implicit representation node.
[0084] Mask label embedding vector It can also be called a mask label embedding matrix or a mask label embedding vector. It carries time sequence information.
[0085] S303 uses a decoder to perform multi-head self-attention calculation on the spliced result to obtain the change magnitude of each implicit representation node in the future time period.
[0086] The decoder can be a Transformer decoder.
[0087] In one example, the above S302-S303 can be achieved by the following formula.
[0088] in, Represents the temporal characteristics of each implicit representation node. and mask label embedding vector Then, the parts are assembled. This indicates the magnitude of change of each implicit representation node in the future. It contains the global macroscopic dynamic evolution law deduced by the model at the implicit semantic level.
[0089] Understandably, thanks to the previous quantization operations, the computational space of global attention is confined to a fixed-capacity K-dimensional space, thus reducing its time and space complexity. a precipitous drop to .
[0090] In the above technical solution, after projecting the variable and discrete physical graph data into continuous and dense implicit representation nodes, a self-attention-based Transformer architecture is introduced to capture higher-order semantic interactions. Compared to local message passing on a restricted physical adjacency graph, the above technical solution can spontaneously uncover long-range global dependencies that are not limited by geographical connectivity, such as implicit collaboration between two residential areas that are far apart but exhibit synchronized morning rush hour patterns.
[0091] In one example, the above-mentioned reverse mapping of the change magnitude of each implicit representation node in future time periods determines the flow prediction result of each logistics node in future time periods. Specifically, this can be implemented as follows: Figure 4 S401-S402 are shown.
[0092] S401, based on the allocation matrix corresponding to each logistics node, reverse the change range of each implicit representation node in the future time period to obtain the change range of each logistics node in the future time period.
[0093] S402, based on the change range of each logistics node in the future period and the node spatial fusion characteristics of the logistics nodes, determine the traffic prediction result of each logistics node in the future period.
[0094] In one example, residual fusion can be performed on the change magnitude of each logistics node in the future time period and the node spatial fusion characteristics of the logistics node to obtain fusion characteristics. Based on the volume characteristics and fusion characteristics, corresponding prediction experts are assigned to logistics nodes with different traffic volumes from the expert hybrid prediction network. Finally, the traffic prediction results of each logistics node in the future time period are determined by the prediction experts corresponding to each logistics node.
[0095] In this context, the expert hybrid prediction network comprises multiple prediction experts used to fit the characteristics of logistics nodes with different flow rates. For example, the expert hybrid prediction network can be a set of experts consisting of M independent proprietary feedforward neural networks. Each feedforward neural network acts as a prediction expert. In one example, M can take the value 6.
[0096] It is understandable that, without any explicit human annotation intervention, each prediction expert can adaptively tend to fit the flow dynamics of logistics nodes in the corresponding flow range during the training process.
[0097] To preserve the macro-evolutionary trend while restoring fine-grained node specificity, this scheme utilizes the allocation matrix S to inversely project the implicit representation of node changes over future periods back into the physical domain, i.e., back into each logistics node. Considering that the projection process inevitably leads to the attenuation of local high-frequency information, the following formula can be used to fuse the future change range of each logistics node with the node space characteristics of the logistics nodes. Perform residual fusion to obtain fusion features This scheme enables the reconstructed physical representation (i.e., fused features) to be integrated. It integrates the global collaborative context from the latent space and effectively compensates for the unique local micro-patterns of each logistics node.
[0098] After obtaining the fusion features through the above method, based on the fusion features and volume features, the weight vector of each logistics node belonging to each prediction expert in the expert hybrid prediction network is determined using the following formula. : in, This indicates the number of forecasting experts.
[0099] Obtain the weight vector of each logistics node belonging to each prediction expert in the expert hybrid prediction network. Then, the following formula can be used to calculate the weight vector of each prediction expert in the expert hybrid prediction network to which each logistics node belongs. This allows us to obtain traffic forecasts for each logistics node in the future.
[0100] in, This indicates the traffic forecast results for each logistics node in the future period. This represents the weight by which a logistics node is assigned to the m-th prediction expert.
[0101] In one example, a Mixture of Experts (MoE) gating network can be deployed in the logistics prediction network (or electronic device). The gating network is used to assign corresponding prediction experts to logistics nodes with different traffic volumes.
[0102] Zero-value inflation under long-tailed distributions is another core pain point restricting the accuracy of logistics forecasting. When dealing with spatiotemporal data with extreme heteroscedasticity, traditional single fully connected regression heads often produce significant mean regression bias under the gradient pull of massive zero-value or low-value samples, thus severely underestimating high-value, instantaneous bursts of demand. To completely decouple the forecasting logic of extreme high-frequency traffic from daily stable traffic, the above-mentioned technical solution can assign corresponding forecasting experts to logistics nodes with different traffic volumes based on the size characteristics of the logistics nodes, realizing a size-adaptive expert hybrid network reconstruction mechanism. At the same time, the model can automatically identify high-intensity potential demand burst patterns and accurately guide them to an expert network specifically for handling large-value regression, eradicating the mean regression problem caused by single parameter sharing at the architectural level.
[0103] In one example, to ensure that the spatiotemporal prediction model not only has high-precision alignment capability at the micro-time step, but also further corrects the fitting bias caused by long-tail data, this scheme can also introduce a flow volume hierarchical reweighting strategy based on the mean square error. Specifically, the actual flow of each logistics node in the future period can be obtained, an adaptive weight function can be constructed, the adaptive weight function is positively correlated with the actual flow, and the spatiotemporal prediction model can be optimized based on the actual flow and the flow prediction results through the adaptive weight function.
[0104] For example, the basic regression loss function can be defined by the following formula, and the spatiotemporal prediction model can be optimized by the basic regression loss function.
[0105] in, Indicates the basic regression loss. This represents the adaptive weighting function. Representing the actual traffic, the adaptive weight function is positively correlated with the actual traffic, which can amplify the penalty for high-throughput samples during gradient backpropagation. This represents the traffic forecast result, where i represents the i-th logistics node.
[0106] In one example, to accelerate the convergence of the MoE gated network routing strategy, this scheme can also optimize the gated network. Specifically, after obtaining the actual traffic of each logistics node in the future time period, the gated network can be optimized based on the prediction experts corresponding to the actual traffic and by assigning corresponding prediction experts to logistics nodes with different traffic volumes.
[0107] For example, the cross-entropy supervised loss function can be defined by the following formula, and the gating network can be optimized using the cross-entropy supervised loss function.
[0108] in, Represents the cross-entropy supervision loss. Let be the flow interval corresponding to the flow of the i-th logistics node at time t. This represents the probability that the gating network selects the m-th prediction expert. This indicates whether the flow interval corresponding to the flow of the i-th logistics node at time t matches the flow interval corresponding to the m-th prediction expert.
[0109] In one example, the overall sequence optimization objective of the logistics forecasting model in this application is... .
[0110] In one example, the aforementioned electronic device can communicate with the scheduling system. After obtaining the traffic forecast results for each logistics node in the future time period, it can send the traffic forecast results for each logistics node in the future time period to the scheduling system. The scheduling system can then perform pre-deployment of transportation capacity, rider scheduling, peak resource scheduling, or route planning based on the traffic forecast results.
[0111] As can be seen from the above embodiments, this solution provides a graph-latent inductive dynamics estimation for dynamic logistics network (GLIDE) method. By acquiring the historical characteristics of multiple real logistics nodes, it maps the dynamically changing and variable number of physical logistics nodes to a smaller and fixed number of pre-defined implicit representation nodes. This completely decouples the modeling from the size and topology of the original logistics graph, effectively adapting to the dynamic generation and disappearance of active flow units in the logistics network. This enables inductive generalization of unseen subgraphs in dynamic logistics networks. Furthermore, this latent space modeling method abandons the traditional approach of using a uniform prediction mechanism for all long-tail demand nodes. It can accurately infer the future change magnitude of each implicit representation node based on its characteristics, avoiding mean regression bias and significantly improving the prediction accuracy of high-demand nodes under long-tail distribution. Furthermore, by compressing high-dimensional, large-scale physical node operations into a low-dimensional, fixed-number implicit node space for time-series extrapolation, the computational volume and resource consumption are significantly reduced, effectively lowering the computational complexity in large-scale graph prediction tasks. Finally, by mapping the evolutionary changes of implicit nodes back to the physical node space, the traffic prediction results for each logistics node in the future are accurately output, comprehensively improving the generalization, accuracy, and efficiency of dynamic logistics network prediction, thereby better meeting the online deployment needs in scenarios such as instant delivery, shared mobility, and urban logistics scheduling.
[0112] Figure 5 A schematic diagram of a flow prediction method for a logistics node provided in an embodiment of this application is shown below. Figure 5 As shown, this scheme can include three parts: inductive graph quantization mechanism, latent space global sequence modeling, and volume adaptive reconstruction.
[0113] See also Figure 5 As shown, in the inductive graph quantization mechanism, a local encoder (including a temporal convolutional layer and a graph convolutional network) can be used to process the historical feature tensor of each logistics node. and spacetime subgraph The process is performed to obtain the node spatial fusion features. Node spatial fusion features through MLP and size characteristics The process yields an allocation matrix S. Then, based on the allocation matrix S, each logistics node is mapped to a set of predefined implicit representation nodes, resulting in a latent space representation. .
[0114] In the latent space global sequence modeling part, latent space representation can be used... The Transformer encoder obtains the temporal features of each implicit representation node. The Transformer decoder then obtains the magnitude of change of each implicit representation node in future time periods. .
[0115] In the volume adaptive reconstruction part, the allocation matrix S can be used to project the implicit representation of the node's change magnitude in future time periods back into the physical domain, and to fuse the node spatial characteristics of each logistics node with the change magnitude in future time periods. Perform residual fusion to obtain fusion features After that, it can be done through Based on fusion characteristics and volume characteristics, a forecasting expert obtained the traffic forecast results for each logistics node in the future period. .
[0116] To verify the effectiveness and superiority of the proposed spatiotemporal prediction model in dynamic urban traffic flow prediction, a comprehensive comparative experiment was conducted. The experimental design, implementation details, and results analysis are as follows: I. Experimental Dataset This experiment uses six real large-scale industrial datasets, covering three typical cities (City A, City B, and City C). To capture the heteroscedasticity of urban human activities, the datasets are further divided into weekday (Wd) and weekend (Weekend) subsets. Detailed statistics of the datasets are shown in Table 1 below. Due to the dynamic birth and death of AFUs, the network generates dynamic subgraphs of varying sizes. To enhance the model's inductive generalization ability and simulate incomplete observation scenarios, two spatial sampling strategies are adopted during the training phase: (1) Uniform structure sampling: randomly sampling 20% to 80% of the nodes in the subgraph; (2) Flow prior sampling: preferentially sampling the head nodes whose single-volume is in the top 20% to 80%. The huge fluctuations in the size of the subgraph nodes highlight the inductive inference challenges faced by the direct-push baseline model.
[0117] Figure 6 This application provides a schematic diagram of urban characteristic distribution. Figure 6Figure 6(a) shows the dramatic dynamic fluctuations in the number of active flow unit nodes in cities A, B, and C over a given date. Figure 6(b) shows the changes in the logarithmic frequency of cities A, B, and C with the flow of active flow unit nodes. The logarithmic frequency distribution exhibits typical long-tail and zero-value expansion characteristics of business flow. Figure 6 (a) and (b) in the figure intuitively reflect the dynamic generation and extinction characteristics of active flow units and the high degree of unevenness of flow distribution in the dynamic logistics spatiotemporal system, further illustrating the difficulty of adapting the traditional direct-push spatiotemporal model in this scenario.
[0118] Table 1
[0119] II. Comparison with baseline models Five mainstream baseline models were selected for comparison in the experiment: (1) Statistical methods: Historical Average (HA). (2) Temporal models based on Multilayer Perceptron (MLP): Spatio-Temporal Identity Decay (STID) and Decomposed Linear Model (DLinear). (3) Temporal models based on Transformer: Informer (efficient long sequence Transformer model), Autoformer (autocorrelation decomposition Transformer model), and iTransformer (inverted Transformer model). (4) Traditional spatio-temporal graph models: Spatio-Temporal Graph Convolutional Networks (STGCN) and Dynamic Spatio-Temporal Graph Neural Network (DSTSGNN). (5) Inductive basic model: Unified Flow Forecasting Model (UniFlow). This model supports cross-graph training and adopts the same parameter sharing mechanism as the spatiotemporal prediction model in this scheme to ensure the fairness of the comparison.
[0120] III. Experiment Implementation Details All models are implemented using the PyTorch deep learning framework and trained and tested on a computing device equipped with an NVIDIA Tesla V100 (32GB) Graphics Processing Unit (GPU). Both the encoder and decoder in this scheme employ a 2-layer network structure, with the hidden layer feature dimension D set to 64 and the number of implicit representation nodes K set to 100. The Adam optimizer is used uniformly for model training, with an initial learning rate set to 10. -4 The weight decay term is set to 10. -5 The maximum number of training epochs was set to 200, and an early stopping strategy with a tolerance of 10 was implemented to prevent overfitting. To ensure fairness in the comparison, all baseline models were within {10}. -2 10 -3 10 -4 A grid search was performed on the candidate set to select the optimal learning rate configuration.
[0121] IV. Experimental Results and Analysis Experimental results are as follows Figure 7 As shown, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are used as the core evaluation metrics; lower metric values indicate higher model prediction accuracy. Comparative results show that the proposed scheme achieves optimal prediction performance on four key test subsets, specifically as follows: 1. In a weekday scenario (stable dynamic), this solution significantly outperforms all baseline models in terms of MAE and RMSE in three city subsets, achieving accurate modeling of stable traffic patterns in cities.
[0122] 2. In the weekend scenario (sudden dynamics), this solution achieves the best results on two city subsets, and its performance on the remaining subset is also significantly better than most baseline models, demonstrating strong robustness to sudden traffic changes.
[0123] 3. Compared with the results of the optimal baseline model, this scheme has achieved a stable decrease in prediction error on multiple datasets, and the overall performance has been significantly improved in terms of inductive generalization ability and dynamic scene adaptability.
[0124] In summary, this solution, through targeted architectural design, effectively solves the inductive reasoning problem of existing direct-propagation spatiotemporal models on large-scale dynamic subgraphs. It demonstrates superior prediction accuracy and generalization ability compared to existing technologies in both stable and bursty traffic scenarios, verifying the advanced nature and industrial application value of the technical solution of this invention.
[0125] The flow prediction device for logistics nodes provided in this application is described below. The flow prediction device for logistics nodes described below can be referred to in correspondence with the flow prediction method for logistics nodes described above.
[0126] Figure 8 This is a schematic diagram of the structure of the flow prediction device for a logistics node provided in an embodiment of this application, with reference to... Figure 8 As shown, the flow prediction device for logistics nodes includes: The acquisition module 801 is used to acquire historical features corresponding to multiple real logistics nodes; The mapping module 802 is used to map each of the logistics nodes to a preset plurality of implicit representation nodes based on the historical features; the number of implicit representation nodes is less than the number of logistics nodes. The determination module 803 is used to determine the magnitude of change of each implicit representation node in a future time period based on the characteristics of each implicit representation node. The reverse mapping module 804 is used to reverse map the change magnitude of each of the implicit representation nodes in the future time period to determine the flow prediction result of each of the logistics nodes in the future time period.
[0127] Figure 9 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 9 As shown, the electronic device may include a processor 910, a communications interface 920, a memory 930, and a communication bus 940. The processor 910, communications interface 920, and memory 930 communicate with each other via the communication bus 940. The processor 910 can call logical instructions in the memory 930 to execute a flow prediction method for logistics nodes. This method includes: acquiring historical features corresponding to multiple real logistics nodes; mapping each logistics node to multiple preset implicit representation nodes based on the historical features (the number of implicit representation nodes is less than the number of logistics nodes); determining the change range of each implicit representation node in the future time period based on the features of each implicit representation node; performing a reverse mapping on the change range of each implicit representation node in the future time period; and determining the flow prediction result for each logistics node in the future time period.
[0128] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0129] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the flow prediction method for logistics nodes provided by the above methods. The method includes: acquiring historical features corresponding to multiple real logistics nodes; mapping each logistics node to a preset number of implicit representation nodes based on the historical features, wherein the number of implicit representation nodes is less than the number of logistics nodes; determining the change range of each implicit representation node in the future time period based on the features of each implicit representation node; performing a reverse mapping on the change range of each implicit representation node in the future time period; and determining the flow prediction result of each logistics node in the future time period.
[0130] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program is implemented to perform the flow prediction method for logistics nodes provided by the above methods. The method includes: acquiring historical features corresponding to multiple real logistics nodes; mapping each logistics node to a preset set of multiple implicit representation nodes based on the historical features, wherein the number of implicit representation nodes is less than the number of logistics nodes; determining the change range of each implicit representation node in the future time period based on the features of each implicit representation node; performing a reverse mapping on the change range of each implicit representation node in the future time period; and determining the flow prediction result of each logistics node in the future time period.
[0131] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0132] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for predicting the flow of goods at a logistics node, characterized in that, Applied to electronic devices, the method includes: Obtain historical features corresponding to multiple real logistics nodes; Based on the historical characteristics, each logistics node is mapped to a preset set of implicit representation nodes; the number of implicit representation nodes is less than the number of logistics nodes. Based on the characteristics of each implicit representation node, the magnitude of change of each implicit representation node in future time periods is determined; By performing a reverse mapping on the change magnitude of each implicit representation node in the future time period, the flow prediction result of each logistics node in the future time period is determined.
2. The method according to claim 1, characterized in that, The step of mapping each logistics node to a preset set of implicit representation nodes based on the historical characteristics includes: The historical features of each logistics node are processed by a temporal convolutional layer to obtain the temporal features of the logistics node. By fusing the temporal features of each logistics node and the features of its neighboring nodes through a graph convolutional network, a node spatial fusion feature is obtained. Determine the allocation matrix corresponding to each of the logistics nodes; the allocation matrix is used to characterize the probability that each of the logistics nodes belongs to a different implicit representation node; Based on the allocation matrix, the node spatial fusion features are weighted and aggregated to obtain the features of each implicit representation node.
3. The method according to claim 2, characterized in that, The determination of the allocation matrix corresponding to each logistics node includes: Based on the node spatial fusion features and volume features, the allocation matrix is generated, whereby the volume features are used to characterize the volume level of the historical flow of each logistics node.
4. The method according to claim 3, characterized in that, The process of generating the allocation matrix based on the node spatial fusion features and volume features includes: The node spatial fusion feature is then combined with the volume feature; A nonlinear transformation is performed on the spliced result to obtain the matching degree between each logistics node and each implicit representation node; The allocation matrix is generated based on the matching degree between each logistics node and each implicit representation node.
5. The method according to claim 4, characterized in that, The process of generating the allocation matrix based on the matching degree between each logistics node and each implicit representation node includes: Top-k screening is performed on the matching degree between each logistics node and each implicit representation node; The results after Top-k filtering are normalized to generate the allocation matrix.
6. The method according to any one of claims 1-5, characterized in that, The step of determining the magnitude of change of each implicit representation node in future time periods based on the characteristics of each implicit representation node includes: The features of each implicit representation node are input into the encoder to extract the temporal features of each implicit representation node; The temporal features of each implicit representation node and the mask label embedding vector are concatenated; the mask label embedding vector carries temporal information. The decoder performs multi-head self-attention calculation on the spliced result to obtain the change magnitude of each implicit representation node in the future time period.
7. The method according to any one of claims 1-5, characterized in that, The step of inversely mapping the change magnitude of each of the implicit representation nodes in future time periods to determine the flow prediction results of each of the logistics nodes in future time periods includes: Based on the allocation matrix corresponding to each logistics node, the change range of each implicit representation node in the future time period is reverse-mapped to obtain the change range of each logistics node in the future time period; the allocation matrix is used to represent the probability that each logistics node belongs to a different implicit representation node; Based on the magnitude of change of each logistics node in the future time period and the node spatial fusion characteristics of the logistics node, the traffic prediction result of each logistics node in the future time period is determined.
8. The method according to claim 7, characterized in that, The process of determining the traffic forecast result for each logistics node in the future time period based on the change range of each logistics node in the future time period and the node spatial fusion characteristics of the logistics node includes: Residual fusion is performed on the change range of each logistics node in future time periods and the node spatial fusion characteristics of the logistics node to obtain the fusion characteristics; Based on the volume characteristics and the fusion characteristics, corresponding prediction experts are assigned to the logistics nodes with different traffic volumes from the expert hybrid prediction network; the expert hybrid prediction network contains multiple prediction experts for fitting the characteristics of logistics nodes with different traffic volumes. The predicted flow forecast results for each logistics node in the future period are determined by the prediction experts corresponding to each logistics node.
9. The method according to claim 8, characterized in that, The electronic device is equipped with a gating network, which is used to assign corresponding prediction experts to the logistics nodes with different traffic volumes. The method further includes: Obtain the actual traffic flow of each of the aforementioned logistics nodes in future time periods; The gating network is optimized based on the prediction experts corresponding to the actual traffic flow and the prediction experts assigned to the logistics nodes with different traffic flows.
10. The method according to any one of claims 1-5, characterized in that, The electronic device is equipped with a spatiotemporal prediction model, which is used to determine the traffic prediction results of each logistics node in the future time period. The method further includes: Obtain the actual traffic flow of each of the aforementioned logistics nodes in future time periods; Construct an adaptive weight function; the adaptive weight function is positively correlated with the actual traffic. Based on the actual traffic flow and the traffic prediction results, the spatiotemporal prediction model is optimized using the adaptive weighting function.
11. A flow prediction device for a logistics node, characterized in that, The device includes: The acquisition module is used to acquire historical features corresponding to multiple real logistics nodes; A mapping module is used to map each of the logistics nodes to a preset plurality of implicit representation nodes based on the historical features; the number of implicit representation nodes is less than the number of logistics nodes. The determination module is used to determine the magnitude of change of each implicit representation node in future time periods based on the characteristics of each implicit representation node. The reverse mapping module is used to reverse map the change magnitude of each of the implicit representation nodes in the future time period to determine the flow prediction result of each of the logistics nodes in the future time period.
12. The apparatus according to claim 11, characterized in that, The mapping module is specifically used for: The historical features of each logistics node are processed by a temporal convolutional layer to obtain the temporal features of the logistics node. By fusing the temporal features of each logistics node and the features of its neighboring nodes through a graph convolutional network, a node spatial fusion feature is obtained. Determine the allocation matrix corresponding to each of the logistics nodes; the allocation matrix is used to characterize the probability that each of the logistics nodes belongs to a different implicit representation node; Based on the allocation matrix, the node spatial fusion features are weighted and aggregated to obtain the features of each implicit representation node.