A multi-vendor collaborative store digitization management system and method

Abstract

The present application belongs to the field of big data processing, and specifically relates to a multi-supplier cooperative store digitized management system and method. Multi-supplier historical delivery time sequence data and store warehouse space time-varying occupation data are collected; a dynamic space-time graph is constructed with supplier nodes and store physical storage location nodes as vertices and logistics time delay and supply-demand dependency as edges; a time sequence attention mechanism is embedded into a graph convolution network to extract space-time coupling features in a concurrent performance state, representing the dynamic mapping correlation between supplier delivery time sequence and store storage location release time sequence; a storage location allocation and delivery time sequence joint optimization model is constructed based on the space-time coupling features, with physical storage location capacity constraints and logistics distribution time sequence constraints as boundary conditions. Dynamic adaptation of cross-supplier logistics time sequence and store physical space is achieved, physical warehouse space occupation and blockage defects caused by concurrent warehousing are overcome, and inventory accumulation and shortage phenomena caused by supply-demand space-time mismatch are eliminated.

Description

Technical Field

[0001] This invention belongs to the field of big data processing, specifically a multi-supplier collaborative digital management system and method for stores. Background Technology

[0002] Existing digital management systems for stores typically employ centralized data aggregation and static threshold warning mechanisms when handling collaborative fulfillment with multiple suppliers. Each supplier's logistics and distribution data, along with store warehousing data, are stored independently on the server and processed sequentially according to document arrival order. The system allocates fixed physical storage areas and fixed delivery time windows for each supplier based on their historical shipment volume. When the system receives a shipment instruction, it compares the volume of the goods with the remaining space threshold of the corresponding fixed storage location. If the remaining space is greater than the goods volume, warehousing is allowed; otherwise, a queuing instruction is triggered. No correlation calculation link is established between the logistics time-series data of each supplier and the occupancy status data of the store storage locations. The occupancy and release of storage space are only compared and determined as static values, and concurrent warehousing requests from different suppliers are mechanically queued according to their timestamp order.

[0003] In the aforementioned existing technologies, each supplier's fulfillment data is processed independently and the warehouse space is statically allocated. This results in the inability to perceive and process the dynamic mapping relationship between the supplier's delivery sequence and the release sequence of the physical storage space in the store when multiple suppliers ship goods concurrently. This leads to conflicts in the grabbing of physical storage space in the store and the problem of supply and demand mismatch in time and space when multiple sources of goods arrive at the same time. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-supplier collaborative digital management system and method for stores, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multi-supplier collaborative digital management system for stores, comprising: The data acquisition component collects historical shipment time-series data from multiple suppliers and time-varying occupancy data of store storage space; The graph construction component, based on the historical delivery time series data and the time-varying occupancy data, constructs a dynamic spatiotemporal graph with supplier nodes and store physical storage location nodes as vertices and logistics delay and supply-demand dependency relationships as edges. The feature extraction component embeds a temporal attention mechanism into a graph convolutional network to process the dynamic spatiotemporal graph and extract spatiotemporal coupling features under the concurrent fulfillment state of multiple suppliers. The spatiotemporal coupling features characterize the dynamic mapping relationship between the supplier's delivery time sequence and the store's warehouse release time sequence. The optimization solution component is constructed based on the spatiotemporal coupling characteristics to build a joint optimization model for warehouse location allocation and delivery timing. The physical warehouse location capacity constraint and logistics delivery timing constraint are used as the boundary conditions of the joint optimization model for warehouse location allocation and delivery timing. The system solves and outputs the dynamic allocation vector of physical warehouse locations and the delivery timing scheduling instructions of suppliers.

[0006] Preferably, in the graph construction component, constructing a dynamic spatiotemporal graph using logistics delay and supply-demand dependency as edges includes: The historical average delivery time from different supplier nodes to physical warehouse nodes is calculated as the edge weight of the logistics delay. The difference between the frequency of shipments from supplier nodes to physical warehouse nodes and the frequency of returns within a preset time window is calculated as the edge weight of the supply and demand dependency relationship. The edge weights of the logistics delay and the edge weights of the supply and demand dependency relationship are concatenated to generate a dynamic edge feature matrix of a dynamic spatiotemporal graph. The dynamic edge feature matrix is ​​periodically updated as the time sliding window progresses. The vertices contain the product attribute feature vector of the supplier node and the spatial geometric attribute feature vector of the physical warehouse node.

[0007] Preferably, in the feature extraction component, embedding a temporal attention mechanism into a graph convolutional network to extract spatiotemporally coupled features includes: In the spatial feature aggregation layer of the graph convolutional network, the features of adjacent nodes are weighted and aggregated based on the dynamic edge feature matrix of the dynamic spatiotemporal graph to output spatial topological features. In the temporal feature extraction layer of the temporal attention mechanism, an attention weight matrix containing a time decay factor is constructed. The spatial topological features within consecutive time slices are weighted and summed. The time decay factor decreases exponentially as the time span increases. The weighted and summed temporal features are then matrix-multiplied with the spatial topological features of the current time slice to output a spatiotemporal coupling feature that represents the dynamic mapping relationship between the supplier's delivery time sequence and the store's warehouse space release time sequence.

[0008] Preferably, in the optimization solution component, constructing the joint optimization model for warehouse location allocation and delivery timing includes: The objective function is to minimize the sum of the idle time of physical storage locations in stores and the delivery time offset of suppliers. The physical storage location capacity constraints include the upper limit threshold of the three-dimensional spatial capacity of physical storage locations and the maximum inbound throughput threshold per unit time. The logistics delivery timing constraints include the latest delivery time of supplier nodes and the fixed time consumption of transportation routes. The physical storage location capacity constraints and the logistics delivery timing constraints are transformed into penalty terms of the objective function. The objective function with penalty terms is iteratively solved based on the alternating direction multiplier method, and the dynamic allocation vector of physical storage locations in stores and the supplier delivery timing scheduling instructions are output.

[0009] Preferably, before the optimization solution component outputs the supplier delivery timing scheduling instruction, it further includes a conflict resolution component: The concurrent conflict feature is extracted from the spatiotemporal coupling feature. The concurrent conflict feature represents that multiple supplier delivery time-series scheduling instructions point to the same physical warehouse node at the same time. The maximum allowable concurrent inbound volume is calculated based on the number of inbound channels and the concurrency of sorting processing of the same physical warehouse node. When the number of supplier delivery time-series scheduling instructions pointing to the same physical warehouse node at the same time exceeds the maximum allowable concurrent inbound volume, a delay offset of a preset time step is applied to the excess supplier delivery time-series scheduling instructions in ascending order of edge weights of the supply and demand dependency relationship of the supplier nodes until the concurrent inbound volume meets the constraint.

[0010] Preferably, the graph construction component further includes an incremental evolution component for constructing a dynamic spatiotemporal graph: The system monitors the data flow distribution drift status of the historical shipment time-series data and the time-varying occupancy data, calculates the KL divergence between the data flow characteristics in the current time window and the historical baseline data flow characteristics, and triggers incremental reconstruction of the dynamic spatiotemporal graph when the KL divergence exceeds a preset drift threshold. It locks the abnormal supplier nodes and abnormal store physical storage location nodes that cause data distribution drift in the current time window, and only recalculates and replaces the local edge weights and local node characteristics related to the abnormal supplier nodes and abnormal store physical storage location nodes, while keeping the rest of the topology and feature parameters of the dynamic spatiotemporal graph unchanged.

[0011] Preferably, the periodic update of the dynamic edge feature matrix includes: Data on abnormal events in the external logistics environment is acquired, and the impact range and duration of the abnormal events are extracted. Target edges between affected supplier nodes and physical warehouse location nodes are selected based on the impact range. Based on the duration and a preset environmental impact attenuation coefficient, the logistics delay increment weight and supply-demand dependence attenuation weight of the target edges are calculated. The logistics delay increment weight is added to the original logistics delay edge weight, and the supply-demand dependence attenuation weight is deducted from the original supply-demand dependence edge weight, generating a dynamic edge feature matrix under sudden environmental impact.

[0012] Preferably, the temporal feature extraction layer of the temporal attention mechanism includes: A multi-head temporal attention subspace for parallel computing is constructed. Each temporal attention subspace independently learns spatial topological feature dependencies at different time scales. The sub-temporal features output by each temporal attention subspace are concatenated and linearly projected to generate multi-scale fused temporal features. The spatial path topological distance between supplier nodes and store physical storage location nodes in the dynamic spatiotemporal graph is extracted. A spatial topological location encoding vector is generated based on the spatial path topological distance. The multi-scale fused temporal features and the spatial topological location encoding vector are added bit-by-bit to output spatiotemporal coupling features that strengthen spatial constraints.

[0013] Preferably, the penalty term of the objective function includes an adaptive penalty coefficient adjustment component: The system collects real-time data on the current occupancy rate of physical storage locations in stores and the current fulfillment urgency of supplier nodes. It then constructs a penalty balance function with these two data points as input. When the current occupancy rate increases, the penalty coefficient corresponding to the physical storage location capacity constraint is dynamically increased to suppress inbound allocation. When the current fulfillment urgency increases, the penalty coefficient corresponding to the logistics delivery timing constraint is dynamically decreased to relax delivery time restrictions. Based on the output value of the penalty balance function, the penalty coefficients corresponding to the physical storage location capacity constraint and the logistics delivery timing constraint are dynamically updated and iteratively solved using the alternating direction multiplier method.

[0014] A multi-supplier collaborative digital store management method includes: Collect historical shipment time-series data from multiple suppliers and time-varying occupancy data of store storage space; Based on the historical delivery time series data and the time-varying occupancy data, a dynamic spatiotemporal graph is constructed with supplier nodes and store physical storage location nodes as vertices and logistics delay and supply-demand dependency relationships as edges. The temporal attention mechanism is embedded into the graph convolutional network to process the dynamic spatiotemporal graph and extract the spatiotemporal coupling features under the concurrent fulfillment state of multiple suppliers. The spatiotemporal coupling features represent the dynamic mapping relationship between the supplier's delivery time sequence and the store's warehouse release time sequence. Based on the aforementioned spatiotemporal coupling characteristics, a joint optimization model for warehouse location allocation and delivery timing is constructed. The physical warehouse location capacity constraint and logistics delivery timing constraint are used as the boundary conditions of the joint optimization model for warehouse location allocation and delivery timing. The method for solving and outputting the dynamic allocation vector of physical warehouse locations and the supplier delivery timing scheduling instructions is then used.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a dynamic spatiotemporal graph with supplier nodes and store physical storage location nodes as vertices and logistics latency and supply-demand dependency relationships as edges. It embeds a temporal attention mechanism into a graph convolutional network to extract spatiotemporal coupling features, establishing a dynamic mapping relationship between supplier delivery timing and store storage location release timing. Based on these spatiotemporal coupling features, a joint optimization model for storage location allocation and delivery timing is constructed. The model uses store physical storage location capacity constraints and logistics delivery timing constraints as boundary conditions to solve for and output a dynamic allocation vector for store physical storage locations and supplier delivery timing scheduling instructions. This achieves dynamic adaptation between cross-supplier logistics timing and store physical space, overcoming the defects of physical storage space contention and congestion caused by concurrent warehousing from multiple suppliers, and eliminating inventory backlog and stockout phenomena caused by spatiotemporal mismatch between supply and demand.

[0016] 2. This invention constructs an attention weight matrix containing a time decay factor in the temporal attention mechanism to perform weighted summation of spatial topological features within continuous time slices, reflecting the decay law of the impact of historical occupancy on current inbound storage; it uses the alternating direction multiplier method to iteratively solve the objective function with store physical storage capacity constraints and logistics delivery time-series constraints, and applies delay offsets in ascending order of edge weights according to the supply and demand dependency relationship when the concurrent inbound volume exceeds the maximum allowed concurrent inbound volume, avoiding physical congestion of the inbound channel; it monitors the data flow distribution drift state, and when the probability distribution difference exceeds the preset drift threshold, it only recalculates and replaces the local edge weights and local node features related to abnormal nodes, reducing the computational resource consumption of dynamic spatiotemporal graph reconstruction and improving the system's robustness in high-concurrency interaction scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall execution process of the multi-supplier collaborative store digital management system of the present invention. Figure 2 This is a flowchart of the dynamic spatiotemporal graph construction and dynamic edge feature matrix generation of the present invention; Figure 3 This is a flowchart of the spatiotemporal coupling feature extraction process of the graph convolutional network with embedded temporal attention according to the present invention. Figure 4 This is a flowchart illustrating the construction and solution process of the joint optimization model for warehouse location allocation and delivery timing in this invention. Figure 5 This is a flowchart illustrating the concurrent data entry conflict resolution process of the present invention. Figure 6 This is a flowchart illustrating the dynamic spatiotemporal graph incremental evolution and adaptation to external environmental shocks in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please refer to Figure 1 This embodiment provides a multi-supplier collaborative store digital management system and method. The multi-supplier collaborative store digital management system acquires historical delivery time-series data from multiple suppliers and time-varying occupancy data of store storage space through a data acquisition component. The data acquisition component establishes data connections with the ERP systems, logistics management systems, and store WMS systems of each supplier through a standardized API interface, acquiring real-time and historical data according to a preset acquisition frequency. The acquisition frequency is set to once per minute. Real-time data includes current pending order information, in-transit transportation information, and real-time occupancy status of store storage locations; historical data spans the past 12 months and is used to construct an initial dynamic spatiotemporal diagram and model training. After acquiring the raw data, the data acquisition component performs data preprocessing operations, including missing value imputation, outlier removal, and data normalization. Missing value imputation uses linear interpolation. For continuous data such as cargo volume and occupancy time, interpolation calculations are performed based on values ​​from adjacent time points; for discrete data such as supplier ID and storage location ID, mode imputation is used. Outlier removal employs the 3σ principle, calculating the mean and standard deviation of each data field. Data exceeding the mean ± 3 times the standard deviation are marked as outliers and removed. Data normalization uses the min-max normalization method, mapping all continuous data to the [0,1] interval to eliminate the influence of different units on subsequent calculations.

[0020] Table 1 illustrates the specific meaning and format of the various types of data acquired by the data acquisition component. Table 1. Field Definitions for Historical Shipment Time-Series Data from Multiple Suppliers and Time-Varying Occupancy Data at Stores refer to Figure 2The data acquisition component transmits preprocessed historical shipment time-series data and time-varying occupancy data to the graph construction component. Based on this data, the graph construction component constructs a dynamic spatiotemporal graph. The vertex set of the dynamic spatiotemporal graph includes a set of supplier nodes and a set of store physical storage location nodes, while the edge set is the set of directed edges from supplier nodes to store physical storage location nodes. The initial feature vector of supplier nodes includes four dimensions: average product volume, average product weight, average product turnover rate, and historical average fulfillment rate. The initial feature vector of store physical storage location nodes includes four dimensions: three-dimensional storage location volume, maximum storage location load capacity, inbound aisle distance, and sorting processing rate. The initial edge weights include two dimensions: logistics latency weight and supply-demand dependency weight.

[0021] The logistics delay weight is calculated as follows: All historical delivery times from a supplier node to a physical warehouse location node within a preset time window are statistically analyzed, and their arithmetic mean is calculated as the logistics delay weight for that edge. The preset time window is the past 30 days, and the delivery time is the difference between the actual arrival timestamp and the dispatch timestamp. The formula for calculating the logistics delay weight is:

[0022] in, Indicates supplier node physical storage location nodes at stores The weight of logistics delay, This indicates the number of times the supplier node sends goods to the warehouse node within a preset time window. Indicates the first The actual arrival timestamp of the next shipment. Indicates the first The shipping timestamp for the next shipment.

[0023] The supply-demand dependency weight is calculated as follows: Within a preset time window, the frequency of shipments and returns from a supplier node to a physical warehouse location at a specific store is statistically analyzed, and the difference between the two is used as the supply-demand dependency weight for that edge. The shipment frequency is the number of shipments within the preset time window, and the return frequency is the number of returns within the preset time window. The formula for calculating the supply-demand dependency weight is:

[0024] in, Indicates supplier node physical storage location nodes at stores The weight of supply and demand dependence This indicates the frequency of shipments from the supplier node to the warehouse node within a preset time window. This indicates the frequency of returns from the supplier node to the warehouse node within a preset time window.

[0025] refer to Figure 3 The graph construction component transmits the constructed dynamic spatiotemporal graph to the feature extraction component. The feature extraction component embeds a temporal attention mechanism into the graph convolutional network to process the dynamic spatiotemporal graph and extract spatiotemporal coupling features under multi-supplier concurrent fulfillment conditions. The graph convolutional network contains two spatial feature aggregation layers. Each spatial feature aggregation layer weights and aggregates the features of adjacent nodes based on the edge weights of the dynamic spatiotemporal graph, outputting spatial topological features. The input of the first spatial feature aggregation layer is the initial feature matrix of the nodes in the dynamic spatiotemporal graph, and the output is a first-order spatial topological feature; the input of the second spatial feature aggregation layer is the first-order spatial topological feature, and the output is a second-order spatial topological feature.

[0026] The formula for calculating spatial feature aggregation is: in, Indicates the first The node feature matrix of the layer, Indicates the first The node feature matrix of the layer, To add a self-loop adjacency matrix, The adjacency matrix of the dynamic spatiotemporal graph. It is the identity matrix. for The degree matrix, , For the first The trainable weight matrix of the layer, This is the ReLU activation function.

[0027] The temporal attention mechanism includes a temporal feature extraction layer, which receives continuous... The spatial topological features of each time slice are used to construct an attention weight matrix that includes a time decay factor. The spatial topological features within consecutive time slices are then summed using weighted averages. The time decay factor decreases exponentially with increasing time span, reflecting the gradual weakening of the influence of historical occupancy status on current warehousing decisions over time.

[0028] The formula for calculating the time decay factor is: in, Indicates the time decay factor. This indicates the time span between the historical time slice and the current time slice. is the attenuation coefficient, with a value range of [0.01, 0.1].

[0029] The formula for calculating the attention weight matrix is: in, Indicates the current time slice Attention weight matrix, This represents the spatial topological feature matrix of the current time slice. Indicates continuity Spatial topological feature matrix of each time slice To query the vector weight matrix, The key vector weight matrix, Let be the dimension of the key vector. For time decay mask matrix, Representing time slices Relative to time slice Time decay factor, This indicates a bitwise multiplication operation. It is a normalized exponential function.

[0030] The time feature extraction layer performs matrix multiplication of the weighted summed time features with the spatial topological features of the current time slice, outputting a spatiotemporal coupling feature. The spatiotemporal coupling feature represents the dynamic mapping relationship between the supplier's delivery time sequence and the store's warehouse release time sequence, and its dimension is the sum of the number of supplier nodes and the number of warehouse nodes multiplied by the feature dimension.

[0031] The formula for calculating the spatiotemporal coupling characteristics is: in, Represents the spatiotemporal coupling characteristic matrix. The value vector weight matrix, This represents the time feature matrix after weighted summation. This represents matrix multiplication.

[0032] refer to Figure 4 The feature extraction component transmits the extracted spatiotemporal coupled features to the optimization and solution component. Based on the spatiotemporal coupled features, the optimization and solution component constructs a joint optimization model for warehouse location allocation and delivery timing, using the physical warehouse location capacity constraints and logistics delivery timing constraints as boundary conditions for the joint optimization model, and solves to output the dynamic allocation vector of physical warehouse locations and the supplier delivery timing scheduling instructions.

[0033] The objective function of the joint optimization model is to minimize the sum of the store's physical storage space idle time and the supplier's delivery time offset. The store's physical storage space idle time is the difference between the actual release time of the storage space and the actual arrival time of the next batch of goods. If the arrival time of the next batch of goods is earlier than the storage space release time, the idle time is negative, indicating that the storage space has been preempted. The supplier's delivery time offset is the absolute value of the difference between the actual delivery time and the planned delivery time.

[0034] The formula for calculating the objective function is: in, The objective function value, The number of physical storage location nodes in the store. For storage location The number of inbound batches within the optimization cycle. For storage location No. The actual arrival time of the shipment. For storage location No. The actual release time of the batch of goods. The number of supplier nodes, For suppliers The actual delivery time, For suppliers The planned delivery time.

[0035] The physical storage space capacity constraints for stores include the upper limit threshold for the three-dimensional spatial capacity of the storage space and the maximum inbound throughput threshold per unit time. The upper limit threshold for the three-dimensional spatial capacity of the storage space is the product of the length, width, and height of the storage space, i.e., the total volume of the storage space; the maximum inbound throughput threshold per unit time is the sum of the total volume of goods that can be processed by all inbound channels in the store per unit time. Logistics and distribution timing constraints include the latest delivery time of supplier nodes and the fixed time of transportation routes. The latest delivery time is the product sales cutoff time minus the sum of the fixed transportation route time and the store's inventory processing time.

[0036] The optimization solution component transforms the above constraints into penalty terms for the objective function. When a constraint is violated, the penalty term increases, thereby increasing the objective function value and guiding the solution process to converge in the direction that satisfies the constraints. The formula for calculating the penalty term is:

[0037] in, For the total value of the penalty items, , , These are the penalty coefficients corresponding to storage capacity constraints, inbound throughput constraints, and delivery time constraints, respectively. For storage location No. The total volume of goods received into the warehouse. For storage location Maximum capacity, For a moment The total volume of goods entering the warehouse. The maximum inbound throughput per unit time. For suppliers The latest shipping time.

[0038] The objective function with penalty terms is: The optimization solution component employs the alternating direction multiplier method to iteratively solve the objective function with a penalty term. The alternating direction multiplier method decomposes the original problem into multiple subproblems, solves each subproblem separately, and then alternately updates the original and dual variables until the convergence condition is met.

[0039] The iterative update formula for the alternating direction multiplier method is: in, Dynamically allocate vectors to storage locations. This is the supplier delivery timing scheduling vector. As dual variables, For penalty parameters, To augment the Lagrange function, , , These are the coefficient matrix and constant vector corresponding to the constraints.

[0040] The convergence condition for iterative solution is that the absolute value of the difference between the objective function values ​​of two consecutive iterations is less than a preset convergence threshold, which is set to... When the convergence condition is met, the iteration process stops, and the current dynamic allocation vector of storage locations and the supplier's delivery time scheduling instruction are output. Each element of the dynamic allocation vector of storage locations represents the probability that a certain batch of goods from a certain supplier will be allocated to a certain storage location, and the storage location with the highest probability is taken as the final allocation result; the supplier's delivery time scheduling instruction contains the optimal delivery timestamp for each supplier.

[0041] This embodiment constructs a dynamic spatiotemporal graph to associate multi-supplier logistics data with store warehousing data. It uses a graph convolutional network with embedded temporal attention mechanism to extract spatiotemporal coupling features, establishes a dynamic mapping relationship between supplier delivery time sequence and store storage space release time sequence, and obtains a dynamic allocation scheme for storage space and delivery time sequence through a joint optimization model, thereby realizing dynamic adaptation between cross-supplier logistics time sequence and store physical space.

[0042] In a preferred embodiment, the graph construction component further concatenates the edge weights of logistics delay and supply-demand dependency relationships to generate a dynamic edge feature matrix. This dynamic edge feature matrix is ​​periodically updated as a time sliding window advances. The time sliding window is 7 days in size, with a sliding step of 1 day, meaning the dynamic edge feature matrix is ​​updated daily. Each element of the dynamic edge feature matrix corresponds to a feature vector of an edge, and the feature vector contains two dimensions: the logistics delay weight and the supply-demand dependency relationship weight.

[0043] The formula for calculating the dynamic edge feature matrix is: in, This is a dynamic edge feature matrix. This is the logistics delay weight matrix. This is a weight matrix representing the supply and demand dependency relationship. This indicates a feature splicing operation.

[0044] The characteristic composition of nodes and edges in the dynamic spatiotemporal graph is shown in Table 2: Table 2 Definition of Node and Edge Features in Dynamic Spatiotemporal Graph refer to Figure 6 The graph construction component also includes an incremental evolution component. This component monitors the data flow distribution drift of historical shipment time-series data and time-varying occupancy data, calculates the KL divergence between the data flow characteristics within the current time window and the historical baseline data flow characteristics, and triggers incremental reconstruction of the dynamic spatiotemporal graph when the KL divergence exceeds a preset drift threshold. The historical baseline data flow characteristics are the average of the data flow characteristics over the past three months, and the preset drift threshold is set to 0.2.

[0045] The formula for calculating the KL divergence is: in, Indicates the current data stream distribution Distribution of historical baseline data streams KL divergence between them Indicates features within the current time window The probability distribution, Indicates historical baseline characteristics The probability distribution.

[0046] When the KL divergence exceeds a preset drift threshold, the incremental evolution component locks down abnormal supplier nodes and abnormal store physical location nodes that have caused data distribution drift within the current time window. The method for locking down abnormal nodes is as follows: calculate the KL divergence between the local data flow characteristics of each node and the local characteristics of the historical baseline; mark nodes whose KL divergence exceeds the preset node drift threshold as abnormal nodes, with the preset node drift threshold set to 0.3. The incremental evolution component only recalculates and replaces the local edge weights and local node features related to abnormal supplier nodes and abnormal store physical location nodes, keeping the rest of the dynamic spatiotemporal graph's topology and feature parameters unchanged. Local edge weights include the weights of all edges connected to abnormal nodes, and local node features include the abnormal node's own feature vector.

[0047] The graph construction component also acquires data on abnormal events in the external logistics environment. These events include severe weather events such as heavy rain, blizzards, and typhoons; traffic incidents such as traffic control and road construction; and emergencies such as supply chain disruptions. This data is obtained through third-party meteorological data interfaces, traffic data interfaces, and news information interfaces. The data includes event type, occurrence time, scope of impact, duration, and degree of impact.

[0048] The graph construction component extracts the impact range and duration of abnormal events in the external logistics environment, and filters the target edges between affected supplier nodes and physical warehouse location nodes based on the impact range. The impact range is centered on the geographic coordinates with a radius of [missing information]. The circular area The value is determined based on the event type and impact level. For example, the impact radius of a rainstorm event is 50 kilometers, and the impact radius of a traffic control event is 10 kilometers. If the geographical coordinates of a supplier node or a physical warehouse location node are within the impact range, the edge between the supplier node and the warehouse location node is marked as the target edge.

[0049] The graph construction component calculates the incremental weight of logistics delay and the attenuation weight of supply and demand dependence on the target edge based on the duration of abnormal events in the external logistics environment and a preset environmental impact attenuation coefficient. The preset environmental impact attenuation coefficient ranges from [0.1, 0.5] and is determined according to the event type and degree of impact. For example, the attenuation coefficient for a typhoon event is 0.5, and the attenuation coefficient for a road construction event is 0.2.

[0050] The formulas for calculating the weighting of logistics delay increment and the weighting of supply-demand dependence decay are as follows: in, As a weight for incremental logistics delays, As supply and demand depend on decay weights, To preset the environmental impact attenuation coefficient, The duration of abnormal events in the external logistics environment. The base delivery time is 24 hours.

[0051] The graph construction component adds incremental logistics delay weights to the original edge weights of logistics delay, and subtracts supply-demand dependency attenuation weights from the original edge weights of supply-demand dependencies, generating a dynamic edge feature matrix under sudden environmental shocks. The updated logistics delay edge weights are: The updated supply and demand dependency edge weights are: .

[0052] The temporal feature extraction layer of the temporal attention mechanism in the feature extraction component further constructs a parallel multi-head temporal attention subspace. Each temporal attention subspace independently learns the spatial topological feature dependencies at different time scales. The number of multi-head temporal attention subspaces is set to four, corresponding to four different time scales: 1 hour, 4 hours, 1 day, and 1 week. The input of each temporal attention subspace is the spatial topological features of continuous time slices at the corresponding time scale, and the output is the sub-temporal features at that time scale.

[0053] The computation process for each temporal attention subspace is the same as that for the basic temporal attention layer, both using an attention weight matrix containing a temporal decay factor to perform a weighted summation of the spatial topological features. The temporal decay coefficients at different time scales are also discussed. Different values, 1-hour scale The value is 0.1, on a 4-hour timescale. The value is 0.05, representing a 1-day timescale. The value is 0.02, on a 1-week scale. The value is 0.01.

[0054] The temporal feature extraction layer concatenates and linearly projects the sub-temporal features output from each temporal attention subspace to generate multi-scale fused temporal features. The dimension of the concatenated sub-temporal features is the sum of the dimensions of each individual sub-temporal feature, and the weight matrix of the linear projection is a trainable parameter that maps the concatenated high-dimensional features to the same dimension as the spatial topological features.

[0055] The formula for calculating the temporal features of multi-scale fusion is as follows: in, For multi-scale fusion of temporal features, , , , Sub-time features at scales of 1 hour, 4 hours, 1 day, and 1 week, respectively. For feature splicing operations, It is a linear projection weight matrix.

[0056] The feature extraction component also extracts the spatial path topological distance between supplier nodes and store physical storage location nodes in the dynamic spatiotemporal graph, and generates a spatial topological location encoding vector based on the spatial path topological distance. The spatial path topological distance is the shortest path length between two nodes, that is, the minimum number of edges traversed from the supplier node to the store physical storage location node. The spatial topological location encoding vector adopts a sinusoidal location encoding method, and its calculation formula is as follows:

[0057] in, The position encoding vector of the first dimension, This represents the topological distance of the spatial path. The dimension of the position encoding vector. For dimensional indexing.

[0058] The feature extraction component performs a bitwise addition of multi-scale fused temporal features and spatial topological location encoding vectors to output spatiotemporal coupled features that enhance spatial constraints. The bitwise addition operation integrates the spatial information of the location encoding into the temporal features, so that the spatiotemporal coupled features simultaneously include temporal dependencies and spatial topological relationships.

[0059] The calculation formula for the spatiotemporal coupling characteristics of enhanced spatial constraints is as follows: in, To enhance the spatiotemporal coupling characteristics of spatial constraints, Based on the spatiotemporal coupling characteristics, It is a spatial topological location encoding vector.

[0060] This embodiment improves the adaptability of the dynamic spatiotemporal graph to changes in data distribution by generating a dynamic edge feature matrix and updating it periodically; it reduces computational resource consumption by realizing local reconstruction of the dynamic spatiotemporal graph through incremental evolution components; it enables the dynamic spatiotemporal graph to reflect the impact of sudden environmental changes on logistics and supply and demand by introducing abnormal event data from the external logistics environment; and it improves feature representation capabilities by constructing a multi-head temporal attention subspace and spatial topological location encoding, extracting multi-scale, spatially constrained spatiotemporal coupling features.

[0061] In another preferred embodiment, the optimization solution component further refines the constraints and solution process of the joint optimization model of warehouse location allocation and delivery timing, and adds a conflict resolution component to handle the conflict problem caused by concurrent warehousing from multiple suppliers.

[0062] The joint optimization model's physical storage capacity constraint further includes dimensional constraints on the three-dimensional space of the storage location, meaning the length, width, and height of the goods cannot exceed the length, width, and height of the storage location, rather than just constraining the total volume. The maximum inbound throughput constraint per unit time further includes the maximum throughput constraint per unit time for each inbound channel, meaning the total volume of goods that can be processed per unit time for each inbound channel cannot exceed the maximum throughput of that channel. The logistics delivery timing constraint further includes supplier production capacity constraints, meaning the maximum shipment volume of a supplier per unit time cannot exceed its production capacity.

[0063] The penalty term of the objective function of the optimization solution component includes an adaptive penalty coefficient adjustment component. The adaptive penalty coefficient adjustment component collects the current occupancy rate of the physical storage location of the store and the current fulfillment urgency of the supplier node in real time, constructs a penalty balance function with the current occupancy rate and the current fulfillment urgency as input, and dynamically updates the penalty coefficients corresponding to the physical storage location capacity constraint and the logistics delivery time sequence constraint of the store.

[0064] The current occupancy rate of a store's physical storage locations is the ratio of the sum of the current occupied volumes of all storage locations to the sum of the total capacity of all storage locations. The calculation formula is as follows: in, This represents the current occupancy rate of the store's physical storage space. For storage location The current occupied volume, For storage location Maximum capacity.

[0065] The current fulfillment urgency of a supplier node is calculated based on the remaining shelf life of the goods and the inventory level, using the following formula: in, For suppliers The current urgency of fulfilling the contract, For the shelf life of the product, This refers to the remaining shelf life of the product. To maintain a safe inventory level for goods, This represents the current inventory level of the product.

[0066] The formula for calculating the penalty balance function is: in, This refers to the updated storage location capacity constraint penalty coefficient. This is the initial penalty coefficient for storage space capacity constraints. This is the occupancy rate adjustment coefficient, with a value of 2.0. The baseline occupancy rate is set at 0.7. The updated shipping time constraint penalty coefficient. This is the initial penalty coefficient for delivery time constraints. This is an urgency adjustment factor, with a value of 1.5. The baseline urgency level is set to 1.0.

[0067] When the current occupancy rate of physical storage locations in a store increases, the penalty coefficient corresponding to the storage location capacity constraint... Increase the penalty coefficient corresponding to the delivery time constraint to suppress inbound allocation and avoid excessive storage space occupancy; when the current fulfillment urgency of a supplier node increases, the penalty coefficient corresponding to the delivery time constraint is increased. This reduces the time constraints on delivery, allowing for priority inbound processing of urgent goods. The penalty coefficient corresponding to the inbound throughput constraint... It remains unchanged, with a value of 10.0.

[0068] Table 3 shows the parameter settings for the iterative solution process using the alternating direction multiplier method. Table 3. Parameter settings for iterative solution using the alternating direction multiplier method. The optimization solution component uses the alternating direction multiplier method to iteratively solve the objective function with an adaptive penalty coefficient. In each iteration, the adaptive penalty coefficient adjustment component updates the penalty coefficient based on the current inventory occupancy rate and supplier fulfillment urgency, and then substitutes it into the augmented Lagrangian function for the solution. During the iteration, the penalty parameter... The penalty parameter is updated by gradually increasing the coefficient to accelerate convergence. When the number of iterations reaches the maximum number of iterations or the convergence condition is met, the iteration process stops, and a preliminary dynamic allocation vector for warehouse locations and supplier delivery timing instructions are output.

[0069] refer to Figure 5 Before the optimization solution component outputs the initial scheduling instructions, a conflict resolution component handles concurrent inbound conflicts. This component extracts concurrent conflict features from the spatiotemporal coupling characteristics, which represent multiple supplier delivery timing scheduling instructions pointing to the same physical warehouse location node at the same time. The conflict resolution component first calculates the maximum allowed concurrent inbound volume for the same physical warehouse location node. This maximum allowed concurrent inbound volume is calculated based on the number of inbound channels to which the location belongs and the concurrency of the sorting process.

[0070] The formula for calculating the maximum allowed concurrent data ingestion is: in, For storage location Maximum allowed concurrent data ingestion This refers to the number of inbound channels belonging to this storage location. This refers to the concurrency of goods handling per unit time for each inbound channel, i.e., the number of inbound batches that can be processed simultaneously per unit time. The time step is 1 hour.

[0071] When the number of supplier delivery scheduling instructions pointing to the same physical storage location node at the same time exceeds the maximum allowed concurrent inbound volume, the conflict resolution component applies a delay offset of a preset time step to the excess supplier delivery scheduling instructions, in ascending order of the edge weights of the supply-demand dependency relationships between supplier nodes. The preset time step is 1 hour. After each delay offset is applied, the concurrent inbound volume at that time is recalculated until the concurrent inbound volume meets the constraint. The smaller the edge weight of the supply-demand dependency relationship, the weaker the supply-demand dependency relationship between the supplier and the storage location. Therefore, its delivery sequence is delayed first, and the impact on the normal operation of the store is smaller. After processing all concurrent conflicts, the conflict resolution component outputs the final dynamic allocation vector of the store's physical storage location and the supplier delivery scheduling instructions.

[0072] The final dynamic allocation vector of physical storage locations for stores is distributed to store warehouse management personnel through the store's WMS system, guiding them in storage location planning and goods receiving operations. The final supplier delivery scheduling instructions are distributed to each supplier's ERP and logistics management systems via a standardized API interface, guiding them in arranging production and delivery plans. The system also monitors the execution of scheduling instructions in real time. When the deviation between actual execution and the scheduling instructions exceeds a preset threshold, a re-optimization process is triggered, updating the storage location allocation and delivery scheduling scheme. The preset deviation threshold is set to 2 hours; that is, when the deviation between the actual delivery time or actual arrival time and the scheduling instructions exceeds 2 hours, re-optimization is triggered.

[0073] This embodiment improves the feasibility of optimization results by refining the constraints of the joint optimization model; it achieves a dynamic balance between storage capacity constraints and logistics timing constraints through an adaptive penalty coefficient adjustment component; it solves the problems of warehouse space grabbing and channel congestion caused by concurrent warehousing from multiple suppliers through a conflict resolution component; and it ensures the stable operation of the system in a dynamically changing environment through a real-time monitoring and re-optimization mechanism.

Claims

1. A multi-supplier collaborative store digital management system, characterized in that, include: The data acquisition component collects historical shipment time-series data from multiple suppliers and time-varying occupancy data of store storage space; The graph construction component, based on the historical delivery time series data and the time-varying occupancy data, constructs a dynamic spatiotemporal graph with supplier nodes and store physical storage location nodes as vertices and logistics delay and supply-demand dependency relationships as edges. The feature extraction component embeds a temporal attention mechanism into a graph convolutional network to process the dynamic spatiotemporal graph and extract spatiotemporal coupling features under the concurrent fulfillment state of multiple suppliers. The spatiotemporal coupling features characterize the dynamic mapping relationship between the supplier's delivery time sequence and the store's warehouse release time sequence. The optimization solution component is constructed based on the spatiotemporal coupling characteristics to build a joint optimization model for warehouse location allocation and delivery timing. The physical warehouse location capacity constraint and logistics delivery timing constraint are used as the boundary conditions of the joint optimization model for warehouse location allocation and delivery timing. The system solves and outputs the dynamic allocation vector of physical warehouse locations and the delivery timing scheduling instructions of suppliers.

2. The multi-supplier collaborative store digital management system according to claim 1, characterized in that, The graph construction component, which uses logistics delay and supply-demand dependency as edges to construct a dynamic spatiotemporal graph, includes: The historical average delivery time from different supplier nodes to physical warehouse nodes is calculated as the edge weight of the logistics delay. The difference between the frequency of shipments from supplier nodes to physical warehouse nodes and the frequency of returns within a preset time window is calculated as the edge weight of the supply and demand dependency relationship. The edge weights of the logistics delay and the edge weights of the supply and demand dependency relationship are concatenated to generate a dynamic edge feature matrix of a dynamic spatiotemporal graph. The dynamic edge feature matrix is ​​periodically updated as the time sliding window progresses. The vertices contain the product attribute feature vector of the supplier node and the spatial geometric attribute feature vector of the physical warehouse node.

3. The multi-supplier collaborative store digital management system according to claim 1, characterized in that, In the feature extraction component, embedding a temporal attention mechanism into a graph convolutional network to extract spatiotemporally coupled features includes: In the spatial feature aggregation layer of the graph convolutional network, the features of adjacent nodes are weighted and aggregated based on the dynamic edge feature matrix of the dynamic spatiotemporal graph to output spatial topological features. In the temporal feature extraction layer of the temporal attention mechanism, an attention weight matrix containing a time decay factor is constructed. The spatial topological features within consecutive time slices are weighted and summed. The time decay factor decreases exponentially as the time span increases. The weighted and summed temporal features are then matrix-multiplied with the spatial topological features of the current time slice to output a spatiotemporal coupling feature that represents the dynamic mapping relationship between the supplier's delivery time sequence and the store's warehouse space release time sequence.

4. The multi-supplier collaborative store digital management system according to claim 1, characterized in that, The optimization solution component includes the following steps: Constructing a joint optimization model for warehouse location allocation and shipping timing. The objective function is to minimize the sum of the idle time of physical storage locations in stores and the delivery time offset of suppliers. The physical storage location capacity constraints include the upper limit threshold of the three-dimensional spatial capacity of physical storage locations and the maximum inbound throughput threshold per unit time. The logistics delivery timing constraints include the latest delivery time of supplier nodes and the fixed time consumption of transportation routes. The physical storage location capacity constraints and the logistics delivery timing constraints are transformed into penalty terms of the objective function. The objective function with penalty terms is iteratively solved based on the alternating direction multiplier method, and the dynamic allocation vector of physical storage locations in stores and the supplier delivery timing scheduling instructions are output.

5. The multi-supplier collaborative store digital management system according to claim 1, characterized in that, Before the optimization solution component outputs the supplier delivery timing scheduling instruction, it also includes a conflict resolution component: The concurrent conflict feature is extracted from the spatiotemporal coupling feature. The concurrent conflict feature represents that multiple supplier delivery time-series scheduling instructions point to the same physical warehouse node at the same time. The maximum allowable concurrent inbound volume is calculated based on the number of inbound channels and the concurrency of sorting processing of the same physical warehouse node. When the number of supplier delivery time-series scheduling instructions pointing to the same physical warehouse node at the same time exceeds the maximum allowable concurrent inbound volume, a delay offset of a preset time step is applied to the excess supplier delivery time-series scheduling instructions in ascending order of edge weights of the supply and demand dependency relationship of the supplier nodes until the concurrent inbound volume meets the constraint.

6. The multi-supplier collaborative store digital management system according to claim 1, characterized in that, The graph construction component also includes an incremental evolution component for constructing dynamic spatiotemporal graphs. The system monitors the data flow distribution drift status of the historical shipment time-series data and the time-varying occupancy data, calculates the KL divergence between the data flow characteristics in the current time window and the historical baseline data flow characteristics, and triggers incremental reconstruction of the dynamic spatiotemporal graph when the KL divergence exceeds a preset drift threshold. It locks the abnormal supplier nodes and abnormal store physical storage location nodes that cause data distribution drift in the current time window, and only recalculates and replaces the local edge weights and local node characteristics related to the abnormal supplier nodes and abnormal store physical storage location nodes, while keeping the rest of the topology and feature parameters of the dynamic spatiotemporal graph unchanged.

7. A multi-supplier collaborative store digital management system according to claim 2, characterized in that, The periodic update of the dynamic edge feature matrix includes: Data on abnormal events in the external logistics environment is acquired, and the impact range and duration of the abnormal events are extracted. Target edges between affected supplier nodes and physical warehouse location nodes are selected based on the impact range. Based on the duration and a preset environmental impact attenuation coefficient, the logistics delay increment weight and supply-demand dependence attenuation weight of the target edges are calculated. The logistics delay increment weight is added to the original logistics delay edge weight, and the supply-demand dependence attenuation weight is deducted from the original supply-demand dependence edge weight, generating a dynamic edge feature matrix under sudden environmental impact.

8. A multi-supplier collaborative store digital management system according to claim 3, characterized in that, The temporal feature extraction layer of the temporal attention mechanism includes: A multi-head temporal attention subspace for parallel computing is constructed. Each temporal attention subspace independently learns spatial topological feature dependencies at different time scales. The sub-temporal features output by each temporal attention subspace are concatenated and linearly projected to generate multi-scale fused temporal features. The spatial path topological distance between supplier nodes and store physical storage location nodes in the dynamic spatiotemporal graph is extracted. A spatial topological location encoding vector is generated based on the spatial path topological distance. The multi-scale fused temporal features and the spatial topological location encoding vector are added bit-by-bit to output spatiotemporal coupling features that strengthen spatial constraints.

9. A multi-supplier collaborative store digital management system according to claim 4, characterized in that, The penalty term of the objective function includes an adaptive penalty coefficient adjustment component: The system collects real-time data on the current occupancy rate of physical storage locations in stores and the current fulfillment urgency of supplier nodes. It then constructs a penalty balance function with these two data points as input. When the current occupancy rate increases, the penalty coefficient corresponding to the physical storage location capacity constraint is dynamically increased to suppress inbound allocation. When the current fulfillment urgency increases, the penalty coefficient corresponding to the logistics delivery timing constraint is dynamically decreased to relax delivery time restrictions. Based on the output value of the penalty balance function, the penalty coefficients corresponding to the physical storage location capacity constraint and the logistics delivery timing constraint are dynamically updated and iteratively solved using the alternating direction multiplier method.

10. A multi-supplier collaborative digital management method for stores, characterized in that, include: Collect historical shipment time-series data from multiple suppliers and time-varying occupancy data of store storage space; Based on the historical delivery time series data and the time-varying occupancy data, a dynamic spatiotemporal graph is constructed with supplier nodes and store physical storage location nodes as vertices and logistics delay and supply-demand dependency relationships as edges. The temporal attention mechanism is embedded into the graph convolutional network to process the dynamic spatiotemporal graph and extract the spatiotemporal coupling features under the concurrent fulfillment state of multiple suppliers. The spatiotemporal coupling features represent the dynamic mapping relationship between the supplier's delivery time sequence and the store's warehouse release time sequence. Based on the aforementioned spatiotemporal coupling characteristics, a joint optimization model for warehouse location allocation and delivery timing is constructed. The physical warehouse location capacity constraint and logistics delivery timing constraint are used as the boundary conditions of the joint optimization model for warehouse location allocation and delivery timing. The method for solving and outputting the dynamic allocation vector of physical warehouse locations and the supplier delivery timing scheduling instructions is then used.

Images