An order feature driven warehouse storage location adaptive prediction method and system

Abstract

The application discloses an order feature driven warehouse storage site adaptive prediction method and system, and belongs to the technical field of warehouse logistics. The application inputs a space-time heterogeneous graph into a space-time graph neural network to generate a space-time feature hidden vector table and issue it to an edge side computing node. The node extracts corresponding target commodity features, extracts corresponding feature data in the space-time feature hidden vector table to generate a current state vector according to the target commodity features, inputs the current state vector into a reinforcement learning strategy network to output a target storage site transfer probability distribution matrix, generates a physical topology action mask matrix to process the target storage site transfer probability distribution matrix, extracts a candidate storage site transfer scheme, and finally obtains a library moving potential energy change corresponding to the scheme. If the library moving potential energy change is greater than a preset anti-oscillation threshold, a scheduling instruction is generated, the real-time performance of warehouse storage site scheduling is improved, and the problem that the real-time performance of warehouse storage site scheduling is reduced when a highly dynamic real work environment is faced in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of warehousing and logistics technology, and in particular to an order feature-driven adaptive prediction method and system for warehouse storage locations. Background Technology

[0002] With the continuous development of intelligent logistics systems, the requirements for the efficiency of inbound and outbound operations in warehousing systems are increasing. In order to reduce picking costs and optimize internal logistics flow, the industry has widely introduced order feature-driven adaptive prediction technology for warehouse locations. The core motivation for designing this technology is to dynamically adjust the storage location of goods by mining the co-occurrence patterns and physical spatial attributes of goods hidden in historical order data. This allows the warehouse layout to adapt to changes in order structure in advance, shortening the physical addressing distance in subsequent operations.

[0003] Existing technologies for warehouse location prediction typically rely on two main implementation paths. One path involves building a complex end-to-end deep learning model or global graph neural network, using the full warehouse topology and historical operation logs as input, and performing deep feature extraction and objective solving within a single system to obtain the globally optimal location transfer solution. The other path relies on lightweight heuristic rule algorithms, which, upon receiving the current order flow, perform real-time calculations and issue instructions based solely on local available space and surface association rules.

[0004] For example, Chinese invention patent CN121329281A discloses an AI-based dynamic optimization method for warehouse locations, which includes: first, acquiring multi-dimensional data on goods, warehousing, and business; then, calculating a "time-sensitive location-goods matching score" that integrates the future outbound frequency of goods, the correlation with neighboring goods, physical distance, and a dynamic time urgency factor that changes as the expected picking time approaches; using a reinforcement learning model, taking the matching score as the key decision basis, and guided by a composite reward function that takes into account the overall layout gain, physical handling costs, and high-urgency task rewards, outputting a dynamic location optimization scheme with globality and foresight.

[0005] For example, Chinese invention patent CN113636254B discloses a method and system for adaptive optimization adjustment of storage slots in an intelligent warehousing system. The method includes: first, constructing an intelligent warehousing system capable of adaptively adjusting space, mainly comprising a storage rack structure, an information acquisition unit, and a storage slot adaptive optimization control system; then, predicting the future storage space requirements of different raw materials and obtaining the storage slot space dimensions required by all raw materials; further, reading the overall information of the storage rack and setting constraints to construct a mathematical model for adaptive optimization adjustment of storage slots; based on the mathematical model, using a genetic algorithm to solve the storage slot optimization problem to obtain the optimal storage slot configuration structure scheme; finally, the control system of the intelligent warehousing system adjusts the storage rack structure according to the obtained optimal storage slot configuration structure scheme to achieve adaptive adjustment of each storage slot.

[0006] However, existing technologies suffer from a critical flaw: a deep architectural coupling between the global deep computation mechanism and real-time physical execution response. Under this tightly coupled framework, the system faces a dilemma where computational overhead and decision quality mutually constrain each other. Forcing a heavy-duty global feature traversal of the model during order influx results in massive computational latency that cannot match the rapid response demands of on-site operations, leading to severe control flow congestion. Conversely, if lightweight local computation is adopted to ensure real-time performance, the prediction process remains a pure software logic black box, detached from verification of the hard safety boundaries of the real physical space and consideration of the marginal cost of equipment allocation. This inevitably leads to frequent outputs of illegal scheduling commands that violate underlying physical safety constraints, or ineffective oscillating warehouse relocations where expected returns cannot compensate for the physical execution costs. This severe disconnect between the information processing level and physical constraint boundaries results in reduced real-time performance of warehouse location scheduling in highly dynamic real-world operating environments. Summary of the Invention

[0007] To address the technical problem of reduced real-time performance of warehouse location scheduling in highly dynamic real-world operating environments, this invention provides an order feature-driven adaptive prediction method and system for warehouse locations. The technical solution is as follows:

[0008] On the one hand, an order feature-driven adaptive prediction method for warehouse storage locations is provided. This method includes: acquiring historical order logs, SKU basic attribute data, and storage location topology status data from the warehousing system; constructing a spatiotemporal heterogeneous graph containing product nodes, order nodes, and storage location nodes on a cloud server, where the spatiotemporal heterogeneous graph represents product co-occurrence relationships, order association relationships, and storage location spatial adjacency relationships; in the cloud server, inputting the spatiotemporal heterogeneous graph into a spatiotemporal graph neural network for periodic iterative calculations, extracting global-dimensional product-order-storage location association features, generating a spatiotemporal feature latent vector table, and asynchronously distributing the spatiotemporal feature latent vector table to edge computing nodes; the edge computing nodes receive real-time incoming current order data, extract the target product features corresponding to the current order data, and retrieve the corresponding product features from the spatiotemporal feature latent vector table based on the target product features. The system concatenates the corresponding feature data with the current order data to generate a current state vector representing the current scheduling environment. This current state vector is then input into a pre-deployed reinforcement learning strategy network for inference, outputting a target storage location transfer probability distribution matrix. Real-time acquisition of the underlying equipment's operating status and spatial carrying capacity limits generates a physical topology action mask matrix with the same dimension as the target storage location transfer probability distribution matrix. The physical topology action mask matrix and the target storage location transfer probability distribution matrix are then multiplied element-wise to filter out illegal transfer probabilities that violate physical safety constraints, extracting the operation path corresponding to the highest effective probability as a candidate storage location transfer scheme. The system obtains the change in transfer potential energy corresponding to the candidate storage location transfer scheme. If the change in transfer potential energy exceeds a preset anti-vibration threshold, a scheduling instruction is generated and executed based on the candidate storage location transfer scheme.

[0009] On the other hand, an order feature-driven adaptive prediction system for warehouse storage locations is provided, including: a cloud-based graph computation module, used to acquire historical order logs, SKU basic attribute data, and storage location topology status data of the warehousing system, and construct a spatiotemporal heterogeneous graph containing product nodes, order nodes, and storage location nodes on a cloud server. The spatiotemporal heterogeneous graph represents product co-occurrence relationships, order association relationships, and storage location spatial adjacency relationships. In the cloud server, the spatiotemporal heterogeneous graph is input into a spatiotemporal graph neural network for periodic iterative computation, extracting global-dimensional product-order-storage location association features, generating a spatiotemporal feature latent vector table, and asynchronously distributing the spatiotemporal feature latent vector table to edge computing nodes; and an edge-side state parsing module, used by edge computing nodes to receive real-time incoming current order data, extract the target product features corresponding to the current order data, and retrieve the corresponding feature data from the spatiotemporal feature latent vector table based on the target product features, and compare it with the current order data. Previous order data is concatenated to generate a current state vector representing the current scheduling environment. The edge-side policy inference module inputs the current state vector into a pre-deployed reinforcement learning policy network for inference and outputs a target storage location transfer probability distribution matrix. The physical topology mask filtering module obtains real-time information on the working status of the underlying equipment and the spatial carrying capacity limits of the warehousing system, and generates a physical topology action mask matrix with the same dimension as the target storage location transfer probability distribution matrix. The physical topology action mask matrix and the target storage location transfer probability distribution matrix are multiplied element-wise to filter out illegal transfer probabilities that violate physical safety constraints and extract the operation path corresponding to the highest effective probability as a candidate storage location transfer scheme. The potential energy judgment and instruction scheduling module obtains the change in transfer potential energy corresponding to the candidate storage location transfer scheme. If the change in transfer potential energy is greater than the preset anti-vibration threshold, a scheduling instruction is generated and issued for execution based on the candidate storage location transfer scheme.

[0010] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0011] 1. The present invention provides an order feature-driven adaptive prediction method for warehouse storage locations. By constructing a spatiotemporal heterogeneous graph containing product nodes, order nodes, and storage location nodes on a cloud server, and using a spatiotemporal graph neural network to perform periodic iterative calculations to generate a spatiotemporal feature latent vector table, which is asynchronously distributed to the edge computing nodes, the edge side only needs to perform a hash lookup with O(1) complexity to obtain a dense feature vector containing global spatiotemporal correlation information. This decouples the heavy global feature calculation from the real-time scheduling decision at the architecture level, and realizes that warehouse storage location scheduling has millisecond-level real-time response capability while maintaining global feature perception capability. This effectively solves the problem of reduced real-time performance of warehouse storage location scheduling caused by the architecture-level coupling between the global deep calculation mechanism and the real-time physical execution response in the prior art.

[0012] 2. In traditional warehousing systems, the storage and scheduling of goods often rely on fixed rules or simple priority sorting, which cannot fully consider the complex relationship between goods and orders, and also lacks the ability to dynamically adjust to the real-time changing environment. However, this invention constructs a spatiotemporal heterogeneous graph containing product nodes, order nodes, and storage location nodes, which can more comprehensively represent the co-occurrence relationship of goods, the association between orders, and the spatial adjacency relationship of storage locations. This graph modeling method effectively integrates the multi-dimensional information of the warehousing system, providing a solid data foundation for subsequent intelligent analysis and decision-making.

[0013] 3. By running a spatiotemporal graph neural network on a cloud server, the system can perform iterative calculations periodically to extract global correlation features between products, orders, and storage locations, generating a spatiotemporal feature latent vector table. In this way, the system can not only achieve comprehensive mining of historical data, but also continuously optimize feature vectors in each iteration, providing accurate feature data support for real-time scheduling decisions. This feature data is asynchronously distributed to edge computing nodes, ensuring that the system can quickly obtain and apply the most relevant historical data when processing real-time orders, thus improving response speed.

[0014] 4. Upon receiving real-time order data, the edge computing nodes quickly retrieve relevant data from the spatiotemporal feature latent vector table based on the characteristics of the target product, generating a state vector for the current scheduling environment. This state vector is input into the reinforcement learning policy network for reasoning and decision-making, outputting a probability distribution matrix for the transfer of the target storage location. By combining the working status information of the underlying equipment of the warehousing system and the spatial carrying capacity limits, the generated physical topology action mask matrix can effectively filter out illegal transfer probabilities that violate physical safety constraints, thereby ensuring the safety and feasibility of scheduling operations. Through the calculation of the change in potential energy of the warehouse transfer, the system can determine the stability of the candidate storage location transfer scheme. If it exceeds the preset anti-vibration threshold, the system will generate a scheduling instruction and issue it for execution, avoiding instability in warehouse operations caused by unreasonable scheduling. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating an order feature-driven adaptive prediction method for warehouse location is provided in this application embodiment;

[0017] Figure 2This is a schematic diagram of the structure of an order feature-driven adaptive prediction system for warehouse location provided in an embodiment of this application. Detailed Implementation

[0018] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present disclosure are shown in the drawings, it should be understood that embodiments of the present disclosure may be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure.

[0019] It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure. In the description of the embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "this embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] The core architecture of this invention is a three-layer asynchronous collaborative system of "cloud-edge-thing": the cloud server is responsible for offline calculation of global spatiotemporal features and periodically distributes spatiotemporal feature latent vector tables (update cycle approximately 4 hours); the edge computing nodes are responsible for receiving real-time order streams and executing lightweight strategy inference (single inference latency less than 10ms); the underlying physical execution layer is responsible for providing microsecond-level physical constraint masks and driving AGVs (Automated Guided Vehicles) to perform warehouse transfer actions. The three layers are decoupled through asynchronous communication, ensuring that heavy computing in the cloud does not block real-time responses on the edge, and that high-frequency inference on the edge is not limited by the response rate of the underlying hardware. The following uses a specific warehouse scenario as a general example: the warehouse has 200 storage locations, manages 500 SKUs (StockKeeping Units), is equipped with 10 AGVs, and processes an average of 3,000 orders per day.

[0022] like Figure 1 The diagram shown is a flowchart of an order feature-driven adaptive prediction method for warehouse storage locations provided in an embodiment of this application. The method includes the following steps:

[0023] Step 1: Obtain historical order logs, SKU basic attribute data, and storage location topology status data from the warehousing system. Construct a spatiotemporal heterogeneous graph containing product nodes, order nodes, and storage location nodes on the cloud server. The spatiotemporal heterogeneous graph represents the co-occurrence relationship of products, the association relationship of orders, and the spatial adjacency relationship of storage locations.

[0024] It should be noted that by abstracting the co-occurrence relationships of goods contained in historical order logs and the physical spatial adjacency relationships in the topological status data of storage locations into a spatiotemporal heterogeneous graph, the heterogeneous graph can simultaneously express multiple semantic relationships between different types of entities (goods, orders, storage locations). Compared with traditional relational databases or isomorphic graph models, it can more completely preserve the spatiotemporal relationship structure in the warehousing system, providing a high-quality input basis for the subsequent feature aggregation of spatiotemporal graph neural networks.

[0025] Specifically, historical order logs record all inbound and outbound operations of the warehousing system over a period of time (usually 60 to 180 days), including fields such as order identifier, order timestamp, list of goods included in the order, and quantity; SKU basic attribute data records the static physical attributes of each SKU (Stock Keeping Unit) in the warehousing system, including fields such as product identifier (SKU code), product weight, product volume, and category label; and warehouse topology status data records the physical space information of each warehouse location, including fields such as warehouse location coordinate identifier (e.g., "Area A-03 Shelf-02"), three-dimensional spatial coordinates, load limit, and current occupancy status.

[0026] The process of constructing a spatiotemporal heterogeneous graph containing product nodes, order nodes, and storage location nodes on a cloud server is as follows:

[0027] Order nodes are generated by extracting order identifiers from historical order logs, product nodes are generated by extracting product identifiers from SKU basic attribute data, and storage location nodes are generated by extracting coordinate identifiers from storage location topology status data.

[0028] Specifically, node generation is the first step in graph construction. Order nodes are indexed by their unique order identifier, product nodes by their SKU code, and storage location nodes by their three-dimensional coordinates. The initial feature vector of a node typically includes its static attributes: product nodes may include physical attributes such as product weight, volume, and category label; storage location nodes may include spatial attributes such as the storage location's weight limit, three-dimensional coordinates, and current occupancy status; and order nodes may include business attributes such as order timestamp and order urgency.

[0029] For example, in the example scenario, the historical order log contains 270,000 order records from the past 90 days, which are extracted to generate 270,000 order nodes, 500 product nodes, and 200 storage location nodes.

[0030] Based on the product information contained in the historical order logs, order association edges with corresponding preset weights are constructed between order nodes and product nodes that have an inclusion relationship.

[0031] It should be noted that order association edges represent the semantic relationship of "an order contains a certain product," and their preset weights are usually set as the frequency or quantity percentage of that product in the order. This type of edge is a key structure for mining product co-occurrence patterns: if two product nodes are frequently associated with each other through a large number of order nodes, they have a high degree of structural similarity in the graph. During the aggregation process, the spatiotemporal graph neural network will encode this co-occurrence information into their respective latent vectors. For example, product SKU_A (shampoo) and product SKU_B (shower gel) co-occurred in the same order 18,000 times in 270,000 historical orders. Based on this, a high-frequency co-occurrence association structure with a weight of 18,000 / 270,000 = 0.067 is established between the product nodes corresponding to SKU_A and SKU_B through the commonly associated order nodes. This means that when the system needs to find the optimal storage location for SKU_A, its latent vector has implicitly encoded the global co-occurrence knowledge that "it should be stored nearby with SKU_B."

[0032] Based on the coordinate distance in the cargo location topology status data, construct spatial adjacency edges with corresponding preset weights between cargo location nodes whose spatial distance is less than a preset distance threshold.

[0033] It should be noted that spatial adjacency edges represent the physical proximity between storage locations. Their preset weight is typically set to the reciprocal of the Euclidean distance between the center points of two storage locations; the closer the distance, the greater the weight. The preset distance threshold is usually set based on the warehouse rack spacing and AGV aisle width, for example, 2.0 meters. This edge type allows the spatiotemporal graph neural network to propagate topological congestion information between physically adjacent storage locations during spatial feature aggregation. This implicitly encodes the spatial occupancy status of its surrounding area in the latent vector of each storage location node, providing a spatial awareness basis for optimizing subsequent candidate storage location transfer schemes.

[0034] For example, in the example scenario, the Euclidean distance between storage location Loc_05 (coordinates (3.0, 1.5, 2.0) meters) and storage location Loc_06 (coordinates (3.0, 1.5, 3.5) meters) is 1.5 meters, which is less than the preset distance threshold of 2.0 meters. Therefore, a spatial adjacency edge is constructed between the two, with a weight of 1 / 1.5 = 0.667.

[0035] Merge order nodes, product nodes, storage location nodes, as well as order-related edges and spatial adjacency edges to generate a spatiotemporal heterogeneous graph.

[0036] Specifically, the final generated spatiotemporal heterogeneous graph is a heterogeneous directed weighted graph containing three types of nodes (order nodes, product nodes, and storage location nodes) and two types of edges (order-related edges and spatial adjacency edges). In the example, the graph size is: a total of 270,000 + 500 + 200 = 270,700 nodes, and the total number of edges is approximately the number of order-related edges (corresponding to the order-product inclusion relationship) plus the number of spatial adjacency edges (the number of storage location pairs with a distance of less than 2.0 meters between each of the 200 storage locations, approximately 600 edges). This graph is stored on a cloud server in the form of a graph database or a sparse adjacency matrix, serving as the input basis for the spatiotemporal graph neural network. It should be noted that, in addition to the graph database, alternative storage methods include sparse matrix representations based on COO (Coordinate) or CSR (Compressed Sparse Row) formats, the choice of which depends on the input interface requirements of the subsequent graph neural network framework (such as PyTorch Geometric, DGL, etc.).

[0037] Step 2: On the cloud server, the spatiotemporal heterogeneous graph is input into the spatiotemporal graph neural network for periodic iterative calculations to extract global-dimensional product-order-location association features, generate a spatiotemporal feature latent vector table, and asynchronously distribute the spatiotemporal feature latent vector table to the edge computing nodes.

[0038] In this embodiment, heavy global feature computation is performed offline in the cloud. A spatiotemporal graph neural network is used to compress the high-dimensional, sparse, heterogeneous graph into a low-dimensional, dense spatiotemporal feature latent vector table, achieving the architectural goal of "no information loss and a dramatic increase in speed." The spatiotemporal feature latent vector table, as a distillation product of cloud intelligence, is asynchronously distributed to the edge, allowing the edge to complete global feature retrieval without accessing the original graph, reducing the lookup complexity to O(1). The spatiotemporal graph neural network updates its computations approximately every 4 hours, pushing the new latent vector table to the edge nodes after each update, achieving a gradual iterative update of global knowledge.

[0039] It should be added that the spatiotemporal heterogeneous graph is input into the spatiotemporal graph neural network for periodic iterative calculations to extract global-dimensional product-order-location association features and generate a spatiotemporal feature latent vector table. The specific steps are as follows:

[0040] First, a spatial feature aggregation operation is performed on the spatiotemporal heterogeneous graph using a spatiotemporal graph neural network, aggregating the feature data of adjacent nodes to the central node to obtain a spatial topological feature vector.

[0041] Spatial feature aggregation is a core operation in the graph convolutional layer of a spatiotemporal graph neural network. For each node in the graph, this operation aggregates the weighted sum of the features of all its neighboring nodes to the central node, thus enabling the feature vector of the central node to encode the structural information within its first-order or even multi-order neighborhood. Since this invention uses heterogeneous graphs, different types of edges employ different learnable weight matrices, the specific calculation formula of which is as follows:

[0042] ;

[0043] in, For nodes In the The output feature vector of the layer has a dimension of (In this embodiment) ); This invention comprises a set of relationship types, including two types: "order association" and "spatial adjacency." For nodes In relation The set of neighboring nodes; The normalization constant (usually taken as ) (i.e., the number of neighboring nodes). For relationship In the The learnable weight matrix of the layer has a dimension of ; This is a self-connection weight matrix used to preserve the feature information of the node itself from the previous layer; For the activation function (such as ReLU). This formula indicates that the node... The updated feature is the output after activation, which is the sum of the weighted aggregation of neighbor features under each type of relationship and the linear transformation of its own features.

[0044] For example, the updated feature vector of product node SKU_A (shampoo) will simultaneously integrate the following information: features of all product nodes that co-occur with SKU_A in the same order (such as SKU_B shower gel, SKU_C conditioner, etc.) aggregated through order association edges, and features of the current storage location node of SKU_A and its adjacent storage location nodes aggregated through spatial adjacency edges. Assuming SKU_A has 50 neighboring product nodes under the relationship "order association", then... The feature vector of each neighbor is first processed After linear transformation, the vector is normalized by dividing by 50 and finally aggregated into a 64-dimensional dense vector that reflects the global co-occurrence semantics. After layer-by-layer aggregation through 3 to 4 layers of graph convolution, the latent vector of each node has encoded the global structural information in its 3rd to 4th order neighborhood.

[0045] It should be added that the specific implementation of spatiotemporal graph neural networks is not limited to the heterogeneous graph convolution form mentioned above. Alternative techniques include graph attention networks (GAT), relational graph convolutional networks (R-GCN), etc., as long as they can perform feature aggregation operations of multiple relation types on heterogeneous graphs.

[0046] Next, a temporal convolution operation is performed on the spatial topological feature vector to extract the temporal evolution features of historical order logs within a preset time window, resulting in a multidimensional fused feature vector.

[0047] It should be noted that the purpose of temporal convolution is to capture the dynamic trends of product co-occurrence relationships and storage location usage patterns over time. Spatial feature aggregation only processes the spatial topological information of a snapshot of the map at a certain moment, but the co-occurrence frequency of products and the usage popularity of storage locations are time-dependent in different time periods (such as promotional seasons, weekends, and holidays). Temporal convolution extracts trend and periodic features in the time dimension by performing one-dimensional convolution on the spatial topological feature vectors of multiple consecutive moments within a preset time window (usually set to 7 to 30 days), and fuses them into the spatial topological feature vector to generate a multi-dimensional fused feature vector that simultaneously encodes spatial structure information and temporal evolution information. In addition, there are other alternative techniques, such as LSTM (Long Short-Term Memory) networks, Transformer temporal encoders, and other temporal modeling methods.

[0048] Finally, the multidimensional fusion feature vectors are reduced in dimensionality to obtain dense feature vectors that correspond one-to-one with each commodity identifier and each storage location identifier. The dense feature vectors are then aggregated to construct a spatiotemporal feature latent vector table.

[0049] Specifically, the dimensionality reduction process uses a fully connected layer to map the multi-dimensional fused feature vector from a high-dimensional space to a low-dimensional dense space (e.g., reducing from 128 dimensions to 64 dimensions), resulting in the final dense feature vector. The spatiotemporal feature latent vector table is physically a hash lookup table with product identifiers or location identifiers as keys and the corresponding dense feature vectors as values. In the example, this table contains 500 product latent vectors and 200 location latent vectors, each a 64-dimensional floating-point array. The total storage space is approximately (500+200)×64×4 bytes = 175KB, allowing for network transmission from the cloud to the edge within milliseconds.

[0050] Step 3: The edge computing node receives the real-time influx of current order data, extracts the target product features corresponding to the current order data, and retrieves the corresponding feature data from the spatiotemporal feature latent vector table based on the target product features. This data is then concatenated with the current order data to generate a current state vector representing the current scheduling environment.

[0051] It should be noted that the core operation is executed in real time on the edge computing node. It uses the target product identifier and the starting storage location identifier as keys to perform a hash lookup in the cached spatiotemporal feature latent vector table with O(1) complexity to obtain the corresponding dense feature vector. Then, it is concatenated with the real-time features of the current order to integrate the global spatiotemporal knowledge distilled in the cloud with the real-time business information captured by the edge into a unified state representation.

[0052] The specific process for generating the current state vector representing the current scheduling environment is as follows:

[0053] Extract the identifier of the target product to be processed from the characteristics of the target product.

[0054] Specifically, when an order enters the edge computing node, the system first parses the product list contained in the order data, and then performs subsequent location optimization evaluation for each product in the list. The target product identifier is the SKU code currently being evaluated. For example, in a typical scenario, an order contains three products: SKU_A (shampoo), SKU_B (body wash), and SKU_C (conditioner). The system first extracts SKU_A as the target product identifier to be processed.

[0055] Retrieve the target product latent vector corresponding to the target product identifier and the starting storage location latent vector corresponding to the current starting storage location of the target product from the spatiotemporal feature latent vector table.

[0056] It should be noted that the latent vector of the target product encodes the co-occurrence relationship characteristics and temporal evolution trend of the product in the global spatiotemporal dimension, while the latent vector of the starting storage location encodes the neighborhood occupancy status and historical usage frequency of the storage location in the spatial topology. The combination of the two latent vectors provides dual global perception information at the "product level" and "spatial level". For example, in connection with the previous example, the system retrieves the 64-dimensional latent vector L(SKU_A) of the target product in the spatiotemporal feature latent vector table using SKU_A as the key, and retrieves the 64-dimensional latent vector L(Loc_12) of the starting storage location using the current storage location Loc_12 of SKU_A as the key.

[0057] The order urgency value, target product latent vector, and starting storage location latent vector in the current order data are concatenated and merged according to a pre-defined dimension order to generate a current state vector with a unified dimension.

[0058] Specifically, the formula for concatenating the current state vector is:

[0059] ;

[0060] in, This is the current order's instant feature vector (containing scalar features such as order urgency value, with dimensions of...). ), The latent vector of the target product (dimension: ), The initial storage location latent vector (dimension: Concat is a dimension-based concatenation operation. The final generated current state vector... The dimension is In the example, let the order urgency be a scalar value of 0.8 (1.0 represents the most urgent, and 0.0 represents a regular order), then... =Concat([0.8],L(SKU_A),L(Loc_12)), the total dimensions are 1+64+64=129.

[0061] Step 4: Input the current state vector into the pre-deployed reinforcement learning policy network for inference, and output the target cargo location transfer probability distribution matrix.

[0062] It should be noted that the reinforcement learning policy network (i.e., the Actor network) receives the current state vector. As input, forward inference is performed using a lightweight multilayer perceptron (MLP), outputting a transition probability distribution matrix covering all target storage locations. After the network is pre-trained in the cloud using policy gradient algorithms (such as PPO, A2C, etc.), only the network parameters for the inference phase are deployed to the edge. Since the forward inference of MLP only involves matrix multiplication and activation function operations, the computational overhead is low, and the latency of a single inference at the edge can be controlled within 10ms.

[0063] Specifically, the inference formula for reinforcement learning policy networks is as follows:

[0064] ;

[0065] in, The parameter is In a multilayer perceptron, the Softmax function maps the output of the MLP to a probability distribution, making... The sum of all elements in the equation is 1. The dimension is ( (i.e., the total number of storage locations), of which the first... element This indicates that the target goods will be transferred from the initial storage location to the next location. The recommendation probability for each storage location. For example, in the through-example, this is represented by a 129-dimensional current state vector. enter Then, a 200-dimensional probability distribution matrix is ​​output via Softmax. ,in (Loc_15 has the highest recommendation probability) , The probability of other cargo locations being unavailable is relatively low.

[0066] It should be added that the training process of the reinforcement learning policy network is completed offline in the cloud. The training objective is to maximize the long-term cumulative savings in picking time, and the training data comes from the historical operation logs of the warehouse system in a simulated environment. The architecture of the policy network is not limited to MLP; alternative technologies include attention-based Transformer policy networks, as long as they can meet the millisecond-level inference latency requirements at the edge.

[0067] Step 5: Obtain real-time information on the working status of the underlying equipment and the spatial carrying capacity limit conditions of the warehousing system, and generate a physical topology action mask matrix with the same dimension as the target storage location transfer probability distribution matrix; perform element-wise multiplication of the physical topology action mask matrix and the target storage location transfer probability distribution matrix to filter out illegal transfer probabilities that violate physical safety constraints, and extract the operation path corresponding to the maximum effective probability as a candidate storage location transfer scheme.

[0068] It should be noted that the target cargo location transfer probability distribution matrix The policy network, generated based on mathematical fitting, lacks physical common sense—it might assign a high probability of transferring a space that is already occupied, a shelf that is already full, or a space that needs to be kept at a safe distance from hazardous materials. Directly executing scheduling based on the network output would inevitably generate illegal instructions that violate physical safety constraints. The introduction of a physical topology action mask matrix, by mathematically truncating the probability distribution, fundamentally prevents the generation of illegal instructions.

[0069] Specifically, the physical topology action mask matrix with the same dimension as the target cargo location transfer probability distribution matrix is ​​generated. The specific process is as follows:

[0070] The system parses the underlying equipment's operating status information to obtain the current occupancy status, and parses the space carrying capacity limit information to obtain the remaining carrying capacity attributes of the target storage location.

[0071] Specifically, the underlying equipment's operational status information is pushed to the physical execution layer in real time at a microsecond frequency by the warehouse control system (WCS). This includes data such as the current occupancy status (idle / occupied) of each storage location, the real-time location and battery level of each AGV, and the current load capacity of each shelf. Space load-bearing limit information records the physical constraint parameters of each storage location, including the maximum load capacity, the maximum volume of goods that can be stored, and special constraints for specific storage locations (such as hazardous materials isolation areas). The remaining load capacity attribute is the margin obtained by subtracting the total weight of currently stored goods from the maximum load capacity of the storage location.

[0072] Compare the remaining load-bearing attributes with the product weight or volume parameters in the SKU basic attribute data.

[0073] If the comparison result is that the carrying attribute is exceeded, or the current occupancy status is that it is already occupied, then the value zero, which represents illegality, will be assigned to the corresponding position in the matrix to be generated.

[0074] If the comparison result meets the carrying attribute and the current occupancy status is idle, then the corresponding position in the matrix to be generated will be assigned a value representing a valid value.

[0075] Based on all the zero and one values, a physical topology action mask matrix composed of Boolean values ​​is generated.

[0076] For example, in the example scenario, the target product SKU_A weighs 1.5kg, and among the 200 storage locations, storage location Loc_15 is currently free and has a remaining capacity of 8.0kg (1.5kg < 8.0kg, which is acceptable). Storage location Loc_22 is already occupied by other goods. The remaining carrying capacity of storage space Loc_07 is only 0.5kg (1.5kg > 0.5kg, exceeding the limit). The final generated physical topology action mask matrix It is a 200-dimensional Boolean vector, with approximately 120 positions set to 1 (valid) and 80 positions set to 0 (invalid).

[0077] Specifically, the operation path corresponding to the highest effective probability is extracted as a candidate cargo location transfer scheme, as follows:

[0078] Perform element-wise multiplication to set the transition probability corresponding to the value of zero in the target cargo location transfer probability distribution matrix to zero, thus obtaining the intermediate filtering matrix.

[0079] Specifically, the element-wise product operation is called the Hadamard product, and its calculation formula is as follows:

[0080] ;

[0081] in, This is the element-wise multiplication operator; ; Minimum constant (e.g.) This is used to prevent abnormal values ​​where the denominator is zero. For The position with a median of 0, regardless of The probability value of the corresponding location determines the product result to be absolutely zero, thus achieving a hard cutoff of the probability of illegal transfer actions. When all storage locations are illegal, i.e. All zeros, the denominator approaches zero. To ensure the system does not crash, it outputs a probability distribution that is close to all zeros. At this time, the system does not perform any database transfer operations.

[0082] Referring to the previous example, (reserve), (Truncation) (Truncation).

[0083] The non-zero transition probability values ​​in the intermediate filtering matrix are summed and normalized to obtain a normalized safety probability matrix in which the total probability is redistributed to the value of one.

[0084] Referring back to the previous example, assuming the sum of non-zero probabilities after filtering is 0.65, then... It becomes the highest probability after normalization.

[0085] Find the target storage location node with the highest probability value in the normalized safety probability matrix, and determine the data flow direction from the starting storage location to the target storage location node as the candidate storage location transfer scheme.

[0086] For example, referring to the previous example, the storage location with the highest probability in the normalized security probability matrix is ​​Loc_15 ( Therefore, the candidate storage location transfer scheme is determined as follows: transfer SKU_A from the initial storage location Loc_12 to the target storage location Loc_15, i.e., Move(SKU_A,Loc_12,Loc_15).

[0087] Step 6: Obtain the change in transfer potential energy corresponding to the candidate storage location transfer plan. If the change in transfer potential energy is greater than the preset anti-vibration threshold, generate a scheduling instruction based on the candidate storage location transfer plan and issue it for execution.

[0088] It should be noted that this step is the final line of defense in ensuring "economic rationality" within the entire technical solution. While step five above has already ensured the physical legitimacy of candidate solutions through physical masking, physical legitimacy does not equate to economic rationality. Although a warehouse transfer operation may not violate physical constraints, its execution consumes AGV transport resources and occupies warehouse passage time. If the expected savings in picking time cannot offset these costs, then this transfer is an "ineffective oscillation" in an economic sense. The change in transfer potential energy is precisely the quantitative indicator for conducting a unified-dimensional marginal cost-benefit analysis of each candidate transfer operation.

[0089] The formula for calculating the change in potential energy during storage transfer is:

[0090] ;

[0091] in, In order to be within the preset evaluation period Inner The expected saving time per picking at any given moment, in seconds; The mechanical energy consumption cost of calling an AGV to perform a single warehouse transfer includes the AGV's travel distance and battery consumption, and is expressed in the number of equipment operations or joules. The penalty cost for congestion caused to warehouse traffic by warehouse transfer operations, expressed in seconds or congestion index; The revenue conversion factor converts time savings into a uniform standard cost unit (CostUnit), with the dimension [cost unit / second]. This is the equipment cost conversion factor, measured in units of cost per unit or joule. The congestion cost conversion factor has the dimension of [cost unit / second]. These three conversion factors unify different physical quantities into dimensionless standard cost units, allowing for direct addition and subtraction of revenue and cost items.

[0092] Dimensional verification of the formula: Left side The standard cost unit is used; the first term in the right-hand formula... The dimension is [cost unit / second] multiplied by The dimension is [second], multiplying them gives [cost unit]; the second term on the right. The unit of measurement is [cost unit / time] multiplied by The dimension is [times], multiplying them gives [cost unit].

[0093] For example, throughout the example scenario, for the candidate solution Move(SKU_A,Loc_12,Loc_15), assume a preset evaluation period. During this period, it is estimated that there will be 15 orders containing SKU_A. After moving SKU_A from Loc_12 to Loc_15, the picking path for each order will be shortened by an average of 8 seconds. Therefore, the total picking time saving is... Seconds. Let's assume... Cost unit per second, then future outbound operation time revenue. Cost unit. Cost per AGV moving from Loc_12 to Loc_15 in a single transfer. Next scheduling, set Cost unit / trip; Current congestion index of the warehouse passageway traversed by this route. Second delay, set Cost unit: per second. This refers to the energy consumption cost of mechanical equipment. Cost unit. Ultimately, Cost unit.

[0094] If the anti-vibration threshold is preset Cost unit, then The data transfer scheme was deemed economically reasonable, and the system generated a scheduling instruction Move(SKU_A, Loc_12, Loc_15) and issued it to the WCS for execution. Conversely, if... If the scheme is physically legal, it is economically unreasonable. The system intercepts the instruction and does not execute the transfer operation, thus avoiding invalid oscillating transfer.

[0095] Furthermore, the specific steps for obtaining the change in potential energy for warehouse transfer corresponding to the candidate warehouse location transfer scheme are as follows:

[0096] The first step is to input the candidate location transfer plan into the preset time benefit evaluation function, calculate the total picking time saved by the candidate location transfer plan within the preset evaluation period, and multiply it by the preset benefit conversion coefficient to obtain the future outbound operation time benefit.

[0097] It should be noted that the input to the time-benefit evaluation function is the candidate storage location transfer scheme (including the target commodity identifier, the starting storage location identifier, and the target storage location identifier). Its evaluation logic is as follows: based on the commodity co-occurrence frequency information encoded in the spatiotemporal feature latent vector table, the value is estimated within the preset evaluation period. The system estimates how many orders will contain the target product and calculates the difference in picking path length between the starting location and the target location. The path savings for each order are then summed to obtain the total picking time savings.

[0098] The second step is to extract the equipment movement parameters from the underlying equipment working status information, calculate the number of scheduling operations required to move from the starting location to the target location and the expected congestion time, and multiply them by a preset cost conversion factor to obtain the mechanical equipment energy consumption cost.

[0099] The third step is to calculate the numerical change in the potential energy of the warehouse transfer by subtracting the energy consumption cost of mechanical equipment from the future time revenue of the outbound operation.

[0100] It should be added that when the potential energy changes during the transfer... When the calculation result is negative, it indicates that the cost of the transfer operation has exceeded its expected benefits, and the system will directly intercept the instruction without execution. It is a positive value but less than the preset anti-oscillation threshold. When this occurs, indicating a slight net benefit insufficient to cover execution friction costs (such as the cumulative disruption to warehouse passage efficiency caused by frequent inventory transfers), the system will also block the instruction. Only when... Only then will the system determine that the warehouse transfer operation is sufficiently economically justifiable and execute it. Taking the example of a continuous process, if a candidate solution is expected to save only 20 seconds of total picking time, then the time gain... The cost unit is 20, while the cost of transferring inventory remains 20 cost units. The system directly intercepts it.

[0101] It should also be noted that, in obtaining the change in potential energy for warehouse transfer corresponding to candidate storage location transfer plans, the previous process also included dynamically updating the threshold:

[0102] Extract the distribution feature data from the current order data, and compare the divergence of the distribution feature data with the historical order distribution features on which the spatiotemporal feature latent vector table is based to obtain the data divergence offset.

[0103] When the detected data divergence offset is greater than the preset divergence threshold, the threshold update mechanism is triggered, which performs an exponential multiplication operation between the preset anti-oscillation threshold and the data divergence offset to increase the value of the preset anti-oscillation threshold.

[0104] It should be noted that the dynamic threshold update mechanism is a robust defense against concept drift scenarios and is one of the key innovations of this invention. Connecting to step two above, the spatiotemporal feature latent vector table is generated in the cloud based on historical order distribution and is updated every 4 hours. If, between two updates, the distribution characteristics of the real-time order stream shift drastically (for example, during the "Double Eleven" promotion, the order frequency of some products increases by tens of times, while the spatiotemporal feature latent vector table still encodes the normal distribution information before the promotion), the reliability of the scheduling decisions made by the edge side based on the outdated latent vectors will decrease. In this case, if the system still uses the original low threshold... Allowing database migration may result in a large number of invalid migration operations based on error messages.

[0105] The formula for calculating the dynamic threshold update is:

[0106] ;

[0107] in, The initial value for the preset anti-vibration threshold (e.g., 15 cost units). Current real-time order distribution The distribution of historical orders relied upon during cloud-based training The KL divergence (Kullback-Leibler Divergence) between the distributions indicates that the larger the KL divergence, the more severe the distribution shift. This is a penalty coefficient (typically between 1.0 and 3.0), controlling the sensitivity of the threshold increase. The physical meaning of this formula is: when the data distribution is stable ( )hour, The threshold remains unchanged; when the data distribution undergoes a drastic shift ( )hour, The exponential increase forces the system to raise the threshold for data transfer, becoming more conservative and reducing or stopping data transfer operations until the implicit vector table is updated in the next 4-hour cloud update cycle, at which point the new implicit vector table adapts to the current order distribution. The threshold then drops back to normal levels.

[0108] For example, throughout the example scenario, let's assume that in everyday situations... (Stable distribution), take ,but The cost unit and threshold only increased slightly, having almost no impact on normal scheduling. However, once the "Double Eleven" promotion began, the real-time order distribution changed drastically. If it jumps to 2.5, then Cost unit, at this point almost all inventory transfer schemes If none of the operations exceed the threshold, the system automatically freezes the database transfer operation and waits for an update from the cloud. After 4 hours, once the cloud completes a new round of implicit vector table updates and distributes them... It fell back to 0.1. Cost unit, system resumes normal scheduling.

[0109] Furthermore, the calculation process for expected congestion time is as follows:

[0110] Obtain the current set of concurrent scheduling instructions contained in the underlying device's working status information, and parse the current set of concurrent scheduling instructions to extract the planned travel path data of each executing device.

[0111] All planned driving route data are subjected to tensor superposition operations in a preset two-dimensional grid coordinate system to generate a global traffic heat map matrix representing the occupancy density of each node in the warehouse road network.

[0112] The spatial transfer paths corresponding to the candidate cargo location transfer schemes are mapped to the global traffic heat map matrix, and the density feature values ​​corresponding to each grid node crossed by the spatial transfer path are extracted.

[0113] Based on the preset delay estimation kernel function, a weighted summation operation is performed on the numerical values ​​of each density feature to obtain the expected congestion time, which represents the penalty cost in the time dimension.

[0114] It should be noted that the calculation of expected congestion time incorporates the impact of warehouse relocation operations on warehouse traffic into the cost assessment system. This ensures that the system considers not only the "physical cost of warehouse relocation itself" but also the "traffic interference caused by warehouse relocation to other ongoing scheduling tasks" when making decisions. The global traffic heatmap matrix divides the warehouse floor into fine-grained two-dimensional grids (e.g., 0.5m × 0.5m) and overlays all currently executing AGV paths onto the grid, generating an AGV pass density value for each grid node. Higher density indicates greater congestion in the area.

[0115] For example, in the scenario, there are currently 6 AGVs performing scheduling tasks, and their planned travel paths cover the grid nodes of the main warehouse road from area A to area C. The AGV travel path corresponding to the candidate solution Move(SKU_A,Loc_12,Loc_15) needs to pass through the main road in area B, where the grid node density is 3.2 (meaning that 3.2 AGVs / unit of time pass through this node). After weighted summation using the delay estimation kernel function, the expected congestion time is... Second.

[0116] like Figure 2 The diagram shown is a structural schematic of an order feature-driven adaptive warehouse location prediction system provided in an embodiment of this application, comprising:

[0117] The cloud-based graph computation module is used to acquire historical order logs, SKU basic attribute data, and storage location topology status data from the warehousing system. It constructs a spatiotemporal heterogeneous graph containing product nodes, order nodes, and storage location nodes on the cloud server. The spatiotemporal heterogeneous graph represents product co-occurrence relationships, order association relationships, and storage location spatial adjacency relationships. On the cloud server, the spatiotemporal heterogeneous graph is input into a spatiotemporal graph neural network for periodic iterative computation to extract global-dimensional product-order-storage location association features, generate a spatiotemporal feature latent vector table, and asynchronously distribute the spatiotemporal feature latent vector table to the edge computing nodes.

[0118] The edge-side state parsing module is used by edge-side computing nodes to receive real-time incoming current order data, extract the target product features corresponding to the current order data, and retrieve the corresponding feature data from the spatiotemporal feature latent vector table based on the target product features. This data is then concatenated with the current order data to generate a current state vector representing the current scheduling environment.

[0119] The edge-side policy inference module is used to input the current state vector into a pre-deployed reinforcement learning policy network for inference and output the target cargo location transfer probability distribution matrix.

[0120] The physical topology mask filtering module is used to obtain the working status information of the underlying equipment and the spatial carrying capacity limit information of the warehousing system in real time, and generate a physical topology action mask matrix with the same dimension as the target storage location transfer probability distribution matrix. The physical topology action mask matrix and the target storage location transfer probability distribution matrix are multiplied element by element to filter out illegal transfer probabilities that violate physical safety constraints, and the operation path corresponding to the highest effective probability is extracted as a candidate storage location transfer scheme.

[0121] The potential energy judgment and instruction scheduling module is used to obtain the change in transfer potential energy corresponding to the candidate storage location transfer plan. If the change in transfer potential energy is greater than the preset anti-vibration threshold, a scheduling instruction is generated and issued for execution based on the candidate storage location transfer plan.

[0122] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the above functions can be divided into different functional modules to complete all or part of the functions described above.

[0123] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A warehouse location adaptive prediction method driven by order features, characterized in that, Includes the following steps: The system acquires historical order logs, SKU basic attribute data, and storage location topology status data from the warehousing system. It then constructs a spatiotemporal heterogeneous graph on a cloud server, which includes product nodes, order nodes, and storage location nodes. The spatiotemporal heterogeneous graph represents product co-occurrence relationships, order association relationships, and storage location spatial adjacency relationships. In the cloud server, the spatiotemporal heterogeneous graph is input into the spatiotemporal graph neural network for periodic iterative calculations to extract global-dimensional product-order-location association features, generate a spatiotemporal feature latent vector table, and asynchronously distribute the spatiotemporal feature latent vector table to the edge computing nodes. The edge computing node receives real-time incoming current order data, extracts the target product features corresponding to the current order data, and retrieves the corresponding feature data from the spatiotemporal feature latent vector table based on the target product features. It then concatenates the target product features with the current order data to generate a current state vector representing the current scheduling environment. The current state vector is input into a pre-deployed reinforcement learning policy network for inference, and the target cargo location transfer probability distribution matrix is ​​output. The system acquires real-time information on the working status of the underlying equipment and the spatial carrying capacity limit conditions of the warehousing system, and generates a physical topology action mask matrix with the same dimension as the target storage location transfer probability distribution matrix. The physical topology action mask matrix and the target storage location transfer probability distribution matrix are multiplied element by element to filter out illegal transfer probabilities that violate physical safety constraints, and the operation path corresponding to the maximum effective probability is extracted as a candidate storage location transfer scheme. Obtain the change in transfer potential energy corresponding to the candidate storage location transfer scheme. If the change in transfer potential energy is greater than the preset anti-vibration threshold, generate a scheduling instruction based on the candidate storage location transfer scheme and issue it for execution.

2. The order feature-driven adaptive prediction method for warehouse location as described in claim 1, characterized in that: The specific process of constructing a spatiotemporal heterogeneous graph containing product nodes, order nodes, and storage location nodes on a cloud server is as follows: The order node is generated by extracting the order identifier from the historical order log, the product node is generated by extracting the product identifier from the SKU basic attribute data, and the storage location node is generated by extracting the coordinate identifier from the storage location topology status data. Based on the product information contained in the historical order logs, order association edges with corresponding preset weights are constructed between the order nodes and the product nodes that have an inclusion relationship; Based on the coordinate distance in the cargo location topology status data, construct spatial adjacency edges with corresponding preset weights between the cargo location nodes whose spatial distance is less than a preset distance threshold; The order node, the product node, the storage location node, the order association edge, and the spatial adjacency edge are merged to generate the spatiotemporal heterogeneous graph.

3. The order feature-driven adaptive prediction method for warehouse location as described in claim 1, characterized in that: The process of inputting the spatiotemporal heterogeneous graph into a spatiotemporal graph neural network for periodic iterative calculations to extract global-dimensional product-order-location association features and generate a spatiotemporal feature latent vector table is as follows: The spatiotemporal graph neural network performs a spatial feature aggregation operation on the spatiotemporal heterogeneous graph, aggregating the feature data of adjacent nodes to the central node to obtain a spatial topological feature vector. Temporal convolution operation is performed on the spatial topology feature vector to extract the temporal evolution features of the historical order log within a preset time window, resulting in a multi-dimensional fusion feature vector; The multidimensional fusion feature vector is subjected to dimensionality reduction processing to obtain dense feature vectors that correspond one-to-one with each commodity identifier and each cargo location identifier, and the dense feature vectors are summarized to construct the spatiotemporal feature latent vector table.

4. The order feature-driven adaptive prediction method for warehouse location as described in claim 1, characterized in that: The specific process for generating the current state vector representing the current scheduling environment is as follows: The target product identifier to be processed is parsed from the target product features; Retrieve the target product latent vector corresponding to the target product identifier and the starting storage location latent vector corresponding to the current starting storage location of the target product from the spatiotemporal feature latent vector table; The order urgency value, the target product latent vector, and the starting storage location latent vector in the current order data are concatenated and merged according to a preset dimension order to generate the current state vector with a unified dimension.

5. The order feature-driven adaptive prediction method for warehouse location as described in claim 1, characterized in that: The specific process for generating the physical topology action mask matrix with the same dimension as the target cargo location transfer probability distribution matrix is ​​as follows: The underlying equipment's working status information is analyzed to obtain the current occupancy status, and the space carrying capacity limit information is analyzed to obtain the remaining carrying capacity attribute of the target cargo space; The remaining load-bearing attributes are compared with the product weight or volume parameters in the SKU basic attribute data; If the comparison result is that the carrying attribute is exceeded, or the current occupancy status is that it is already occupied, then the value zero, which represents illegality, is assigned to the corresponding position of the matrix to be generated. If the comparison result is consistent with the carrying attribute, and the current occupancy status is idle, then a value representing a valid value, one, is assigned to the corresponding position in the matrix to be generated. Based on all the numerical zeros and ones, the physical topology action mask matrix composed of Boolean values ​​is generated.

6. The order feature-driven adaptive prediction method for warehouse location as described in claim 5, characterized in that: The operation path corresponding to the highest effective probability is extracted as a candidate cargo location transfer scheme, and the specific method is as follows: Perform the element-wise multiplication operation to directly set the transfer probability corresponding to the value of zero in the target cargo location transfer probability distribution matrix to zero, and obtain the intermediate filtering matrix. The non-zero transition probability values ​​in the intermediate filtering matrix are summed and normalized to obtain a normalized safety probability matrix in which the total probability is redistributed to the value of one. The target cargo location node with the highest probability value is found in the normalized safety probability matrix, and the data flow direction from the starting cargo location to the target cargo location node is determined as the candidate cargo location transfer scheme.

7. The order feature-driven adaptive prediction method for warehouse location as described in claim 1, characterized in that: The specific steps for obtaining the change in potential energy for warehouse transfer corresponding to the candidate warehouse location transfer scheme are as follows: The candidate location transfer scheme is input into a preset time benefit evaluation function to calculate the total picking time saved by the candidate location transfer scheme within a preset evaluation period, and multiplied by a preset benefit conversion coefficient to obtain the future outbound operation time benefit. Extract the equipment movement parameters from the working status information of the underlying equipment, calculate the number of scheduling times and expected congestion time required from the starting location to the target location, and multiply by a preset cost conversion coefficient to obtain the energy consumption cost of the mechanical equipment; The change in potential energy for warehouse transfer is calculated by subtracting the energy consumption cost of the mechanical equipment from the revenue generated from the future outbound operation time.

8. The order feature-driven adaptive prediction method for warehouse location as described in claim 1, characterized in that: The step of obtaining the change in transfer potential energy corresponding to the candidate storage location transfer scheme also includes dynamically updating the threshold: Extract the distribution feature data from the current order data, and compare the distribution feature data with the historical order distribution features on which the spatiotemporal feature latent vector table was generated to obtain the data divergence offset. When the data divergence offset is detected to be greater than the preset divergence threshold, a threshold update mechanism is triggered, which performs an exponential multiplication operation between the preset anti-oscillation threshold and the data divergence offset to increase the value of the preset anti-oscillation threshold.

9. The order feature-driven adaptive prediction method for warehouse location as described in claim 7, characterized in that: The calculation process for the expected congestion time is as follows: Obtain the current concurrent scheduling instruction set contained in the underlying device working status information, and parse the current concurrent scheduling instruction set to extract the planned driving path data of each executing device; All the planned driving route data are subjected to tensor superposition operation in a preset two-dimensional grid coordinate system to generate a global traffic heat map matrix representing the occupancy density of each node in the warehouse road network; Map the spatial transfer path corresponding to the candidate cargo location transfer scheme to the global traffic heat map matrix, and extract the density feature values ​​corresponding to each grid node crossed by the spatial transfer path; Based on a preset delay estimation kernel function, a weighted summation operation is performed on each of the density feature values ​​to obtain the expected congestion time, which represents the time dimension penalty cost.

10. A warehouse location adaptive prediction system driven by order features, characterized in that, include: The cloud-based graph calculation module is used to acquire historical order logs, SKU basic attribute data, and storage location topology status data of the warehousing system, and to construct a spatiotemporal heterogeneous graph containing product nodes, order nodes, and storage location nodes on the cloud server. The spatiotemporal heterogeneous graph represents the co-occurrence relationship of products, the association relationship of orders, and the spatial adjacency relationship of storage locations. In the cloud server, the spatiotemporal heterogeneous graph is input into the spatiotemporal graph neural network for periodic iterative calculations to extract global-dimensional product-order-location association features, generate a spatiotemporal feature latent vector table, and asynchronously distribute the spatiotemporal feature latent vector table to the edge computing nodes. The edge-side state parsing module is used by the edge-side computing node to receive real-time incoming current order data, extract the target product features corresponding to the current order data, retrieve the corresponding feature data in the spatiotemporal feature latent vector table according to the target product features, and concatenate it with the current order data to generate a current state vector representing the current scheduling environment. The edge-side policy inference module is used to input the current state vector into a pre-deployed reinforcement learning policy network for inference and output the target cargo location transfer probability distribution matrix. The physical topology mask filtering module is used to obtain the working status information of the underlying equipment and the spatial carrying capacity limit information of the warehousing system in real time, generate a physical topology action mask matrix with the same dimension as the target storage location transfer probability distribution matrix, and perform element-wise multiplication operation on the physical topology action mask matrix and the target storage location transfer probability distribution matrix to filter out illegal transfer probabilities that violate physical safety constraints, and extract the operation path corresponding to the maximum effective probability as a candidate storage location transfer scheme. The potential energy judgment and instruction scheduling module is used to obtain the change in transfer potential energy corresponding to the candidate storage location transfer scheme. If the change in transfer potential energy is greater than the preset anti-vibration threshold, a scheduling instruction is generated and issued for execution according to the candidate storage location transfer scheme.

Images