A steel storage location allocation optimization method fusing graph representation learning and improved meme algorithm

Abstract

This invention discloses a steel warehouse location allocation optimization method that integrates graph representation learning and an improved meme algorithm, belonging to the field of steel warehouse location management and intelligent allocation technology. The invention includes: constructing a SKU-attribute heterogeneous graph and using an improved Node2Vec algorithm to mine implicit business relationships between steel products, extracting semantic prior knowledge that can quantify "business similarity"; constructing a multi-objective optimization model that balances outbound efficiency and semantic regularity, and solving it using an improved meme algorithm that integrates semantic priors. This algorithm achieves fast and high-quality solutions to warehouse location allocation problems under complex constraints through mechanisms such as cluster-guided heuristic initialization, SDE-based elite selection, dual adaptive crossover, and semantically aware directional mutation. It effectively solves the problems of data sparsity, objective conflict, and slow algorithm convergence in steel warehouse location allocation, realizing intelligent and efficient warehouse location allocation, and has broad practical application value.

Description

Technical Field

[0001] This invention belongs to the field of steel warehouse location management and intelligent allocation technology. More specifically, it relates to an optimization method for steel warehouse location allocation that integrates graph representation learning and improved meme algorithms in scenarios with sparse data and multi-objective conflicts. Background Technology

[0002] Large-scale steel warehousing, as a key hub in the supply chain, generally faces challenges such as high non-standardization of SKU (stock keeping unit) specifications, complex attribute combinations, and stringent constraints on heavy cargo stacking. Existing warehouse location allocation decisions rely heavily on the experience of on-site dispatchers, resulting in the spatial dispersion of related steel products, fragmented inventory structure, and consequently, increased subsequent repacking and outbound costs.

[0003] Currently, academia and industry face two major bottlenecks in solving this problem: First, the sparsity of data. Historical order data is sparse and exhibits a long-tail distribution, making it difficult for traditional statistical rule-based methods to uncover the implicit, deep business semantic relationships between steel SKUs (such as which different specifications of steel are frequently purchased in the same project). Second, multi-objective conflicts at the optimization level. Warehouse allocation needs to strike a balance between "improving outbound efficiency" and "maintaining the semantic regularity of inventory." Existing multi-objective evolutionary algorithms (such as NSGA-II) often face slow convergence speed, uneven solution set distribution, and a tendency to get trapped in local optima when dealing with such strongly constrained, multi-conflict problems.

[0004] A search revealed several patents related to warehouse location allocation. For example, Chinese patent application number 202510405443.7, filed on April 2, 2025, discloses an intelligent allocation method and system for bulk cargo warehouse locations in ports. This method includes a process of "historical data collection—constraint design—constructing a multi-objective model using a genetic algorithm—genetic pre-screening to generate initial feasible solutions that satisfy hard constraints—improved NSGA-II solution—and warehouse location allocation based on the optimization results." This approach aims to improve warehouse location utilization, balance loads, and reduce outbound costs in various port cargo scheduling scenarios. By introducing greedy pre-screening and multi-objective evolutionary solving, the feasibility and optimization efficiency of the initial solution can be improved to some extent. However, its potential drawbacks include: the lack of an association modeling mechanism for "data sparsity / cold start" scenarios, making it difficult to characterize the implicit business relationships between goods; the objective system focuses on short-term operating costs and load balancing, without explicitly constraining the long-term regularity of the inventory structure, which can easily lead to the fragmentation of the storage location structure over time; and the optimization process mainly relies on a general evolutionary framework, lacking deep enhancement strategies such as semantic prior guidance and local search, which may limit the convergence speed, uniformity of solution distribution, and interpretability of the solution.

[0005] For example, Chinese patent application number 202410679628.2, filed on May 29, 2024, discloses an optimization method for collaboratively solving warehouse location allocation and inbound route planning. This method includes: acquiring cargo data and candidate warehouse location information; calculating cargo priority and generating an initial warehouse location allocation solution accordingly; constructing a matching function to measure the degree of matching between cargo and warehouse location; constructing an objective function to evaluate the merits of the warehouse location allocation scheme; using simulated annealing to iteratively search the warehouse location allocation scheme to obtain multiple sets of candidate allocation results; performing inbound route planning for each set of candidate allocation results and calculating the total inbound distance; and finally selecting the warehouse location allocation result with the shortest total inbound path as the output. This patented method, by incorporating the inbound path distance into the result filtering, reduces the walking distance for inbound operations and improves the utilization rate of warehouse space to some extent. However, its potential drawbacks include: the target system focuses on minimizing the inbound distance, making it difficult to cover the more critical outbound efficiency and long-term operational costs such as stacking in steel warehousing; it lacks a cargo association modeling mechanism for scenarios with sparse historical order data, making it difficult to characterize implicit business relationships and cold-start semantic similarity between SKUs; the solution framework mainly relies on simulated annealing with single-solution iteration, lacking systematic control over the diversity and uniformity of Pareto solution sets, making it difficult to output multiple optional trade-off solutions simultaneously; and it does not reflect targeted handling of complex constraints such as load-bearing capacity, stacking stability, and crane topology in heavy-duty non-standard product scenarios. Summary of the Invention

[0006] The purpose of this invention is to address the problems of non-standard inventory units, sparse orders, difficulty in mining business relationships, and the conflict between outbound efficiency and spatial regularity in existing large-scale steel warehouses. It provides a steel warehouse location allocation optimization method that integrates graph representation learning and an improved meme algorithm. This method solves the relationship mining problem under sparse data by using heterogeneous graph representation learning to obtain semantic priors. Then, an improved meme algorithm that integrates semantic information is designed to improve solution efficiency and solution set quality through multiple collaborative mechanisms.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a method for steel storage location allocation that integrates graph representation learning and improved meme algorithms, comprising the following steps: Step S1: Obtain and preprocess historical order data and steel SKU attribute data from the steel warehousing system, and construct an SKU-attribute heterogeneity graph; Step S2: Apply the improved Node2Vec graph representation learning model with meta-path constraints to the SKU-attribute heterogeneous graph to perform feature learning, obtain low-dimensional semantic vectors of steel SKU attribute data, and generate semantic similarity matrix and semantic clusters by calculating the similarity between vectors, as semantic prior knowledge for subsequent decision-making. Step S3: Combining the warehouse physical topology and operational constraints, construct a weighted total travel objective function that minimizes the steel outbound operation. F 1 and the objective function of maximizing the semantic relevance of storage locations F 2. Multi-objective optimization model; Step S4: Solve the multi-objective optimization model constructed in step S3 using the improved meme algorithm that incorporates the semantic prior knowledge described in step S2, and output the Pareto optimal storage location allocation scheme to optimize the allocation of steel storage locations. The improved meme algorithm integrates a heuristic initialization mechanism based on semantic clustering, an elite selection mechanism based on SDE, a dual adaptive crossover mechanism, a semantically aware directional mutation mechanism, and a simulated annealing local search mechanism.

[0008] As a possible implementation of the first aspect of the present invention, in step S1, a heterogeneous graph G=(V,E) is defined, wherein the node set V includes SKU entity nodes and attribute concept nodes, and the attribute concept nodes include at least category nodes, material nodes and specification nodes; the edge set E includes associated edges connecting SKU nodes and their corresponding attribute nodes. The specific steps of step S2 are as follows: Step S2.1: Set the meta-path constraint to SKU-attribute-SKU, restricting the random walk process to only jump between SKU nodes with the same attribute, and generate a sequence of walk nodes; Step S2.2: Input the wandering node sequence into the Skip-gram model to train and obtain semantic vectors, denoted as SKU vectors; Step S2.3: Calculate the cosine similarity between any two SKU vectors, construct the semantic similarity matrix S, and use the K-Means algorithm to cluster all SKU vectors to obtain semantic clusters.

[0009] As a possible implementation of the first aspect of the present invention, in step S3, the weighted total travel objective function F1 for minimizing the steel outbound operation is calculated based on the average lateral and longitudinal running speed of the crane, the lateral deviation of the outbound channel, and the SKU heat weight with a time decay coefficient; the time decay coefficient is determined by the steel outbound heat half-life. The objective function F2 for maximizing the semantic relevance of storage locations is calculated by accumulating the semantic similarity of steel stored in adjacent storage locations.

[0010] As a possible implementation of the first aspect of the present invention, in step S4, the heuristic initialization mechanism processing flow includes: constructing cluster blocks using semantic clusters, generating initial individuals guided by cluster blocks according to a preset ratio, forcibly allocating steel belonging to the same semantic cluster to physically adjacent storage locations, and randomly generating the remaining proportion of individuals.

[0011] As a possible implementation of the first aspect of the present invention, in step S4, the processing flow of the elite selection mechanism of the SDE includes: Map the individuals of the population to the objective function space, and for each individual, translate its position according to its convergence relationship with the non-dominated front. The Euclidean distance between the translated individual and its nearest neighbor is calculated as the density estimate. Individuals with smaller density estimates are preferentially removed to maintain the uniformity of the solution set distribution in the high-dimensional target space while preserving convergence.

[0012] As a possible implementation of the first aspect of the present invention, in step S4, the crossover probability of the dual adaptive crossover mechanism is dynamically calculated by combining the evolutionary stage, the non-dominated sorting level and the global semantic isolation rate; the global semantic isolation rate is obtained by traversing the storage locations and calculating the complement of the local average semantic similarity, and is used to characterize the degree of semantic fragmentation of the storage location allocation scheme.

[0013] As a possible implementation of the first aspect of the present invention, in step S4, the processing flow of the semantically aware directional mutation mechanism includes: Traverse the chromosome and calculate the average semantic similarity between each gene locus and its physically adjacent library sites. If the similarity is lower than a set threshold, it is marked as a semantically isolated point. Based on the semantic similarity matrix, retrieve the top-K SKUs with the highest similarity to the isolated SKU. If they exist, perform targeted insertion and replacement; otherwise, perform random swapping.

[0014] As a possible implementation of the first aspect of the present invention, in step S4, the processing flow of the simulated annealing local search mechanism includes: screening the set of elite individuals with non-dominated ranking Rank 1 and constructing an energy function containing outbound travel, semantic relevance, and normalized weighted power penalty constraint violation degree; performing neighborhood perturbation on the elite individuals and accepting new solutions according to the Metropolis criterion to complete the fine-tuning within the feasible region; the constraint violation degree adopts the normalized weighted power penalty form to eliminate the difference in the dimensions of the upper limit of the storage location's load-bearing capacity and to apply a nonlinear penalty to overload behavior to ensure that the storage location allocation meets the load-bearing constraints.

[0015] The improved meme algorithm of this invention features optimized algorithm mechanism and operator design, making it more advantageous in multi-objective optimization of steel storage locations. Traditional multi-objective optimization algorithms (such as NSGA-II) are prone to failure of the crowding distance mechanism when facing high-dimensional objective spaces, resulting in uneven distribution of solution sets. The improved meme algorithm introduces an elite selection strategy based on SDE, comprehensively considering the convergence and sparsity of individuals, effectively improving the uniformity of Pareto solution set distribution. At the same time, the dual adaptive crossover strategy can dynamically adjust the crossover probability according to the evolutionary stage and individual quality, balancing global exploration and local exploitation, and avoiding premature convergence of the algorithm.

[0016] Furthermore, traditional evolutionary algorithms typically employ random swapping as their mutation operator, resulting in extremely low efficiency and a tendency to disrupt desirable structures. The improved meme algorithm utilizes a semantically aware directional mutation operator, leveraging a semantic similarity matrix mined through graph representation learning as prior knowledge, to purposefully repair "semantic outliers," significantly enhancing the algorithm's optimization efficiency for the "semantic regularity of storage locations" objective. Moreover, the embedded simulated annealing local search mechanism allows for fine-tuning of elite individuals, further improving the quality of the solution set. This approach, integrating domain knowledge (semantic prior) and intelligent optimization (meme algorithm), fully utilizes the implicit value of data to achieve precise solutions to storage location allocation problems under complex constraints.

[0017] A second aspect of the present invention also provides a steel warehouse location allocation optimization system that integrates graph representation learning and improved meme algorithms, comprising: The heterogeneous graph construction module is used to collect historical inbound and outbound orders and steel SKU attribute data to construct an SKU-attribute heterogeneous graph; The graph representation learning module is used to perform improved Node2Vec training with meta-path constraints, and outputs SKU semantic vectors, semantic similarity matrices and semantic clusters; The multi-objective model building module is used to combine warehouse topology and operational constraints to establish a steel warehouse location allocation optimization model that minimizes the weighted total outbound travel and maximizes the semantic correlation of warehouse locations. An improved meme algorithm solution module is used to perform semantic prior-based heuristic initialization, SDE elite selection, double adaptive crossover, semantically aware directional mutation, and simulated annealing local search, and output a Pareto optimal storage location allocation scheme. The storage location allocation output module is used to display and execute the final storage location allocation scheme that meets the load-bearing constraints and operation optimization objectives.

[0018] A third aspect of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of the present invention.

[0019] A fourth aspect of the present invention also provides a computer-readable storage medium for storing a computer program that, when run on a computer, causes the computer to perform the method described in the present invention.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention constructs a SKU-attribute heterogeneous graph oriented towards steel business characteristics by combining heterogeneous graph representation learning with multi-objective evolutionary computation, and mines SKU semantic vectors using an improved Node2Vec model, thereby effectively capturing implicit business associations between steel products under sparse data conditions and solving the cold start problem. At the same time, this method can extract high-quality semantic prior knowledge from limited historical orders, providing a solid data foundation for subsequent intelligent allocation.

[0021] (2) This invention creatively proposes an improved meme algorithm based on SDE elite selection and double adaptive crossover. This algorithm optimizes the location of steel storage in multiple objectives. Compared with the shortcomings of the traditional NSGA-II algorithm, such as slow convergence and uneven distribution of solution set, this invention maintains population diversity through SDE strategy and balances exploration and exploitation through double adaptive crossover, thereby effectively improving the convergence speed and solution set quality of the algorithm. Especially in the case of a significant conflict between efficiency and regularity objectives, it can also obtain a uniform and high-quality Pareto front.

[0022] (3) This invention designs a semantically aware directional mutation operator and integrates it into the evolution process. In the mutation operation, the semantic similarity matrix is ​​queried to repair the "semantic isolated points", which can significantly improve the semantic correlation of the storage location and effectively reduce the fragmentation of the inventory, thereby helping to reduce the subsequent turnover rate and operation cost. At the same time, the operator avoids blind random search and greatly improves the optimization efficiency of the algorithm.

[0023] (4) When performing refined optimization of Pareto frontier elite individuals, the present invention embeds a local search mechanism based on simulated annealing. By constructing a weighted energy function to guide the search direction, and using the Metropolis criterion to escape local optima, it is beneficial to further improve the accuracy of the solution set and ensure that the final output storage location allocation scheme has higher execution efficiency in actual operation. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall process of the warehouse location allocation optimization method of the present invention; Figure 2 This is a schematic diagram of the SKU-attribute heterogeneity graph in an embodiment of the present invention; Figure 3This is a schematic diagram illustrating the present invention of performing meta-path constraint random walks on SKU-attribute heterogeneous graphs to discover implicit business associations; Figure 4 This is a schematic diagram of the process using the improved meme algorithm of the present invention; Figure 5 This is a comparison chart of the convergence curves of the objective function F1 (minimizing outbound costs) in the embodiments of the present invention; Figure 6 This is a comparison chart of the convergence curves of the objective function F2 (maximizing semantic relevance) in the embodiments of the present invention; Figure 7 This is a comparison chart of the final Pareto front distribution in an embodiment of the present invention; Figure 8 This is a comparison chart of the marginal contributions of each improved module in the embodiments of the present invention; Figure 9 This is a comparison chart of the total outbound operation costs in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] This invention provides a steel warehouse location allocation optimization method that integrates graph representation learning and improved meme algorithms. Addressing the challenges of highly non-standard SKU (stock keeping unit) specifications, difficulties in business association mining due to sparse historical order data, and significant conflicts between outbound efficiency and spatial regularity objectives in warehouse location allocation within large steel warehouses, this invention first collects historical orders and steel type attributes to construct an SKU-attribute heterogeneous graph. On this heterogeneous graph, an improved Node2Vec algorithm incorporating meta-path constraints is used to learn SKU semantic vectors, generating a semantic similarity matrix and semantic clustering as priors. Finally, a method is established based on warehouse topology and operational constraints, aiming to minimize weighted outbound travel distances and maximize the semantic correlation of warehouse locations. The target is a multi-objective model; an improved meme solving strategy that integrates semantic priors is proposed: including a dual adaptive crossover mechanism based on individual and population characteristics to dynamically adjust the crossover probability and improve the efficiency of solution space exploration and high-quality gene recombination; and a semantically similarity-based semantic perception mutation mechanism to guide the directional rearrangement of storage locations, improving the effectiveness of mutation and semantic regularity; and further combining heuristic initialization, SDE elite selection and simulated annealing local search to achieve a fast and high-quality solution to the storage location allocation problem under complex constraints, and finally output a storage location allocation scheme that takes into account both high efficiency and high regularity.

[0027] The present invention will now be described in detail with reference to specific embodiments.

[0028] Example 1 Combination Figure 1 As shown, this embodiment provides a method for optimizing steel storage location allocation by fusing graph representation learning and improved meme algorithms, including the following steps: Step S1: Obtain and preprocess historical order data and steel SKU (stock keeping unit) attribute data from the steel warehousing system, and construct an SKU-attribute heterogeneous graph oriented towards the characteristics of steel business; The historical order data collected in this embodiment comes from the steel warehousing management system. The collected data includes at least the following fields: order number or work batch number, SKU identifier, quantity / weight, timestamp, work type (inbound / outbound), and picking unit / item (optional).

[0029] The collected steel SKU (stock keeping unit) attribute data includes steel attribute data and warehouse topology data. The steel attribute data includes at least the fields of category, material, and specifications (optionally including length, strength grade, surface condition, etc.). The warehouse topology data includes at least the storage area / aisle information, storage location number and its geometric / coordinate information, aisle / entrance / exit reference location, storage location capacity and load-bearing limit, mixed storage rules, and other constraint information.

[0030] The preprocessing of the collected raw data mainly includes the following steps: 1) Standardization and noise reduction: Unify the SKU representation method (preferably using the "category-material-specification" triple as the SKU primary key), handle missing and outlier values; unify the units and dimensions of the weight / quantity field; 2) Sample construction: Using orders (or outbound batches) as "co-occurrence containers", construct a co-occurrence set of SKUs within the same order / batch; 3) Data partitioning (optional, for vector quality verification): Stratified random sampling can be used to divide historical orders / batch into a training set (e.g., 80%) and a test set (e.g., 20%) to ensure that the distribution ratio of core categories and long-tail categories in the two sets is consistent, thereby avoiding semantic evaluation bias. In addition, a validation set can also be partitioned within the training set for parameter selection.

[0031] Combination Figure 2 To address the sparse nature of historical steel order data, this embodiment does not directly construct an SKU co-occurrence graph, but instead constructs an SKU-attribute heterogeneous graph, specifically including: Define a heterogeneous graph G=(V,E).

[0032] The node set V contains two types of nodes: SKU entity nodes, representing specific steel materials; and attribute concept nodes, including "category," "material," and "specification." The edge set E is defined as a "SKU-ownership-attribute" relationship.

[0033] Combination Figure 2 The structure of the constructed SKU-attribute heterogeneous graph is shown. For example, if a steel SKU (stock quantity unit) is rebar, then an edge is established between the "SKU node" and the "rebar node". Through this graph construction method, even if two SKUs have never appeared in the same historical order (cold start), as long as they share the same material or specifications, a connection is established in the graph through attribute nodes (second-order association).

[0034] In one specific embodiment, a steel warehouse, area A, was selected as the test object. The warehouse contains 12 storage locations, denoted as L1-L12. Each storage location is mapped using a two-dimensional grid, with L1-L4 in the first row, L5-L8 in the second row, and L9-L12 in the third row. Each storage location has a maximum load capacity of 10 tons. There are six SKUs to be assigned, denoted as SKU1-SKU6. SKU1 is Q355 H-beams, SKU2 is Q355 I-beams, SKU3 is Q355 medium-thick plates, SKU4 is Q235 channel steel, SKU5 is Q355 angle steel, and SKU6 is Q235 medium plates. In the order data from the past six months, SKU1, SKU2, and SKU3 frequently co-occur in multiple project orders. Therefore, in the heterogeneous graph, in addition to forming potential connections through order co-occurrence, semantic associations can also be formed by sharing the "Q355" material attribute.

[0035] Step S2: Apply the improved Node2Vec graph representation learning model with meta-path constraints to the SKU-attribute heterogeneous graph constructed in step S1 to perform feature learning and extract semantic prior knowledge. This step specifically includes the following process: Step S2.1: Based on the SKU-attribute heterogeneous graph G constructed in step S1, in order to capture paths with specific business meanings on the graph, this invention sets the meta-path constraint as "SKU-attribute-SKU". During the random walk, the sequence of walking nodes is strictly restricted to follow this pattern. For example... Figure 3 As shown, for example, starting from a SKU node, the next step must be to its attribute node, and the next step after that must be to return to another SKU node that possesses that attribute. This ensures that the wandering node sequence captures "business attribute similarity" rather than random structural connections. The length of the wandering node sequence can be preset according to the size of the SKU-attribute heterogeneous graph, for example, set to a preset number of steps (such as 40 to 80 steps), to ensure that the wandering node sequence can effectively cover small clusters of SKUs that are highly related in terms of business.

[0036] Step S2.2: Treat the wandering node sequence generated in step S2.1 as a "sentence" in natural language and input it into the Skip-gram (Word2Vec) model for training. In this embodiment, the vector dimension d=64 and the window size w=5 are set. After training, each SKU node obtains a 64-dimensional SKU vector. v i (Denotes it as SKU vector), which contains the characteristics of the steel in terms of business logic and captures the implicit business relationships.

[0037] Step S2.3: Calculate the cosine similarity between any two SKU vectors in all SKU vectors, and generate an N×N semantic similarity matrix. S Values ​​in the matrix S ij The larger the value, the better the quality of the steel. i and j Those who are more similar in business should be grouped together.

[0038] By combining the calculated cosine similarity between any two SKU vectors, the K-Means algorithm is used to cluster all SKU vectors, classifying all steel products into... K There are three core semantic clusters. Steel types within the same cluster exhibit high complementarity or similarity (e.g., the "rebar core cluster"). These clustering results will be used to guide the initialization of subsequent optimization algorithms.

[0039] Specifically, in this embodiment, the semantic similarity between some SKUs obtained after training is as follows: the similarity between SKU1 and SKU2 is 0.86, the similarity between SKU1 and SKU3 is 0.82, the similarity between SKU1 and SKU6 is 0.28, and the similarity between SKU4 and SKU6 is 0.74.

[0040] Based on the clustering results, SKU1, SKU2, SKU3, and SKU5 were assigned to the same semantic cluster C1, while SKU4 and SKU6 were assigned to a semantic cluster C2. This result indicates that SKUs made of Q355 material that frequently appear in similar orders are more semantically similar, and their corresponding semantic clusters can be directly used to guide the heuristic initialization operation based on clusters in the subsequent step S4.

[0041] Step S3: Construct a multi-objective optimization model for steel storage locations that integrates semantic similarity. The specific process is as follows: Step S3.1: Model the physical scene of the warehouse.

[0042] First, the warehouse's two-dimensional layout is gridded, with each grid cell serving as an allocatable storage location, and each storage location having storage location coordinates ( X pos , Y pos Attributes such as load capacity and other properties. Define the coordinates of the warehouse's outbound operation points (e.g., entrances). X channel .

[0043] The decision variables are encoded using real number permutation encoding, and the chromosome length is equal to the total number of steel pieces to be allocated. N Gene location corresponds to library order, and gene value corresponds to assigned SKU ID.

[0044] Step S3.2: Construct a multi-objective optimization model. This model includes core objective functions: minimizing the weighted total travel distance of steel outbound operations and maximizing the semantic correlation of storage locations. These are denoted as: objective function ( F 1): Minimize the weighted total travel distance for steel outbound operations, specifically expressed as:

[0045] in: Index for steel SKUs; This indicates the number of SKUs in the set; and These represent the horizontal and vertical coordinates of the storage location assigned to the current SKU (stock unit) in the warehouse's planar coordinate system. This indicates the lateral coordinates of the outbound passage (or the centerline of the main passage) in the coordinate system. This indicates the lateral deviation between the SKU storage location and the outbound channel; , To determine the direction transformation weights in the time dimension based on the mechanical operating characteristics of the overhead crane, satisfying .

[0046]

[0047]

[0048] In the formula, The average or rated speed of the vehicle (lateral movement). The average or rated speed of the vehicle (longitudinal); The popularity weight of SKU (stock keeping unit) is used to characterize the importance of the SKU in historical outbound shipments, and is obtained by exponential decay accumulation and normalization:

[0049]

[0050] for A collection of historical outbound timestamps. For set for The timestamp of any historical outbound record; The timestamp corresponding to the current moment. The difference is expressed in days and satisfies ; This is the time decay coefficient, used to control the rate at which historical data decays in heat. A larger value indicates that the heat decays faster over time. In practical applications, this time decay coefficient... This is not an arbitrarily assigned constant, but rather determined based on the business turnover cycle of a specific steel warehouse. Preferably, the calculation is performed using a mechanism that sets the half-life of steel outbound heat, with the specific formula as follows:

[0051] in, This is the half-life of the steel outbound demand (in this embodiment, the unit is days). For example, in this embodiment, if the average inventory turnover period of a warehouse is 30 days, it means that the guiding significance of historical orders from 30 days ago for current outbound shipments is reduced by half, then we take... Calculation This approach enables the algorithm to precisely couple the mathematical model with the actual physical lifecycle of warehousing. >0 represents a very small positive number, used to avoid the denominator being zero; Indicates in set Original popularity of all SKUs The maximum value; By minimizing the objective function ( This allows SKUs with higher popularity weights to be preferentially assigned to storage locations with smaller vertical coordinates and smaller horizontal deviations from the channel, thereby reducing the overall weighted outbound operation cost.

[0052] It should be noted that traditional algorithms for warehouse optimization typically use a simple geometric distance (such as Manhattan distance) multiplied by the historical outbound frequency.

[0053] The above mechanism has two major drawbacks: ① The actual handling of steel in a warehouse relies on overhead cranes (bridge cranes). The mechanical speed of the overhead cranes in the trolley direction (lateral) and the trolley direction (longitudinal) are completely different. The simple Manhattan distance cannot reflect the real physical time cost.

[0054] ② Traditional historical frequency Freq i It cannot handle long-tail steel products and is not sensitive to time (the frequency of shipments a year ago and the frequency of shipments yesterday have the same weight, which does not conform to the actual business cycle).

[0055] This invention designs an objective function. F 1. Introducing the speed of the overhead crane V x ,V y The directional transformation weights are designed to ensure the model closely fits the actual physical environment; a time decay factor is introduced. η This gives higher popularity weight to SKUs that have recently been shipped, automatically adapting to changes in the business cycle.

[0056] objective function ( F 2) Maximize the semantic relevance of storage locations, specifically expressed as:

[0057] in, This represents a set of physically adjacent (front, back, left, right) storage locations. Storage locations The steel stored above and its warehouse location The steel SKUs (stock keeping units) stored above. S The semantic similarity matrix generated in step S2.3 is calculated by summing the semantic similarity of steel stored in adjacent storage locations, and the matrix elements are... Defined as and semantic vectors Cosine similarity between them.

[0058] It should be noted that current methods for improving storage location regularity involve matching discrete attributes, typically using a simple indicator function: (1 point for the same type, 0 points otherwise) When using the aforementioned indicator function to handle non-standard large steel products, often the steel purchased in the same order is completely different in type and material (for example, a project requires both rebar and I-beams). Simply grouping them by type leads to the dispersion of related goods, greatly increasing the cost of subsequent sourcing and repacking. Therefore, this invention extracts a semantic similarity matrix through a heterogeneous graph (i.e., an improved Node2Vec walking mechanism). S This includes "implicit business relationships" that are not directly visible to the human eye (e.g., which items are frequently purchased together). Maximizing this objective function... F 2. This is equivalent to physically binding and storing goods that are highly related in terms of business, in order to reflect the spatial regularity of the warehouse layout and the degree of business aggregation.

[0059] In this implementation, it is assumed that the warehouse exit is located at the origin of the coordinate system, and the Manhattan distances from L1-L12 to the exit are 2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7 respectively; the outbound frequencies of SKU1-SKU6 in the past 30 days are 18, 15, 14, 7, 10, 6 respectively. If a certain scheme arranges the high-frequency SKU1, SKU2, and SKU3 in the near-end and adjacent storage locations such as L1, L2, and L5, then its Smaller and The frequency is relatively high; if the high-frequency and semantically similar SKUs are dispersed in remote storage locations such as L3, L8, and L12, the weighted outbound travel will increase, and the semantic correlation between adjacent storage locations will decrease.

[0060] Step S4: Solve the multi-objective optimization model using the KG-MNSGA-II algorithm, which incorporates semantic prior knowledge, and combine it with... Figure 4 This includes the following steps: Step S4.1: Perform heuristic initialization based on the semantic clusters obtained in step S2.3. When initializing the population, some SKUs in the same semantic cluster are preferentially allocated to physically adjacent storage locations to generate initial individuals with high semantic regularity; the remaining individuals are generated randomly to implant excellent gene fragments of semantic clustering into the initial population, thereby taking into account both the quality of the initial solution and the diversity of the population.

[0061] In this embodiment, the population size is set to 100, with 80% of the initial individuals generated based on semantic clustering guidance, and 20% generated randomly. For example, for cluster C1={SKU1, SKU2, SKU3, SKU5}, it is preferentially arranged in adjacent neighboring regions such as L1, L2, L5, and L6; cluster C2={SKU4, SKU6} is preferentially arranged in another adjacent region such as L9 and L10. This improves the feasibility and regularity of the initial population.

[0062] Step S4.2: Perform objective function calculation, non-dominated sorting, and SDE-based elite selection on the current population.

[0063] The objective function constructed based on step S3 (Weighted total outbound travel) and objective function (Library location semantic association) Calculate the bi-objective fitness value of each individual in the population, and perform non-dominated ranking of the population based on the fitness values ​​to divide it into tiers. Specifically, the elite selection process based on SDE is as follows: Map the population individuals to the objective function space. For each individual, shift its position according to its convergence relationship with the non-dominated front, so that the dominated individuals are away from the front. The Euclidean distance between the translated individual and its nearest neighbor is calculated as the density estimate. During the environment selection phase, individuals with smaller density estimates are preferentially eliminated to maintain convergence while preserving the uniformity of the solution set distribution in the high-dimensional target space. The SDE-based elite selection strategy of this invention can preferentially retain individuals that are both close to the Pareto front and conducive to a uniform distribution of the solution set.

[0064] S43. Execute the dual adaptive crossover operator. This invention designs a dynamic crossover probability. The calculation formula is as follows:

[0065] in: For the current population individuals to be crossovered (i.e., a certain storage allocation scheme). This represents the individual's rank in the current population's non-dominant ranking; The maximum number of levels in the current population; and These are the preset upper and lower bounds of the crossover probability, respectively, satisfying... ; It is a very small positive number, used to prevent the denominator from being 0; For individuals The global semantic isolation rate is used to characterize the degree of semantic fragmentation of the current storage location allocation scheme in physical space.

[0066] Specifically, the individual Global semantic isolation rate The calculation employs a continuous mapping method without a hard threshold, specifically including the following steps: (1) Calculate the local average semantic similarity: for individuals The represented inventory scheduling scheme iterates through all warehouse locations that have been allocated steel. Based on the semantic similarity matrix generated in step S2 (The matrix elements have been linearly mapped to) (Interval), calculate storage location The average semantic similarity between the steel material on the screen and other steel materials in its physical neighborhood :

[0067] in, For storage location The set of physically adjacent storage locations (e.g., front, back, left, and right). This represents the number of adjacent storage locations (3 or 5 if located at the edge of a storage location, and 8 if not located at the edge).

[0068] (2) Calculate the continuous semantic isolation degree: abandon the discrete hard threshold judgment and define the storage location. Continuous semantic isolation It is the complement of its average semantic similarity, that is:

[0069] because ,therefore The larger the value, the more likely it is to be stored in a warehouse. The weaker the business connection between the steel products on the platform and the surrounding steel products, the higher their degree of isolation.

[0070] (3) Calculate the global semantic isolation rate: This involves assigning individual... The consecutive semantic isolation degrees of all storage locations are summed and divided by the total number of steel SKUs to be assigned. The global semantic isolation rate of this scheme is obtained. :

[0071] The traditional crossover operator:

[0072] Traditional algorithms only consider the overall score of an individual, neglecting its internal structure. In complex storage allocation scenarios, good individuals often have well-organized "high-quality gene blocks" (i.e., physically adjacent and semantically similar steel). If these individuals are still subjected to random crossover with a high probability, the hard-won, well-developed structure can be easily destroyed, causing the algorithm to stagnate or fail to converge.

[0073] The dual adaptive crossover operator proposed in this invention dynamically adjusts the crossover probability by simultaneously utilizing "evolutionary stage information" and "individual level information." In the early stages of evolution, it increases the overall crossover probability to enhance global exploration capabilities; in the later stages, it decreases the overall crossover probability to protect superior structures. This method endows the algorithm with semantic awareness, particularly for high-level elite individuals (i.e., schemes with low semantic fragmentation). Smaller individuals (those with a regular structure and higher hierarchy) are assigned a lower crossover probability to lock in and protect these superior genes; lower-level individuals (those with severe fragmentation and poor performance) are assigned a higher crossover probability to balance the preservation of superior genes and the exploration of new solutions.

[0074] During crossover, sequential crossover or partial mapping crossover methods that maintain the feasibility of the permutation structure are preferred. After crossover, duplicate SKUs are eliminated and missing SKUs are filled in for the offspring to ensure that each SKU appears only once and each storage location corresponds to a valid allocation value.

[0075] Specifically, in this embodiment, it is set , , When the iteration reaches the 20th generation, the maximum number of generations is 200. For a given individual to be crossed... If its non-dominated level is Rank1, and its global semantic isolation rate is calculated based on the semantic similarity matrix, then... For another individual to be crossed If its non-dominated level is Rank3 and its global semantic isolation rate is... Substituting this into the above formula, we can see that the individual The crossover probability is lower than that of individuals. This indicates that for individuals with better hierarchical levels and more regular semantic structures, the algorithm tends to reduce crossover perturbation to protect the already formed good "semantic blocks"; while for individuals with poorer hierarchical levels or higher semantic fragmentation, it increases the crossover probability to enhance the recombination search ability and promote the breaking of invalid structures.

[0076] S44. Semantic-aware directed mutation operator: On the offspring population generated by the crossover operation, perform the mutation based on the semantic similarity matrix obtained in step S2. S This invention utilizes a directed mutation operation. Traditional mutation operators typically employ random swapping, lacking domain knowledge guidance and resulting in low optimization efficiency; this invention, however, leverages a semantic similarity matrix. S The provided prior business information guides the direction of mutation in a targeted manner. The specific execution logic is as follows: 1) Identification of isolated sites in chromosomes: Traverse the chromosome and calculate the average semantic similarity between each gene locus and its physically adjacent sites. If the similarity is below a set threshold, it is marked as a "semantic isolated site" (i.e., a misplaced item).

[0077] 2) Targeted insertion: For the identified isolated SKUs, in the semantic similarity matrix... S Find the top-K SKUs with the highest similarity to it; 3) Location swapping: With a high probability (e.g., 0.9), isolated SKUs are migrated to the physical nearest storage location of the Top-K related SKUs with the highest semantic similarity; otherwise, a random swap is performed. This mutation operator utilizes prior business knowledge and is a purposeful "repair" rather than a blind "trial and error."

[0078] S45. Simulated Annealing Local Search: The meme algorithm emphasizes the synergy between global and local approaches. After each generation of genetic operations, a local search is performed on the set of elite individuals with a non-dominant Rank1 in the current population for fine-tuning.

[0079] Simultaneously, to ensure the strict physical feasibility of the final output solution, the load-bearing constraints of the storage locations are explicitly introduced and addressed at this stage. Specifically, a weighted energy function including a penalty term is constructed:

[0080] in, This is the target balance coefficient, used to adjust the influence strength of the semantic relevance target; To constrain the weighting coefficient of the penalty item, this is used to strengthen the degree to which feasibility and load-bearing constraints are met. , (for example, take) =1000), used to enhance the degree of satisfaction of load-bearing constraints, satisfying Preferably ; To solve for the overall violation of the load-bearing constraints by x, a normalized weighted power penalty form is used for calculation:

[0081] in, For the collection of storage locations; This is the storage location weighting coefficient, used to reflect the importance of different storage locations, satisfying... Preferably ; Used to delineate warehouse locations The overload capacity is 0 when the storage location is not overloaded, and is the weight difference exceeding the upper limit of the load capacity when overloaded.

[0082] This is the penalty order, used to increase the intensity of the penalty for overload, to satisfy... Preferably ; It is a very small positive number, used to prevent the denominator from being zero and to improve numerical stability.

[0083] It should be noted that traditional multi-objective algorithms typically use a "death penalty" method to handle constraints (such as the storage space's load-bearing capacity limit) (once the load is exceeded). (Fitness is directly assigned to infinity and the cullery is eliminated), or a simple linear penalty method:

[0084] The load-bearing constraints of steel storage are extremely stringent. Simply eliminating all intermediate solutions with overload would lead to a sharp decrease in population diversity. Furthermore, a simple linear penalty cannot effectively distinguish between "slight overload" and "severe overload," and it fails to consider the dimensional issues arising from the capacity differences between different storage locations. Therefore, this invention adopts a modified normalized weighted power penalty form, which, on the one hand, divides by... The dimensional differences in the load-bearing caps of different storage locations were eliminated; on the other hand, a penalty order was introduced. It imposes a nonlinear, geometrically amplified penalty on severe overload behavior, while allowing a very small number of excellent individuals with slight overload to survive during the simulated annealing phase using the Metropolis criterion.

[0085] And for any storage location Introducing violation variables And satisfy:

[0086] in, ∈{0,1} means Should it be allocated to a storage location? Each storage location is assigned only one type. ; For SKU weight parameters, The upper limit of the load-bearing capacity of the storage location, the Used to delineate warehouse locations The overload capacity, when the storage space is not overloaded =0, when the storage space is overloaded. The weight difference, equal to or greater than the weight limit exceeding the maximum load capacity, is thus passed. Penalize overloaded solutions.

[0087] Then, a new solution is obtained by applying a neighborhood perturbation to the elite individuals. According to the Metropolis criterion, the Pareto front is refined within a finite number of steps. This yields a non-dominated elite solution set refined by simulated annealing local search, and outputs a warehouse location allocation scheme that satisfies the load-bearing constraints. This scheme reduces the weighted travel distance for outbound shipments while increasing the semantic association clustering degree of steel in adjacent warehouse locations.

[0088] For example: Suppose that after perturbing the neighborhood of a certain elite solution, the resulting new solution attempts to allocate a total of 12 tons of steel to storage location L1 (the upper limit of the load). (tons), then the overload of storage space L1 at this time Tons. Substituting this into the above penalty formula will produce... The violation value is then multiplied by a larger penalty coefficient. This will lead to a decrease in the total energy of the new solution. A surge.

[0089] After the algorithm terminates in step S4, a set of Pareto optimal solutions is output. Warehouse managers can select the optimal solution from these solutions to guide on-site operations based on their current operational preferences (efficiency or regularity).

[0090] Combination Figure 5 and Figure 6 As can be seen, compared with the benchmark algorithm (i.e., the NSGA-II algorithm), the improved meme algorithm of this invention shows excellent performance in algorithm convergence. Figure 5 The convergence curves of the baseline algorithm and the KG-MNSGA-II algorithm of this invention in terms of outbound cost are shown respectively. Figure 6 The convergence curves of semantic relevance of the baseline algorithm and the KG-MNSGA-II algorithm of this invention are shown respectively. The starting point of the algorithm of this invention is higher than that of the baseline algorithm (thanks to heuristic initialization), and the final convergence value is close to 95, which is significantly higher than that of the baseline algorithm, proving the effectiveness of the semantic-aware mutation operator.

[0091] Combination Figure 7 As can be seen, compared with the benchmark algorithm, the Pareto front distribution generated by the algorithm of this invention in the dual-objective space is of higher quality. The red point set (of this invention) clearly dominates the blue point set (of the benchmark algorithm), and its distribution range is wider and more uniform, verifying the advantage of the SDE-based elite selection strategy in maintaining the diversity of solution sets.

[0092] Furthermore, to analyze the marginal contribution of each improved module in this invention to the overall optimization effect, under the same historical order data, warehouse topology data, and parameter settings, a complete algorithm including all improved modules was used as a control group. The semantic clustering-based heuristic initialization module, the dual adaptive cross-multiplication module, the semantically aware directional mutation module, and the simulated annealing local search module were removed for comparative experiments. Statistical analysis was performed on the changes in objective function convergence performance, Pareto front distribution quality, and final solution quality for each comparison group, yielding the following results: Figure 8 The results show the comparison of the marginal contributions of each improved module.

[0093] Figure 8 This indicates that the heuristic initialization module primarily improves the quality of the initial solution and enhances the objective function. The initial level; the dual adaptive crossover module mainly enhances population recombination efficiency and accelerates the objective function. The convergence of the target function; the semantically aware directional mutation module mainly improves the semantic aggregation of adjacent storage locations, and improves the convergence of the target function. The improvement effect is even more significant; the simulated annealing local search module further improves the local refinement of the elite solution and the quality of the final Pareto front. This demonstrates that the various improved modules in this invention work synergistically at different stages, jointly improving the convergence speed and solution quality of the steel warehouse allocation scheme.

[0094] In this implementation, the backtracking simulation results based on real historical orders are as follows: Figure 9 As shown, compared to the actual historical operation, the total outbound operation cost optimized by this invention is significantly reduced, and the overall operation efficiency is improved by approximately 28.6%, further verifying the application value of the method of this invention in actual steel warehousing scenarios.

[0095] In summary, this invention effectively solves the problems of data sparsity, target conflict, and slow algorithm convergence in steel storage location allocation by constructing heterogeneous graph mining semantic priors and deeply integrating them into various aspects of the improved meme algorithm. It achieves intelligent and efficient storage location allocation and has broad practical application value.

[0096] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A steel yard location allocation method fusing graph representation learning and improved meme algorithm, characterized in that: Includes the following steps: Step S1: Obtain and preprocess historical order data and steel SKU attribute data from the steel warehousing system, and construct an SKU-attribute heterogeneity graph; Step S2: Apply an improved Node2Vec graph representation learning model with meta-path constraints to the SKU-attribute heterogeneous graph to perform feature learning and extract semantic prior knowledge. Step S3, constructing a multi-objective optimization model including a weighted total travel objective function of minimizing the steel material outbound operation F 1 and a semantic correlation degree objective function of maximizing the storage location F 2 Step S4: Solve the multi-objective optimization model constructed in step S3 using the improved meme algorithm that incorporates the semantic prior knowledge described in step S2, and output the Pareto optimal storage location allocation scheme to optimize the allocation of steel storage locations. The improved meme algorithm integrates a heuristic initialization mechanism based on semantic clustering, an elite selection mechanism based on SDE, a dual adaptive crossover mechanism, a semantically aware directional mutation mechanism, and a simulated annealing local search mechanism.

2. The method according to claim 1, characterized in that, In step S2, the specific steps are as follows: Step S2.1: Set the meta-path constraint to SKU-attribute-SKU, restricting the random walk process to only jump between SKU nodes with the same attribute, and generate a sequence of walk nodes; Step S2.2: Input the wandering node sequence into the Skip-gram model to train and obtain semantic vectors, denoted as SKU vectors; Step S2.3: Calculate the cosine similarity between any two SKU vectors and construct a semantic similarity matrix. S The K-Means algorithm was used to cluster all SKU vectors to obtain semantic clusters.

3. The method according to claim 1, characterized in that, In step S3, the objective function for minimizing the weighted total travel distance of the steel outbound operation is described. F 1. The calculation is based on the average lateral and longitudinal running speed of the vehicle, the lateral deviation of the outbound channel, and the SKU popularity weight with a time decay coefficient; the time decay coefficient is determined by the half-life of the steel outbound heat. The objective function for maximizing the semantic relevance of storage locations F 2. The semantic similarity of steel stored in adjacent storage locations is calculated by summing them up.

4. The method according to any one of claims 1-3, characterized in that, In step S4, the heuristic initialization mechanism process includes: constructing cluster blocks using semantic clusters, generating initial individuals guided by the cluster blocks according to a preset ratio, forcibly allocating steel belonging to the same semantic cluster to physically adjacent storage locations, and randomly generating the remaining proportion of individuals.

5. The method according to any one of claims 1-3, characterized in that, In step S4, the processing flow of the elite selection mechanism of the SDE includes: Map the individuals of the population to the objective function space, and for each individual, translate its position according to its convergence relationship with the non-dominated front. The Euclidean distance between the translated individual and its nearest neighbor is calculated as the density estimate. Individuals with smaller density estimates are preferentially removed to maintain the uniformity of the solution set distribution in the high-dimensional target space while preserving convergence.

6. The method according to any one of claims 1-3, characterized in that, In step S4, the crossover probability of the dual adaptive crossover mechanism is dynamically calculated by combining the evolutionary stage, the non-dominated sorting level, and the global semantic isolation rate; the global semantic isolation rate is obtained by traversing the storage locations and calculating the complement of the local average semantic similarity, and is used to characterize the degree of semantic fragmentation of the storage location allocation scheme.

7. The method according to any one of claims 1-3, characterized in that, In step S4, the processing flow of the semantically aware directional mutation mechanism includes: Traverse the chromosome and calculate the average semantic similarity between each gene locus and its physically adjacent library sites. If the similarity is lower than a set threshold, it is marked as a semantically isolated point. Based on the semantic similarity matrix, retrieve the top-K SKUs with the highest similarity to the isolated SKU. If they exist, perform targeted insertion and replacement; otherwise, perform random swapping.

8. The method according to any one of claims 1-3, characterized in that, In step S4, the processing flow of the simulated annealing local search mechanism includes: screening the set of elite individuals with non-dominated ranking Rank 1 and constructing an energy function that includes outbound travel, semantic relevance, and normalized weighted power penalty constraint violation; performing neighborhood perturbation on the elite individuals and accepting new solutions according to the Metropolis criterion to complete the fine-tuning within the feasible region; the constraint violation adopts the normalized weighted power penalty form to eliminate the difference in the dimensions of the upper limit of the storage location's load-bearing capacity and to apply a nonlinear penalty to overload behavior to ensure that the storage location allocation meets the load-bearing constraints.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that, when run on a computer, causes the computer to perform the method as described in any one of claims 1-8.

Images