Correlation and proportion based ctu system dispersed storage method

Abstract

The application discloses a CTU system dispersion storage method based on correlation and proportion, which comprises collecting historical order data of an intelligent warehouse, containing the type and corresponding demand of required SKU every day, counting the specific coordinates of the warehouse layout and aisle and workbench of the intelligent warehouse, and calculating the distance list d of each aisle to the nearest workbench; processing the collected historical order data of the intelligent warehouse, using the confidence in the Apriori algorithm as the measurement of the correlation of two SKUs, calculating the confidence between all SKUs in the order, and obtaining the corresponding correlation matrix r; determining the proportion parameter matrix c corresponding to the ratio of the total demand of each two SKUs in all SKUs; based on the warehouse layout information, the parameter matrix r and the proportion parameter matrix c, the following optimization model is established, and the model is solved by using the simulated annealing-variable neighborhood search method. The application stores the SKUs in the same order on the same shelf, which is convenient for picking and improves the picking efficiency.

Description

Technical Field

[0001] This invention relates to the field of warehousing methods, and more particularly to a distributed storage method for a CTU system based on correlation and proportion. Background Technology

[0002] The CTU system is a "goods-to-person" intelligent warehouse, mainly composed of workbenches, shelves, and CTUs (intelligent robots). Workbenches serve as the warehouse's entrance and exit, and can be divided into inbound and outbound workbenches. Shelves are used for storing SKUs (Stock Keeping Units). CTUs handle SKU handling tasks, including inbound, outbound, and return of surplus materials. Unlike traditional warehouses, when an order arrives, the Warehouse Management System (WMS) processes the order and assigns it to the corresponding CTU and workbench. The CTU then moves the target box to the workbench for manual retrieval or picking. After the manual work is completed, the CTU returns the box to its appropriate location. The main characteristics of the CTU system are its large shelf capacity, the ability of a single CTU to carry multiple boxes simultaneously, and its narrow aisles, allowing only one CTU to pass at a time.

[0003] The centralized storage strategy used by enterprises does not take into account the correlation of SKUs, which will lead to two problems: CTUs that need the same SKU will be concentrated together, resulting in congestion; SKUs in the same order are stored in distant locations, resulting in a long total journey for a single CTU to pick all the required SKUs, resulting in low efficiency. Summary of the Invention

[0004] This invention aims to at least partially solve one of the technical problems in related technologies. To this end, one objective of this invention is to propose a distributed storage method for CTU systems based on correlation and proportion, which uses historical data and warehouse layout to determine storage location allocation. This method mainly consists of two parts: using historical data to measure SKU correlation and estimate proportion parameters, and combining these parameters with the warehouse layout to establish and solve an optimization model.

[0005] According to the present invention, a distributed storage method for CTU systems based on correlation and proportion is proposed, the method steps are as follows:

[0006] S1: Collect historical order data from the smart warehouse, including the types of SKUs required daily and their corresponding quantities. Compile a warehouse layout diagram, including the coordinates of aisles and workstations, and calculate the distance d from each aisle to the nearest workstation. i This represents the distance from the i-th shelf to the nearest workbench;

[0007] S2: Process the historical order data collected from the smart warehouse, use the confidence level in the Apriori algorithm as a measure of the correlation between two SKUs, calculate the confidence level between all SKUs in the order, and obtain the corresponding correlation matrix r;

[0008] S3: Determine the proportional parameter matrix c corresponding to the ratio of the total demand of every two SKUs in all SKUs;

[0009] S4: Based on warehouse layout information, parameter matrix r, and proportional parameter matrix c, the following optimization model is established:

[0010]

[0011]

[0012]

[0013]

[0014]

[0015]

[0016]

[0017] Where p ki Let z represent the number of bins for SKU I on shelf k. kij Q represents the maximum number of packages that can be stored proportionally for SKU I and SKU J on shelf k. i q represents the number of bins required to meet the demand for SKU I. i d represents the quantity of SKUI in a bin. k B represents the Manhattan distance of shelf k from the nearest workbench. k This represents the capacity of shelf k;

[0018] S5: Use the SAVNSP algorithm to solve the optimization model in step S4. First, select the corresponding shelves and storage capacity based on the total SKU demand. Then, generate an initial feasible solution and continuously perform variable neighborhood search until the specified number of iterations is reached to obtain the shelf number vector K. d and its corresponding capacity vector delete The elements that are 0 and K d Elements at the same position in the middle;

[0019] S6: After the overall location on the shelf is determined, shelving begins. First, an initial solution is generated, followed by iterative processing based on simulated annealing and variable neighborhood search. The algorithm for generating the initial solution is as follows:

[0020] Initialize matrix P init Let k be |K′| rows, |I| columns, and all elements be 0. Let k = 0.

[0021] while k<|K′|do:

[0022] Let the available capacity of shelf k be b←B′[k];

[0023] Initialize the SKU count vector on the shelf as slf←M[k,:];

[0024] while b > 0 do:

[0025] If sum(slf) == 0, then:

[0026] Select the SKU with the highest weighted relevance:

[0027] else then:

[0028] if|{i|i∈I,Q[i]>0,P[k,i]==0}|>0:

[0029] Prioritize SKUs that are not yet available:

[0030]

[0031] else then:

[0032] Calculate the constrained degree vector:

[0033] Select the SKU with the highest number of restrictions:

[0034] Update the k-th row of P: P[k,i] select ]←P[k,i select +1.

[0035] Update remaining demand: Q[i] select ]←Q[i select +1.

[0036] Update remaining capacity: b←b-1

[0037] Change to the next shelf: k←k+1;

[0038] Based on the initial solution generation algorithm, the preliminary results for the shelf are obtained: storing the quantity matrix P. init ;

[0039] S7: To maximize the effective relevance of the uploaded results, the following simulated annealing-variable neighborhood algorithm is used to iterate the results:

[0040] Initialize the number of iterations and the initial temperature: iter←0,T←T start

[0041] Obtain the initial solution P init ←sol_init(K′,B′,Q,s)

[0042] Define the current solution: P cur ←P init ,P best ←P init

[0043] Calculate the objective function value: f cur ←f(P cur ),f best ←f(P best )

[0044] while iter<max_iter and T>T end do:

[0045] Replication matrix: P vns ←P cur .copy()

[0046] Randomly select two shelves K from {1, 2, ..., |K'|} s ←[k1,k2]

[0047] Calculate the total quantity of each SKU on two shelves: Q′←[Q′1,Q′2,…,Q′ |I| ], where Q′ i ←P vns [k1,i]+P vns [k2,i]

[0048] Find the non-zero column of Q′ and denote its column index as .

[0049] I s Corresponding to the new storage quantity Q s ←Q′[I s ]

[0050] Calculate the quantity stored on each shelf: b1←B′[k1], b2←B′[k2]

[0051] Solving the relaxed model P s ←MODEL(Q s ,I s (b1,b2)

[0052] Integerization of relaxation solution P z ←INTEGER(P s ,Ks ,Q′,B′)

[0053] Use P s Replace P cur Lines k1 and k2: P vns [k1]←P s [1],P vns [k2]←P s [2].

[0054] Find the objective function value: f vns ←f(P vns )

[0055] if f vns >f cur then:

[0056] P cur ←P vns ,f cur ←f vns

[0057] if f cur >f best then

[0058] P best ←P cur ,f best ←f cur

[0059] else do

[0060]

[0061] P cur ←P vns ,f cur ←f vns

[0062] T←T×rec

[0063] iter←iter+1.

[0064] Preferably, the method for integerizing the relaxation solution in step S7 is as follows:

[0065] Initialize integer matrix P z For shape and P s They are the same, but all their elements are 0.

[0066] For P s Round the first line down to the nearest integer and assign it to P. z [1]:P z [1]←Floor(P s [1])

[0067] The first row requires an additional number of bins to meet the capacity constraint.

[0068] diff←B′[K s [1]]-sum(P z [1])

[0069] For P s The first row modulo 1 (i.e., the decimal part): mod ← P z [1]mod1

[0070] Sort the remainders in descending order and record their corresponding index before sorting:

[0071] mod idx ← arg max(mod)

[0072] while i≤diff do

[0073] P z [1,mod idx[i]]←P z [1, mod idx[i]] + 1

[0074] i←i+1

[0075] The remaining bins were allocated to the second shelf: P z [2]←Q′-P z [1].

[0076] Preferably, the method for selecting the appropriate shelf and storage capacity based on the total SKU demand in step S5, and then generating an initial feasible solution, is as follows:

[0077]

[0078]

[0079]

[0080]

[0081] make This is the actual capacity used in the order. If the value is non-zero, the shelf is used, and the solution of this model is the initial feasible solution for subsequent iterations.

[0082] Preferably, the method of continuously changing the neighborhood search until a specified number of iterations is as follows:

[0083] Initialize the shelf using a capacity vector The length is the same as the number of shelves, and all elements are 0;

[0084] Calculate the total number of bins required to meet the total demand.

[0085] Arrange the shelves according to d k Sort by size from smallest to largest, the corresponding shelf number vector is K. d ;

[0086] Circular access to K d element k in i ,if but otherwise And end the loop;

[0087] Obtain the shelf number vector K d and its corresponding capacity vector delete The elements that are 0 and K d Elements at the same position in the text.

[0088] Preferably, in step S2, the confidence level in the Apriori algorithm is used as a measure of the correlation between the two SKUs. The method steps for calculating the confidence level between all SKUs in the order and obtaining the corresponding correlation matrix r are as follows:

[0089] S21: The demand for each SKU each day is taken as the order for that day, and the set of SKU types with non-zero demand in the order is taken as a data set;

[0090] S22: Support is the number of times the target SKU appears in the dataset, and its expression is as follows:

[0091]

[0092] S23: Confidence level. The confidence level of SKU I and SKU J is the proportion of data sets containing SKU J that contain SKU I, which is the ratio of the support of SKU I and SKU J to the support of SKU I. The calculation method is as follows:

[0093]

[0094] S24: Calculate the confidence scores among all SKUs in the order and obtain the corresponding correlation matrix r, where

[0095] r ij =P(SKU) j |SKU j ).

[0096] Preferably, the method for determining the scaling parameter matrix c in step S3 is as follows:

[0097] S31: Select orders that contain both SKU1 and SKU2 from all orders to form an order set O. ij ;

[0098] S32: For order set O ij The sum of the quantities of SKU I and SKU J is s. i and s j ,but

[0099] The beneficial effects of this invention are:

[0100] (1) SKUs in the same order will be stored on the same shelf, which is convenient for picking and improves picking efficiency;

[0101] (2) Since the SKUs on the shelves are stored proportionally, the relevance of the SKUs on the shelves will not gradually degrade with picking, making the system more stable;

[0102] (3) When the same SKU is stored on multiple shelves and multiple CTUs need the same SKU, the target shelves of the CTUs can be different, thereby reducing congestion;

[0103] (4) SKUs are stored on the shelves closest to the outbound workstation, which can reduce the travel distance and time of CTU picking and increase picking efficiency. Attached Figure Description

[0104] In the attached diagram:

[0105] Figure 1 This is a flowchart of a distributed storage method for a CTU system based on correlation and proportion proposed in this invention;

[0106] Figure 2 A bar chart showing the total number of shelf visits for the three shelf-up strategies proposed in this invention under different instances;

[0107] Figure 3 This is a line graph showing the decline rate of shelf access frequency proposed in this invention. Detailed Implementation

[0108] Reference Figure 1 The distributed storage method for CTU systems based on correlation and proportion provided by this invention is applicable to "goods-to-person" smart warehouses with similar characteristics to CTU systems (such as shelf capacity being much larger than the capacity of intelligent handling robots). This method is based on historical data and warehouse layout to solve the storage location allocation problem, and mainly consists of two parts: using historical data to measure SKU correlation and estimate proportion parameters, and combining the parameters and warehouse layout to establish and solve an optimization model.

[0109] I. Data Preparation: This method requires two types of data: historical order data and warehouse layout data.

[0110] (1) Historical order data of the smart warehouse, including the types of SKUs required each day and the corresponding demand.

[0111] (2) The warehouse layout diagram of the smart warehouse, including the specific coordinates of the aisles and workbenches, and the calculated distance d from each aisle to the nearest workbench (where d... i (where i is the distance from the i-th shelf to the nearest workbench).

[0112] II. Measurement of Correlation and Estimation of Proportion Parameter: The confidence level from the Apriori algorithm is used as a measure of the correlation between the two SKUs. The confidence level is determined by the support of the frequent items. The support and confidence level are determined as follows:

[0113] (1) First, the demand for each SKU each day is taken as the order for that day, and the set of SKU types with non-zero demand in the order is taken as a data set.

[0114] (2) Support is the number of times the target SKU appears in the dataset, and its expression is as follows:

[0115]

[0116] (3) Confidence (Relevance): The confidence of SKU I and SKU J is the proportion of data sets containing SKU J in the data set containing SKU I, that is, the ratio of the support of SKU I and SKU J to the support of SKU I. The calculation method is as follows:

[0117]

[0118] (4) Calculate the confidence scores among all SKUs in the order and obtain the corresponding correlation matrix r, where

[0119] r ij =P(SKU) j |SKU j )

[0120] III. The ratio parameter is the ratio of the total demand for two SKUs. The ratio matrix c is estimated using its elements c. ij Obtain it by following these steps:

[0121] (1) Select orders that contain both SKUI and SKUJ from all orders to form an order set O. ij .

[0122] (2) For the order set O ijThe sum of the quantities of SKU I and SKUJ is s. i and s j ,but

[0123] IV. Based on the warehouse layout information and the estimated parameter matrices r and c, the following optimization model is established:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129]

[0130]

[0131] Among them, the decision variable is: p ki Let z be the number of SKUI bins on shelf k; kij The maximum number of packages to be stored proportionally for SKU I and SKU J on shelf k;

[0132] Other variables: Q i Number of bins required to meet the demand for SKU I; q i The quantity of SKU I in a bin; d k Let B be the Manhattan distance between shelf k and the nearest workbench; k Let k be the capacity of the shelf.

[0133] V. Since this problem is an NP (Nondeterminism Polynomial) difficult problem, the SAVNSP algorithm is used to solve this model. SAVNSP decomposes the problem into two parts: first, select the appropriate shelf and storage capacity according to the total demand of SKUs, then generate an initial feasible solution, and continuously perform variable neighborhood search until the specified number of iterations is reached.

[0134] (1) The first part is the selection of shelves and the capacity used, and the corresponding optimization model is:

[0135]

[0136]

[0137]

[0138]

[0139] The meaning of the symbols here remains unchanged; the solution can be used directly, and let... This is the actual capacity used in the order. If the value is non-zero, the shelf is used, and the solution of this model is the initial feasible solution for subsequent iterations.

[0140] (2) This model selects the shelf closest to the workbench that can meet the order capacity, and can be selected step by step by d. k The smallest shelf will be selected until the required capacity for the order is met; specifically, shelf K will be chosen. d and determine capacity The steps are as follows:

[0141] Initialize the shelf using a capacity vector The length is the same as the number of shelves, and all elements are 0.

[0142] Calculate the total number of bins required to meet the total demand.

[0143] Arrange the shelves according to d k Sort by size from smallest to largest, the corresponding shelf number vector is K. d .

[0144] Circular access to K d element k in i ,if but otherwise And then end the loop.

[0145] Obtain the shelf number vector K d and its corresponding capacity vector delete The elements that are 0 and K d Elements at the same position in the text.

[0146] VI. After determining the overall location on the shelf, the shelving process begins. First, an initial solution is generated, followed by iterative steps based on simulated annealing and variable neighborhood search. The algorithm for generating the initial solution is as follows:

[0147] Initialize matrix P init It has |K′| rows and |I| columns, with all elements being 0.

[0148] Let k = 0

[0149] while k<|K′|do:

[0150] Let the available capacity of shelf k be b←B′[k]

[0151] Initialize the SKU quantity vector on the shelf: slf←M[k,:]

[0152] while b > 0 do:

[0153] If sum(slf) == 0, then:

[0154] Select the SKU with the highest weighted relevance:

[0155] else then:

[0156] if|{i|i∈I,Q[i]>0,P[k,i]==0}|>0:

[0157] Prioritize SKUs that are not yet available:

[0158]

[0159] else then:

[0160] Calculate the constrained degree vector:

[0161] Select the SKU with the highest number of restrictions:

[0162] Update the k-th row of P: P[k,i] select ]←P[k,i select +1.

[0163] Update remaining demand: Q[i] select ]←Q[i select +1.

[0164] Update remaining capacity: b←b-1

[0165] Change to the next shelf: k←k+1

[0166] This algorithm yields the initial results for the product listing: storing the quantity matrix P. init .

[0167] VII. To maximize the effective relevance of the uploaded results, the following simulated annealing-variable neighborhood algorithm is used to iterate the results:

[0168] Initialize the number of iterations and the initial temperature: iter←0,T←T start

[0169] Obtain the initial solution P init ←sol_init(K′,B′,Q,s)

[0170] Define the current solution: P cur ←P init ,P best ←Pinit

[0171] Calculate the objective function value: f cur ←f(P cur ),f best ←f(P best )

[0172] while iter<max_iter and T>T end do:

[0173] Replication matrix: P vns ←P cur .copy()

[0174] Randomly select two shelves K from {1, 2, ..., |K'|} s ←[k1,k2]

[0175] Calculate the total quantity of each SKU on two shelves: Q′←[Q′1,Q′2,…,Q′ |I| ],in

[0176] Q′ i ←P vns [k1,i]+P vns [k2,i]

[0177] Find the non-zero column of Q′ and denote its column index as .

[0178] I s Corresponding to the new storage quantity Q s ←Q′[I s ]

[0179] Calculate the quantity stored on each shelf: b1←B′[k1], b2←B′[k2]

[0180] Solving the relaxed model P s ←MODEL(Q s ,I s (b1,b2)

[0181] Integerization of relaxation solution P z ←INTEGER(P s ,K s ,Q′,B′)

[0182] Use P s Replace P cur Lines k1 and k2: P vns [k1]←P s [1],P vns [k2]←P s [2].

[0183] Find the objective function value: f vns ←f(P vns )

[0184] if f vns >f cur then:

[0185] P cur ←P vns ,f cur ←f vns

[0186] if f cur >f best then

[0187] P best ←P cur ,f best ←f cur

[0188] else do

[0189]

[0190] P cur ←P vns ,f cur ←f vns

[0191] T←T×rec

[0192] iter←iter+1

[0193] In this process, MODEL is used as a relaxed optimization model. By continuously optimizing the distribution of SKUs between any two shelves, the objective function value is improved. The simulated annealing criterion in step 21 is used to accept the difference in solutions and prevent falling into local optima.

[0194] The main steps of integerizing the relaxed solution obtained by the variable neighborhood search are as follows:

[0195] Initialize integer matrix P z For shape and P s They are the same, but all their elements are 0.

[0196] For P s Round the first line down to the nearest integer and assign it to P. z [1]:P z [1]←Floor(P s [1])

[0197] The first row requires an additional number of bins to meet the capacity constraint.

[0198] diff←B′[K s [1]]-sum(P z [1])

[0199] For P s The first row modulo 1 (i.e., the decimal part): mod ← P z [1]mod1

[0200] Sort the remainders in descending order and record their corresponding indices before sorting:

[0201] mod idx ← arg max(mod)

[0202] while i≤diff do

[0203] P z [1,mod idx[i]]←P z [1, mod idx[i]] + 1

[0204] i←i+1

[0205] The remaining bins were allocated to the second shelf: P z [2]←Q′-P z [1]

[0206] This algorithm sorts and adds the solutions to the shelf based on the remainders of the optimal solution when rounded down, thereby making the integer solutions as close as possible to the relaxed solutions, and thus making its objective function closer to the optimal solution.

[0207] Example 1:

[0208] Numerical simulation experiments based on real data

[0209] Using historical order data from the CTU system for 90 days, from June 19, 2022 to September 19, 2022, data from June 19, 2022 to August 19, 2022 (60 days) was used for parameter estimation, and the remaining 30 days of data were used for numerical experiments. First, the last 30 days of data were numbered, with each number representing the number of days from August 19, 2022, to the instance's execution date. Instances 11, 22, and 29 had very small total SKU demand, requiring only one shelf to complete a full day's storage; therefore, their shelf access counts were the same and not worth discussing, and will not be shown in the numerical experiments. Only the experimental results for the remaining 27 instances are presented here.

[0210] Since the model only shows the distribution of bins on the shelves, the efficiency differences between different strategies cannot be intuitively seen before picking. Therefore, this paper proposes a greedy order matching algorithm to assign the intelligent robot to select shelves and retrieve the corresponding bins. During the picking process of all orders within a day, the number of times the intelligent robot visits the shelves is recorded, and the total number of shelf visits during the picking process is used to measure the efficiency of the storage location allocation strategy. The higher the number of shelf visits for a storage location allocation strategy, the lower its efficiency, and vice versa.

[0211] Depend on Figure 2 As shown, considering proportional distributed storage, the number of shelf visits is the fewest during each day's picking process; the centralized storage strategy has the highest number of shelf visits; and the distributed storage strategy without proportional consideration falls in between. This result indicates that the distributed storage strategy considering proportionality provides the highest picking efficiency in this warehouse, followed by distributed storage without proportional consideration, and centralized storage is the worst. Figure 3 As can be seen, in 27 instances, distributed storage considering proportions is more than 25% more efficient than centralized storage, with a maximum efficiency of 58%. Even though distributed storage without considering proportions takes into account SKU relevance and outperforms the centralized storage strategy, distributed storage considering proportions can still improve picking efficiency by more than 20% compared to distributed storage without considering proportions.

Claims

1. A distributed storage method for a CTU system based on correlation and proportion, characterized in that, The method steps are as follows: S1: Collect historical order data from the smart warehouse, including the types of SKUs required daily and their corresponding quantities. Compile a warehouse layout diagram, including the coordinates of aisles and workstations, and calculate the distance d from each aisle to the nearest workstation. i This represents the distance from the i-th shelf to the nearest workbench; S2: Process the historical order data collected from the smart warehouse, use the confidence level in the Apriori algorithm as a measure of the correlation between two SKUs, calculate the confidence level between all SKUs in the order, and obtain the corresponding correlation matrix r; S3: Determine the proportional parameter matrix c corresponding to the ratio of the total demand of every two SKUs in all SKUs; S4: Based on warehouse layout information, parameter matrix r, and proportional parameter matrix c, the following optimization model is established: Where p ki Let z represent the number of bins for SKU I on shelf k. kij Q represents the maximum number of packages that can be stored proportionally for SKU I and SKU J on shelf k. i q represents the number of bins required to meet the demand for SKU I. i d represents the quantity of SKUI in a bin. k B represents the Manhattan distance of shelf k from the nearest workbench. k This represents the capacity of shelf k; S5: Use the SAVNSP algorithm to solve the optimization model in step S4. First, select the corresponding shelves and storage capacity based on the total SKU demand. Then, generate an initial feasible solution and continuously perform variable neighborhood search until the specified number of iterations is reached to obtain the shelf number vector K. d and its corresponding capacity vector delete The elements that are 0 and K d Elements at the same position in the middle; S6: After the overall location on the shelf is determined, shelving begins. First, an initial solution is generated, followed by iterative processing based on simulated annealing and variable neighborhood search. The algorithm for generating the initial solution is as follows: Initialize matrix P init Let k = 0, where k' is the number of rows and I is the number of columns, and all elements are 0. while k<|K'|do: Let the available capacity of shelf k be b←B'[k]; Initialize the SKU count vector on the shelf as slf←M[k,:]; while b > 0 do: If sum(slf) == 0, then: Select the SKU with the highest weighted relevance: else then: if|{i|i∈I,Q[i]>0,P[k,i]==0}|>0: Prioritize SKUs that are not yet available: else then: Calculate the constrained degree vector: Select the SKU with the highest number of restrictions: Update the k-th row of P: P[k,i] select ]←P[k,i select +1. Update remaining demand: Q[i] select ]←Q[i select +1. Update remaining capacity: b←b-1 Change to the next shelf: k←k+1; Based on the initial solution generation algorithm, the preliminary results for the shelf are obtained: storing the quantity matrix P. init ; S7: To maximize the effective relevance of the uploaded results, the following simulated annealing-variable neighborhood algorithm is used to iterate the results: Initialize the number of iterations and the initial temperature: iter←0,T←T start Obtain the initial solution P init ←sol_init(K',B',Q,s) Define the current solution: P cur ←P init ,P best ←P init Calculate the objective function value: f cur ←f(P cur ),f best ←f(P best ) while iter<max_iter and T>T end do: Replication matrix: P vns ←P cur .copy() Randomly select two shelves K from {1, 2, ..., |K'|} s ←[k1,k2] Calculate the total quantity of each SKU on two shelves: Q'←[Q′1,Q′2,…,Q′ |I| ], where Q′ i ←P vns [k1,i]+P vns [k2,i] Find the non-zero column of Q′ and denote its column index as . I s Corresponding to the new storage quantity Q s ←Q'[I s ] Calculate the quantity stored on each shelf: b1←B'[k1], b2←B'[k2] Solving the relaxed model P s ←MODEL(Q s ,I s (b1,b2) Integerization of relaxation solution P z ←INTEGER(P s ,K s ,Q',B') Use P s Replace P cur Lines k1 and k2: P vns [k1]←P s [1],P vns [k2]←P s [2]. Find the objective function value: f vns ←f(P vns ) if f vns >f cur then: P cur ←P vns ,f cur ←f vns if f cur >f best then P best ←P cur ,f best ←f cur else do P cur ←P vns ,f cur ←f vns T←T×rec iter←iter+1.

2. The distributed storage method for a CTU system based on correlation and proportion according to claim 1, characterized in that: The steps for integerizing the relaxation solution in step S7 are as follows: Initialize integer matrix P z For shape and P s They are the same, but all their elements are 0. For P s Round the first line down to the nearest integer and assign it to P. z [1]:P z [1]←Floor(P s [1]) The first row requires an additional number of bins to meet the capacity constraint. diff←B'[K s [1]]-sum(P z [1]) For P s The first row modulo 1 (i.e., the decimal part): mod ← P z [1]mod1 Sort the remainders in descending order and record their corresponding index before sorting: modidx ← argmax(mod) while i≤diff do P z [1,modidx[i]]←P z [1,modidx[i]]+1 i←i+1 The remaining bins were allocated to the second shelf: P z [2]←Q'-P z [1].

3. The distributed storage method for a CTU system based on correlation and proportion according to claim 1, characterized in that, The steps for selecting the appropriate shelves and storage capacity based on the total SKU demand in step S5, and then generating the initial feasible solution, are as follows: make This is the actual capacity used in the order. If the value is non-zero, the shelf is used, and the solution of this model is the initial feasible solution for subsequent iterations.

4. The distributed storage method for a CTU system based on correlation and proportion according to claim 3, characterized in that, The steps for continuously searching the variable neighborhood until a specified number of iterations are as follows: Initialize the shelf using a capacity vector The length is the same as the number of shelves, and all elements are 0; Calculate the total number of bins required to meet the total demand. Arrange the shelves according to d k Sort by size from smallest to largest, the corresponding shelf number vector is K. d ; Circular access to K d element k in i ,if but otherwise And end the loop; Obtain the shelf number vector K d and its corresponding capacity vector delete The elements that are 0 and K d Elements at the same position in the text.

5. A distributed storage method for a CTU system based on correlation and proportion according to claim 1, characterized in that, The steps for using the confidence level in the Apriori algorithm as a measure of the correlation between two SKUs in step S2, calculating the confidence level between all SKUs in the order, and obtaining the corresponding correlation matrix r are as follows: S21: The demand for each SKU each day is taken as the order for that day, and the set of SKU types with non-zero demand in the order is taken as a data set; S22: Support is the number of times the target SKU appears in the dataset, and its expression is as follows: S23: Confidence level. The confidence level of SKU I and SKU J is the proportion of data sets containing SKU J that contain SKU I, which is the ratio of the support of SKU I and SKU J to the support of SKU I. The calculation method is as follows: S24: Calculate the confidence scores among all SKUs in the order and obtain the corresponding correlation matrix r, where r ij =P(SKU j |SKU j )。 6. A distributed storage method for a CTU system based on correlation and proportion according to claim 5, characterized in that, The steps for determining the scaling parameter matrix c in step S3 are as follows: S31: Select orders that contain both SKUI and SKUJ from all orders to form an order set O. ij ; S32: For order set O ij The sum of the quantities of SKU I and SKU J is s. i and s j ,but

Images