Logistics carrier adaptive configuration method based on LightGBM and logistics data hybrid drive

Abstract

The invention relates to the field of intelligent manufacturing and warehouse logistics, and discloses a logistics carrier adaptive configuration method based on LightGBM and logistics data hybrid drive, comprising the following steps: preprocessing multi-source heterogeneous data to form a demand sequence and a feature data set, and constructing a LightGBM-based demand prediction model to obtain prediction data; based on the prediction data, constructing a multi-objective optimization decision model, and performing optimization solution by adopting NSGA-II; calculating a required minimum physical space through a particle swarm optimization algorithm to obtain an optimal storage layout; performing an iterative test through the discrete event simulation model to obtain AGV fleet configuration; identifying a flow bottleneck through a constraint theory to obtain an optimal order picking port configuration; according to the method, discrete event simulation and iterative optimization are carried out, optimal adaptive configuration meeting requirements is obtained, and through the steps, dynamic reconstruction and adaptive optimization of the storage unit and the key equipment in the logistics scene are achieved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and warehousing logistics, specifically to an adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data. Background Technology

[0002] In modern production and logistics systems, enterprises commonly face challenges such as diverse product varieties, small batch sizes, shorter delivery times, and increased demand volatility. In warehousing and internal logistics systems, factors such as rack size and layout, AGV fleet configuration, and the number and location of picking bays significantly impact system throughput and response time. Traditional design methods typically rely on experience and static calculations for one-time planning. Once production rhythms or order structures change, system performance can easily degrade significantly, making rapid reconfiguration and adaptive configuration of equipment and storage units difficult.

[0003] Traditional warehouse configurations typically rely on fixed rules and historical experience. However, when faced with constantly changing product demands, order volumes, and warehouse equipment, traditional methods often struggle to achieve rapid response and efficient resource utilization. Furthermore, existing equipment configuration optimization efforts often focus only on specific aspects, such as optimizing the number of AGVs or the rack layout, failing to integrate warehouse scale, equipment configuration, and order structure within a unified optimization framework. This leads to discrepancies between local and global optima, and the system struggles to adapt promptly to changing demands. Summary of the Invention

[0004] To address the technical problems of existing technologies, such as difficulty in achieving rapid response and efficient resource utilization, local and global discrepancies, and difficulty in timely adjustment when demand changes, this invention provides an adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data. The technical solution is as follows:

[0005] Step 1: Preprocess the multi-source heterogeneous data to form demand sequences and feature datasets, construct a demand prediction model based on LightGBM, and obtain prediction data;

[0006] Step 2: Based on the above prediction data, construct a multi-objective optimization decision model and use NSGA-II for optimization solution;

[0007] Step 3: Calculate the minimum physical space required using the particle swarm optimization algorithm to obtain the optimal warehouse layout;

[0008] Step 4: Iterative testing using a discrete event simulation model to obtain the AGV fleet configuration;

[0009] Step 5: Identify flow bottlenecks using constraint theory to obtain the optimal picking bay configuration;

[0010] Step 6: Based on the above optimal warehouse layout, AGV fleet configuration and optimal picking gate configuration, conduct simulation and iterative optimization to obtain the optimal adaptive configuration that meets the requirements.

[0011] Further, step 1 specifically involves: collecting multi-source heterogeneous data, including historical order data, product bills of materials, production plans, material physical properties, and equipment files; cleaning, handling missing values, removing outliers, unifying dimensions, and reconstructing time series data to form a demand sequence and feature dataset with SKU-time period granularity; when constructing the demand forecasting model, firstly dividing the dataset into training and validation sets, then selecting sales data after logarithmic transformation as the target variable, and using the LightGBM regression algorithm for model training; automatically performing feature selection during model training and evaluating the importance of each feature, selecting the top-ranked key features based on their importance; evaluating the performance of the model after training, adjusting and optimizing the key parameters of the demand forecasting model based on the evaluation results, and using the optimized model to predict key information in the future and outputting forecast data.

[0012] Further, step 2 specifically involves: receiving the predicted data, constructing a mathematical programming model, performing system-level configuration multi-objective modeling, using warehouse scale parameters, shelf layout parameters, AGV models and quantities, picking bays and workstations as decision variables, taking the minimum equipment investment and operating costs, the maximum system throughput and the minimum average order response time as objective functions, and taking warehouse building area, inventory capacity, equipment utilization rate and service level as constraints, constructing a system-level equipment and warehousing unit configuration multi-objective optimization decision model, and using NSGA-II for optimization solution.

[0013] Furthermore, step 3 specifically involves: using a particle swarm optimization algorithm to encode and solve the number of rows and columns of the shelves, the location of the aisles, and the setting of high-frequency areas. The fitness function is the weighted sum of the average AGV travel distance and the shelf cost. Each particle represents a layout scheme. By simulating the search behavior of particles in the solution space, their position and speed are continuously adjusted to obtain the optimal warehouse layout.

[0014] Further, step 4 specifically involves: establishing a vehicle selection rule base based on material weight, volume, and handling method; determining candidate AGV types; constructing a discrete event simulation model to simulate the order execution process under different fleet sizes; designing a fitness function based on order completion time, AGV utilization rate, and congestion rate; using the particle swarm optimization algorithm to search for combinations of vehicle numbers; encoding the number of vehicle numbers as particle positions; calling the discrete event simulation model for each candidate solution; statistically analyzing the number of tasks completed per unit time, average task completion time, and AGV utilization rate; and comprehensively considering operational performance and cost to obtain the AGV fleet configuration.

[0015] Further, step 5 specifically involves: identifying bottleneck processes through constraint theory, establishing a mixed-integer linear programming model with the picking gate opening status and order allocation as decision variables, and using Gurobi to solve the model to optimize the number and location of picking gates and obtain the optimal picking gate configuration.

[0016] Further, step 6 specifically involves mapping the optimal warehouse layout, AGV fleet configuration, optimal picking station configuration, and order allocation scheme to the Flexsim simulation environment for discrete event simulation. This allows for the acquisition of performance metrics within the target operating cycle, including throughput, average response time, resource utilization, and work-in-process. If the performance metrics do not meet the preset performance threshold, the continuous parameters in the multi-objective optimization decision model are adjusted based on Bayesian optimization. In each iteration, Bayesian optimization recommends new parameter combinations, the performance is simulated and evaluated, and the results are updated. The particle swarm optimization algorithm and discrete event simulation model are then iteratively called again until the performance metrics meet the requirements or the iteration limit is reached.

[0017] Furthermore, the update formulas for the particle velocity and position are as follows:

[0018] ,

[0019] ,

[0020] in, For the first The particle in the first Storage location distribution in the next iteration. For the first The speed of each particle For the first The optimal solution in the history of a single particle. This is the globally optimal solution in the history of the population. For inertial weights, , As a learning factor, , It is a random number between 0 and 1.

[0021] Furthermore, the function of the mixed-integer linear programming model is:

[0022] ,

[0023] in, For picking port The decision variable for whether to enable it. To open the picking port Fixed costs, For picking port Cost factor per unit load For picking port The amount of tasks processed within the optimization cycle.

[0024] Beneficial effects:

[0025] This invention directly uses demand forecasting results as input to the configuration model, enabling the resource allocation process to fully consider the dynamic changes and uncertainties in demand. This achieves seamless integration from the demand layer to the resource layer, more accurately reflecting the needs of actual logistics scenarios and thus improving the rationality and effectiveness of resource allocation. Regarding the comprehensiveness of resource allocation, this invention fully considers the joint configuration of multiple factors such as warehouse scale and layout, AGV fleet, and picking station resources. By comprehensively considering these factors, it avoids the problem of overall performance degradation caused by local optimization. When determining the warehouse scale and layout, this invention simultaneously considers the operating efficiency of the AGV fleet and the operational capacity of the picking station, ensuring that various resources can work collaboratively to achieve efficient operation of the overall logistics system. Through digital twin simulation and Bayesian optimization, the theoretical optimality of the model is coupled with the dynamic behavior of the system, achieving iterative optimization oriented towards actual operational performance. Attached Figure Description

[0026] Figure 1 This is a flowchart of an adaptive configuration method for logistics carriers driven by a hybrid approach of LightGBM and logistics data. Detailed Implementation

[0027] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0028] like Figure 1 As shown, the adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data of the present invention includes the following steps:

[0029] Step 1: Collect multi-source heterogeneous data, including historical order data, product bill of materials, production plans, material physical properties, and equipment records. Clean the multi-source heterogeneous data, handle missing values, remove outliers, unify dimensions, and reconstruct time series data to form a demand sequence and feature dataset with SKU-time period granularity, as shown below:

[0030] ,

[0031] ,

[0032] in, It is the first The input sequence for each time window has a dimension of . , This refers to the predicted frequency for subsequent times within the corresponding time window. When building the demand forecasting model, the dataset is first divided into training and validation sets. Then, sales data after logarithmic transformation is selected as the target variable. The LightGBM regression algorithm is used for model training. Feature selection is automatically performed during model training, and the importance of each feature is evaluated, selecting the top 15 key features based on their importance. After model training, its performance is evaluated using multiple evaluation metrics, including mean absolute error. Root mean square error Mean absolute percentage error and the coefficient of determination Based on these evaluation results, key parameters of the model, such as maximum depth and number of leaf nodes, are adjusted and optimized. The evaluation index formula is as follows:

[0033] ,

[0034] ,

[0035] ,

[0036] ,

[0037] in, Indicates the number of samples. Indicates the first A true sales figure. Indicates the first One predicted sales volume This represents the mean of the true values. Prevent division by zero. Predict key information over a future period using the optimized model and output predicted data.

[0038] Step 2: Receive the predicted data, construct a mathematical programming model, and perform system-level multi-objective modeling. Use warehouse scale parameters, rack layout parameters, AGV type and quantity, picking bay and workstation quantity as decision variables. The objective functions are minimizing equipment investment and operating costs, maximizing system throughput, and minimizing average order response time. Constraints include warehouse building area, inventory capacity, equipment utilization rate, and service level. Construct a system-level multi-objective optimization decision model for equipment and warehouse unit configuration, as shown below:

[0039] ,

[0040] ,

[0041] ,

[0042] in, These represent equipment investment and operating costs, system throughput, and average order response time, respectively. Represents the cost of shelving and facilities. For the cost of AGV carts, To minimize the cost of the picking area, the model aims to minimize equipment investment and operating costs, maximize system throughput, and minimize average order response time, while maintaining system flexibility and adaptability. NSGA-II (Non-dominated sorting genetic algorithm) is used for optimization. Individuals in the population are sorted non-dominated according to the objective function, and the crowding degree of each individual is calculated, as shown below:

[0043] ,

[0044] in, Representative of individuals The level of congestion, Individuals The maximum and minimum values ​​of the objective function Individuals The objective function values ​​of the next and previous neighbors.

[0045] Step 3: The particle swarm optimization algorithm is used to encode and solve the number of rows and columns of shelves, aisle positions, and high-frequency area settings. The fitness function is the weighted sum of the average AGV travel distance and shelf cost. Each particle represents a layout scheme. By simulating the search behavior of particles in the solution space, their positions and speeds are continuously adjusted to seek the optimal warehouse layout. The update formulas for particle speed and position are as follows:

[0046] ,

[0047] ,

[0048] in, For the first The particle in the first Storage location distribution in the next iteration. For the first The speed of each particle For the first The optimal solution in the history of a single particle. This is the globally optimal solution in the history of the population. For inertial weights, , As a learning factor, , A random number between 0 and 1. Simultaneously, a fitness function is set to optimize the warehouse layout to reduce AGV travel distance and improve racking cost-effectiveness, as shown below:

[0049] ,

[0050] in, The penalty coefficient is... The AGV travel distance. To reduce shelf costs, an adaptive inertia weight control mechanism and a probability mapping transformation mechanism were specifically designed. Through these two mechanisms, particles can dynamically adjust their exploration and convergence capabilities based on their real-time state during the search process. At the same time, the introduction of a reasonable degree of random perturbation factors can effectively prevent the search process from falling into a rigid mode, providing particles with the possibility of escaping local optima, thereby improving global search efficiency.

[0051] Step 4: Establish a vehicle selection rule base based on material weight, volume, and handling method; determine candidate AGV types; construct a discrete event simulation model to simulate the order execution process under different fleet sizes; and design a fitness function based on order completion time, AGV utilization rate, and congestion rate, as shown below:

[0052] ,

[0053] in, The penalty coefficient is... This represents the average task completion time. The number of AGVs is determined by the particle swarm optimization algorithm. Combinations of AGV numbers for each vehicle type are searched, and the number of each vehicle type is encoded as a particle position. For each candidate solution, the discrete event simulation model is called to calculate the number of tasks completed per unit time, the average task completion time, and the AGV utilization rate. The AGV fleet configuration is then obtained by comprehensively considering operational performance and cost.

[0054] Step 5: Identify bottleneck processes using constraint theory, establish a mixed-integer linear programming model with picking bay opening status and order allocation as decision variables, and solve it using Gurobi to optimize the number and location of picking bays, obtaining the optimal picking bay configuration. The function of the mixed-integer linear programming model is:

[0055] ,

[0056] in, For picking port The decision variable for whether to enable it. To open the picking port Fixed costs, For picking port Cost factor per unit load For picking port The amount of tasks processed within the optimization cycle.

[0057] Step 6: Map the optimal warehouse layout, AGV fleet configuration, optimal picking station configuration, and order allocation scheme to the Flexsim simulation environment for discrete event simulation. Obtain performance indicators within the target operating cycle, including throughput, average response time, resource utilization, and work-in-process. If the performance indicators do not meet the preset performance threshold, adjust the continuous parameters in the multi-objective optimization decision model based on Bayesian optimization. Select continuous parameters such as channel width and AGV maximum waiting time threshold as Bayesian optimization variables, and use a Gaussian process as a surrogate model.

[0058] ,

[0059] in, The average system response time. The total system cost, These are the response time weighting coefficient and the cost weighting coefficient, respectively. With average response time as the optimization objective, Bayesian optimization recommends new parameter combinations in each iteration. Simulation evaluates the performance and updates the parameters. The particle swarm optimization algorithm and discrete event simulation model are then iteratively called again until the performance indicators meet the requirements or the iteration limit is reached.

[0060] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for adaptive configuration of logistics carriers based on a hybrid approach driven by LightGBM and logistics data, characterized in that... Includes the following steps: Step 1: Preprocess the multi-source heterogeneous data to form demand sequences and feature datasets, construct a demand prediction model based on LightGBM, and obtain prediction data; Step 2: Based on the above prediction data, construct a multi-objective optimization decision model and use NSGA-II for optimization solution; Step 3: Calculate the minimum physical space required using the particle swarm optimization algorithm to obtain the optimal warehouse layout; Step 4: Iterative testing using a discrete event simulation model to obtain the AGV fleet configuration; Step 5: Identify flow bottlenecks using constraint theory to obtain the optimal picking bay configuration; Step 6: Based on the above optimal warehouse layout, AGV fleet configuration and optimal picking gate configuration, conduct simulation and iterative optimization to obtain the optimal adaptive configuration that meets the requirements.

2. The adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data as described in claim 1, characterized in that: Step 1 specifically involves: collecting multi-source heterogeneous data, including historical order data, product bills of materials, production plans, material physical properties, and equipment files; cleaning, handling missing values, removing outliers, unifying dimensions, and reconstructing time series data to form a demand sequence and feature dataset with SKU-time period granularity; when constructing the demand forecasting model, firstly dividing the dataset into training and validation sets, then selecting sales data after logarithmic transformation as the target variable, and using the LightGBM regression algorithm for model training; automatically performing feature selection during model training and evaluating the importance of each feature, selecting the top-ranked key features based on their importance; evaluating the performance of the model after training, adjusting and optimizing the key parameters of the demand forecasting model based on the evaluation results, and using the optimized model to predict key information in the future and outputting forecast data.

3. The adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data as described in claim 1, characterized in that: Step 2 specifically involves: receiving the predicted data, constructing a mathematical programming model, performing system-level configuration multi-objective modeling, using warehouse scale parameters, shelf layout parameters, AGV models and quantities, picking bays and workstations as decision variables, taking the minimum equipment investment and operating costs, the maximum system throughput and the minimum average order response time as objective functions, and taking warehouse building area, inventory capacity, equipment utilization rate and service level as constraints, constructing a system-level equipment and warehousing unit configuration multi-objective optimization decision model, and using NSGA-II for optimization solution.

4. The adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data as described in claim 1, characterized in that: Step 3 specifically involves: using a particle swarm optimization algorithm to encode and solve the number of rows and columns of the shelves, the location of the aisles, and the setting of high-frequency areas. The fitness function is the weighted sum of the average AGV travel distance and the shelf cost. Each particle represents a layout scheme. By simulating the search behavior of particles in the solution space, their position and speed are continuously adjusted to obtain the optimal warehouse layout.

5. The adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data as described in claim 1, characterized in that: Step 4 specifically involves: establishing a vehicle selection rule base based on material weight, volume, and handling method; determining candidate AGV types; constructing a discrete event simulation model to simulate the order execution process under different fleet sizes; designing a fitness function based on order completion time, AGV utilization rate, and congestion rate; using the particle swarm optimization algorithm to search for combinations of vehicle types; encoding the number of vehicle types as particle positions; calling the discrete event simulation model for each candidate solution; statistically analyzing the number of tasks completed per unit time, average task completion time, and AGV utilization rate; and comprehensively considering operational performance and cost to obtain the AGV fleet configuration.

6. The adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data as described in claim 1, characterized in that: Step 5 specifically involves: identifying bottleneck processes through constraint theory, establishing a mixed-integer linear programming model with the picking gate opening status and order allocation as decision variables, and using Gurobi to solve the model to optimize the number and location of picking gates and obtain the optimal picking gate configuration.

7. The adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data as described in claim 1, characterized in that: Step 6 specifically involves mapping the optimal warehouse layout, AGV fleet configuration, optimal picking station configuration, and order allocation scheme to the Flexsim simulation environment for discrete event simulation. This yields performance metrics within the target operating cycle, including throughput, average response time, resource utilization, and work-in-process. If these performance metrics do not meet the preset performance threshold, the continuous parameters in the multi-objective optimization decision model are adjusted based on Bayesian optimization. In each iteration, Bayesian optimization recommends new parameter combinations, the simulation evaluates the performance, and updates the results. The particle swarm optimization algorithm and discrete event simulation model are then called back for iterative calculations until the performance metrics meet the requirements or reach the iteration limit.

8. The adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data as described in claim 4, characterized in that: The formulas for updating the particle velocity and position are as follows: ， ， in, For the first The particle in the first Storage location distribution in the next iteration. For the first The speed of each particle For the first The optimal solution in the history of a single particle. This is the globally optimal solution in the history of the population. For inertial weights, , As a learning factor, , It is a random number between 0 and 1.

9. The adaptive configuration method for logistics carriers based on a hybrid drive of LightGBM and logistics data as described in claim 6, characterized in that: The function of the mixed-integer linear programming model is: ， in, For picking port The decision variable for whether to enable it. To open the picking port Fixed costs, For picking port Cost factor per unit load For picking port The amount of tasks processed within the optimization cycle.

Images