Production logistics global modeling and joint optimization method based on demand prediction

Abstract

The invention discloses a production logistics global modeling and joint optimization method based on demand prediction, and the method employs a LightGBM model to carry out middle and short term periodic demand prediction, and considers the influence of time lag nonlinear factors on demands. A production-inventory strategy joint optimization model taking periodic replenishment as a core is constructed, and a (T, s, S) strategy framework is adopted to perform joint modeling on a safety inventory level s, a target inventory level S and a production quantity q in each period. The model adopts an improved self-adaptive differential evolution algorithm, minimizes total cost as a target function, comprehensively considers inventory holding cost, stockout cost, production cost and transportation cost, meanwhile, restrains the service level not to be lower than a preset standard, and realizes production-inventory coordinated optimization under the condition of prediction uncertainty. The whole-process production logistics model is constructed, the expandability is good, and the service level and the operation cost control level are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent manufacturing and supply chain management, and particularly relates to a production logistics global modeling and joint optimization method based on demand prediction. BACKGROUND

[0002] In modern manufacturing and logistics systems, enterprises usually face the operating characteristics of multi-variety, small batch, short delivery period, etc., demand volatility is significantly enhanced, production resources are limited, and inventory configuration complexity is improved. In actual operation, production planning and inventory strategy are often affected by uncertain factors, including time series changes, seasonality, promotion interference, etc. of demand, so that single-dimensional local optimization methods are difficult to ensure overall operating efficiency.

[0003] Common inventory control strategies include fixed cycle ordering strategy (T, s, S), random ordering strategy (s, S), etc. These strategies are mostly based on historical experience and theoretical model derivation, and usually rely on assumptions about demand distribution when setting parameters. However, real sales data often shows non-normality, periodicity and mutability, which makes the strategy based on theoretical distribution assumptions have problems such as error accumulation and adjustment lag in actual application.

[0004] In terms of data modeling, machine learning methods have been widely used in sales demand prediction modeling, especially the LightGBM model based on the gradient boosting framework, which can handle large-scale heterogeneous data and has advantages such as modeling non-linear relationships, automatic feature selection, and strong generalization ability. By combining sales history, product attributes, time characteristics and other multi-dimensional information, the prediction accuracy of short-term demand can be improved, providing data support for downstream production planning and inventory control.

[0005] However, although the prediction model has good performance, most current prediction and control modules are still decoupled in actual deployment, lacking joint optimization mechanisms. Production planning is usually based on fixed capacity configuration and preset cycles, and inventory strategy is replenished with static rules, resulting in poor coordination and response lag in the overall system. In particular, in the case of rapid demand fluctuations or resource shortages, it is easy to cause waste of production capacity, stockout or inventory accumulation, etc. SUMMARY

[0006] The present application aims to provide a production logistics global modeling and joint optimization method based on demand prediction to solve the above technical problems.

[0007] To solve the above technical problems, the specific technical scheme of the production logistics global modeling and joint optimization method based on demand prediction of the present application is as follows:

[0008] A production logistics global modeling and joint optimization method based on demand prediction, comprising the following steps:

[0009] Step one, data integration and preprocessing, association and integration of sales log, product information, manufacturing cost, transportation cost and other multi-source structured data, data feature normalization, data type conversion and missing value processing;

[0010] Step two, multi-level feature engineering and time series structure construction, through sliding window, lag variable, time period feature construction method, extract statistical features related to future demand change;

[0011] Step three, demand prediction model construction based on LightGBM, using LightGBM regression model to model daily sales, and rolling prediction output future four weeks sales for each product;

[0012] Step four, joint optimization model of production-inventory strategy is constructed, which is extended on the basis of fixed cycle order strategy (T, s, S), and controllable production q is introduced for joint optimization;

[0013] Step five, parameter optimization based on improved differential evolution algorithm SaDE, using SaDE to jointly optimize the inventory strategy parameters (s, S) and cycle plan production q of each week;

[0014] Step six, simulation and performance evaluation, rolling simulation for the next four weeks, output weekly service level, cost structure change, compared with fixed strategy, verify the effectiveness of dynamic strategy.

[0015] Further, the step one includes: according to the unique identifier Product_ID, the multi-source data is integrated, the daily sales record of each product is constructed, the time field in the data set is converted, and the type of the classification variable is marked. The missing value and abnormal value record are removed.

[0016] Further, the step two includes the following specific steps:

[0017] For each product, the lag feature of sales data is extracted, which represents the difference between two period lag sales and reflects the trend change. For each date t:

[0018] sold_lag k (t)=Sold_Num(t-k)

[0019] Wherein, k is the lag day (3, 7, 21), Sold_Num is the sales of that day,

[0020] Construct moving average time series, smooth the sales sequence, reduce short-term fluctuations, and the average sales of the past w days at the current time point t:

[0021]

[0022] Construct a moving standard deviation time series to represent the fluctuation amplitude of sales in the past w days:

[0023]

[0024] Further, the step three includes the following specific steps:

[0025] Construct the training set and validation set, and use the LightGBM regression model to train the log-transformed sales as the target variable. During the model training process, automatic feature selection and importance evaluation are performed, and the top 15 key features are retained for prediction;

[0026] The performance of the prediction model is evaluated using the mean absolute error MAE, root mean square error RMSE, mean absolute percentage error MAPE, and determination coefficient R 2 The evaluation index formulas are as follows:

[0027]

[0028] Where n represents the number of samples, y i represents the i-th real sales, represents the i-th predicted sales, represents the mean of the true value, and ∈ prevents division by zero.

[0029] Based on the trained model, daily rolling prediction is performed for the next four weeks to generate a daily sales prediction sequence. The prediction results are aggregated according to natural weeks to generate a demand prediction matrix for 4 consecutive natural weeks.

[0030] Further, the step four includes the following specific steps:

[0031] A fixed ordering cycle T = 7 days and lead time L = 3 are adopted. Within each cycle T, the reorder point s, target inventory level S, and cycle production q are defined:

[0032] s = D L + SS

[0033] S = s + Q

[0034] q = min(Q, Max_Capacity-current_inventory)

[0035] Where SS represents safety stock, Q represents economic order quantity, Max_Capacity represents the maximum inventory capacity of the product, and current_inventory represents the current inventory level.

[0036] The model's objective function design comprehensively considers multiple dimensions such as holding costs, stockout costs, production costs, and transportation costs. It also sets constraints on maximum production capacity, minimum order quantity, non-negative inventory levels, and safety stock fluctuation range. The optimization objective is to minimize total cost while maintaining service levels within these constraints.

[0037] Objective = Cost + Penalty

[0038] Where Cost represents the total cost and Penalty represents the service rate penalty.

[0039] Furthermore, step five includes the following specific steps:

[0040] The population is initialized, and an adaptive mutation strategy is used to adjust the differential mutation factor and crossover probability. The cost function and service level index of multiple solutions are evaluated in each generation. The process is iterated weekly, and the inventory and production quantities of each week are carried over to the next week.

[0041] During this period, the optimization objective is to minimize the total cost while ensuring that the service level is greater than the threshold of 0.75.

[0042] Furthermore, step six includes the following specific steps:

[0043] The optimization process is based on Monte Carlo simulation. Each strategy undergoes 200 Monte Carlo simulations based on the inventory replenishment logic to evaluate inventory dynamics, cost distribution, and service level performance under each set of s, S, and q parameters. The optimal strategy combination is searched through improved differential evolution. The optimization results are output weekly, including the weekly reorder point s, order cap S, planned production volume q for the cycle, and the corresponding cost and service level indicators. The formulas for total cost and service level are as follows:

[0044] Cost=StockoutCost+HoldingCost+ProductCost+ShippingCost

[0045]

[0046] The production logistics global modeling and joint optimization method based on demand forecasting proposed in this invention has the following advantages:

[0047] 1. Introduce a machine learning prediction model based on LightGBM and integrate the model into the inventory and production strategy optimization process. The prediction results are no longer supplementary information, but directly drive the control strategy.

[0048] 2. At the optimization level, an improved adaptive differential evolution algorithm is adopted to achieve joint optimization of the (s, S) inventory strategy parameters and production plan quantity under the condition of demand uncertainty. This method can quickly converge to a high-quality solution in a multi-objective and multi-constraint environment, and effectively balance the contradictory relationship between holding costs, stockout costs, production costs, and service levels. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the replenishment logic of a production logistics global modeling and joint optimization method based on demand forecasting in this embodiment;

[0050] Figure 2 This is a flowchart illustrating the joint optimization model of the production-inventory strategy in this embodiment. Detailed Implementation

[0051] To better understand the purpose, structure, and function of this invention, the following description, in conjunction with the accompanying drawings, provides a more detailed account of a production logistics global modeling and joint optimization method based on demand forecasting.

[0052] like Figure 1 and Figure 2 As shown in this embodiment, a method for comprehensive production logistics modeling and joint optimization based on demand forecasting includes the following steps:

[0053] Step 1: Data integration and preprocessing. This involves linking and integrating multi-source structured data such as sales logs, product information, manufacturing costs, and transportation costs of e-commerce FMCG products, and performing feature normalization, data type conversion, and missing value handling on the data.

[0054] Specifically, the multi-source data is integrated based on the unique identifier Product_ID to construct daily sales records for each product. The time field in the dataset is formatted and the categorical variables such as product number and warehouse number are labeled with their types. Missing values ​​and outlier records are removed.

[0055] Step two involves multi-level feature engineering and time series structure construction. This involves extracting statistical features related to future demand changes using methods such as sliding windows, lagged variables, and time period feature construction.

[0056] Specifically, for each product, the lagged characteristics of sales data (sales volume in the previous 3, 7, and 21 days) are extracted to represent the difference between the two periods of lagged sales volume, reflecting trend changes. For each date t:

[0057] sold_lag k (t) = Sold_Num(tk)

[0058] Where k is the lag days (3, 7, 21) and Sold_Num is the sales volume for that day.

[0059] Construct a moving average time series to smooth the sales data and reduce short-term fluctuations. At the current time t, the average sales over the past w days are:

[0060]

[0061] Construct a moving standard deviation time series to represent the fluctuation range of sales over the past w days:

[0062]

[0063] Step 3: Based on the LightGBM demand forecasting model, the LightGBM regression model is used to model daily sales and to generate rolling forecasts for each product for the next four weeks.

[0064] Specifically, a training set and a validation set are constructed, and the sales volume after logarithmic transformation is used as the target variable. The LightGBM regression model is used for training. During the model training process, feature selection and importance evaluation are performed automatically, and the top 15 key features are retained for prediction.

[0065] Using mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) 2 The performance of the prediction model is evaluated using metrics such as maximum depth and number of leaf nodes, and model parameters are adjusted for optimization. The formulas for the evaluation metrics are as follows:

[0066]

[0067]

[0068] Where n represents the number of samples, y i This represents the i-th actual sales volume. This represents the i-th predicted sales volume. This represents the true mean value, and ∈ prevents division by zero.

[0069] Based on the trained model, daily rolling forecasts are made for the next four weeks to generate a daily sales forecast sequence. The forecast results are then aggregated by natural week to generate a demand forecast matrix for four consecutive natural weeks.

[0070] Step four: Construct a joint optimization model for production-inventory strategies. This model extends the fixed-cycle ordering strategy (T, s, S) by introducing a controllable production quantity q for joint optimization.

[0071] Specifically, a fixed ordering cycle T = 7 days and a lead time L = 3 days are adopted. Within each cycle T, the reorder point s, the target inventory level S, and the cycle production quantity q are defined.

[0072] s=D L +SS

[0073] S = s + Q

[0074] q=min(Q,Max_Capacity-current_inventory)

[0075] Where SS represents safety stock, Q represents economic order quantity, Max_Capacity represents the maximum inventory capacity of the product, and current_inventory represents the current inventory level.

[0076] The objective function of the model is designed to take into account multiple dimensions such as holding costs, stockout costs, production costs and transportation costs. It also sets constraints such as maximum production capacity limits, minimum order quantities, non-negative inventory constraints, and safety stock fluctuation ranges. The optimization objective is to minimize total costs while keeping service levels within the constraints.

[0077] Objective = Cost + Penalty

[0078] Where Cost represents the total cost and Penalty represents the service rate penalty.

[0079] Step 5: Optimize parameters based on the improved differential evolution algorithm (SaDE). Use SaDE to jointly optimize the weekly inventory strategy parameters (s,S) and the cycle planned production quantity q.

[0080] Specifically, the population is initialized, and an adaptive mutation strategy is used to adjust the differential mutation factor and crossover probability. The cost function and service level index of multiple solutions are evaluated in each generation. The process is iterated weekly, and the inventory and production of each week are carried over to the next cycle. The optimization objective is to minimize the total cost while ensuring that the service level is greater than the threshold of 0.75.

[0081] Step six: Simulation and performance evaluation. Perform rolling simulations for the next four weeks, outputting weekly changes in service levels and cost structure, and compare them with the fixed strategy to verify the effectiveness of the dynamic strategy.

[0082] Specifically, the optimization process is based on Monte Carlo simulation. Each strategy undergoes 200 Monte Carlo simulations based on the inventory replenishment logic, evaluating inventory dynamics, cost distribution, and service level performance under each set of parameters s, S, and q. The optimal strategy combination is then searched through improved differential evolution. Optimization results are output weekly, including the weekly reorder point s, order cap S, planned production volume q for the cycle, and corresponding cost and service level indicators. The formulas for total cost and service level are as follows:

[0083] Cost=StockoutCost+HoldingCost+ProductCost+ShippingCost

[0084]

[0085] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for global production logistics modeling and joint optimization based on demand forecasting, characterized in that, Includes the following steps: Step 1: Data integration and preprocessing. This involves linking and integrating multi-source structured data such as sales logs, product information, manufacturing costs, and transportation costs, and performing feature normalization, data type conversion, and missing value handling on the data. Step 2: Multi-level feature engineering and time series structure construction. Through methods such as sliding window, lagged variables, and time period feature construction, statistical features related to future demand changes are extracted. Step 3: Based on the LightGBM demand forecasting model, the LightGBM regression model is used to model daily sales and to generate rolling forecasts for each product for the next four weeks. Step 4: Construct a joint optimization model for production-inventory strategies. This model is an extension of the fixed-cycle ordering strategy (T, s, S) by introducing a controllable production quantity q for joint optimization. Step 5: Optimize parameters based on the improved differential evolution algorithm SaDE. Use SaDE to jointly optimize the weekly inventory strategy parameters (s,S) and the cycle plan production quantity q. Step six: Simulation and performance evaluation. Perform rolling simulations for the next four weeks, outputting weekly changes in service levels and cost structure, and compare them with the fixed strategy to verify the effectiveness of the dynamic strategy.

2. The method for comprehensive production logistics modeling and joint optimization based on demand forecasting according to claim 1, characterized in that, Step one includes: integrating multi-source data based on the unique identifier Product_ID, constructing daily sales records for each product, converting the format of the time field in the dataset, labeling the categorical variables, and removing missing and outlier records.

3. The method for comprehensive production logistics modeling and joint optimization based on demand forecasting according to claim 1, characterized in that, Step two includes the following specific steps: For each product, the lagged characteristics of sales data are extracted to represent the difference between two periods of lagged sales, reflecting trend changes. For each date t: sold_lag k (t)=Sold_Num(t-k) Where k is the lag period (3, 7, 21), and Sold_Num is the sales volume for that day. Constructing a moving average time series to smooth the sales series and reduce short-term fluctuations, at the current time t, the past w Average daily sales: Construct a moving standard deviation time series to represent the fluctuation range of sales over the past w days:

4. The method for comprehensive production logistics modeling and joint optimization based on demand forecasting according to claim 1, characterized in that, Step three includes the following specific steps: We constructed training and validation sets, and used the logarithmically transformed sales volume as the target variable to train the LightGBM regression model. During the model training process, feature selection and importance evaluation were performed automatically, and the top 15 key features were retained for prediction. Use mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 The performance of the prediction model is evaluated using the following metrics, and the model parameters are adjusted for optimization. The formula for the evaluation metrics is as follows: Where n represents the number of samples, y i This represents the i-th actual sales volume. This represents the i-th predicted sales volume. This represents the true mean value; ∈ prevents division by zero. Based on the trained model, daily rolling forecasts are made for the next four weeks to generate a daily sales forecast sequence. The forecast results are then aggregated by natural week to generate a demand forecast matrix for four consecutive natural weeks.

5. The method for comprehensive production logistics modeling and joint optimization based on demand forecasting according to claim 1, characterized in that, Step four includes the following specific steps: Using a fixed ordering cycle T = 7 days and a lead time L = 3 days, within each cycle T, define the reorder point s, the target inventory level S, and the cycle production quantity q: s=D L +SS S = s + Q q=min(Q,Max_Capacity-current_inventory) Where SS represents safety stock, Q represents economic order quantity, Max_Capacity represents the maximum inventory capacity of the product, and current_inventory represents the current inventory level; The model's objective function design comprehensively considers multiple dimensions such as holding costs, stockout costs, production costs, and transportation costs. It also sets constraints on maximum production capacity, minimum order quantity, non-negative inventory levels, and safety stock fluctuation range. The optimization objective is to minimize total cost while maintaining service levels within these constraints. Objective = Cost + Penalty Where Cost represents the total cost and Penalty represents the service rate penalty.

6. The method for comprehensive production logistics modeling and joint optimization based on demand forecasting according to claim 1, characterized in that, Step five includes the following specific steps: The population is initialized, and an adaptive mutation strategy is used to adjust the differential mutation factor and crossover probability. The cost function and service level index of multiple solutions are evaluated in each generation. The process is iterated weekly, and the inventory and production of each week are carried over to the next cycle. The optimization objective is to minimize the total cost while ensuring that the service level is greater than the threshold of 0.

75.

7. The method for comprehensive production logistics modeling and joint optimization based on demand forecasting according to claim 1, characterized in that, Step six includes the following specific steps: The optimization process is based on Monte Carlo simulation. Each strategy undergoes 200 Monte Carlo simulations based on the inventory replenishment logic to evaluate inventory dynamics, cost distribution, and service level performance under each set of s, S, and q parameters. The optimal strategy combination is searched through improved differential evolution. The optimization results are output weekly, including the weekly reorder point s, order cap S, planned production volume q for the cycle, and the corresponding cost and service level indicators. The formulas for total cost and service level are as follows: Cost=StockoutCost+HoldingCost+ProductCost+ShippingCost

Images