An inventory optimization model construction method supporting cross-border e-commerce supply chain
By constructing a multi-source data acquisition module for the cross-border e-commerce supply chain and a core framework for an inventory optimization model, the problems of single data and poor model synergy in existing technologies have been solved. This has enabled the dynamic adaptability and accuracy of the inventory optimization model, thereby improving the efficiency and accuracy of inventory management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI SAIPAI INFORMATION CONSULTING CO LTD
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-14
AI Technical Summary
Existing cross-border e-commerce supply chain inventory management suffers from problems such as limited data, poor model coordination, and inability to adapt to dynamic changes, resulting in poor inventory optimization, low sales forecast accuracy, and an inability to meet the needs of refined and intelligent management.
A multi-source data acquisition module for the cross-border e-commerce supply chain was constructed, and a core framework for an inventory optimization model was established, including sub-models for demand forecasting, inventory allocation, replenishment scheduling, and constraint configuration. A hybrid algorithm combining deep learning and statistical analysis was adopted to achieve collaborative training and dynamic calibration of the sub-models, and real-time adjustments were made in conjunction with multi-dimensional data.
It improves the comprehensiveness and accuracy of data in the inventory optimization model, adapts to dynamic changes in the supply chain, enhances the accuracy of sales forecasting and the reliability of inventory management, and reduces the risks of stockouts and inventory backlog.
Smart Images
Figure CN122390630A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cross-border e-commerce supply chain management technology, and more specifically to a method for constructing an inventory optimization model to support cross-border e-commerce supply chains. Background Technology
[0002] With the rapid development of the cross-border e-commerce industry, the collaborative efficiency of each link in the supply chain has become a key factor affecting corporate competitiveness. Among them, inventory management, as a core link in the supply chain, is directly related to a company's cash flow, customer experience, and market competitiveness. Currently, the cross-border e-commerce supply chain faces complex problems such as multi-warehouse layout (domestic and overseas warehouses), fluctuating logistics timeliness, changing customs supervision policies, uncertain market demand, and competition from other products. Traditional inventory management methods and existing intelligent replenishment tools are no longer sufficient to meet the needs of refined and intelligent inventory optimization.
[0003] In existing technologies, most cross-border e-commerce ERP systems have significant shortcomings in their intelligent replenishment functions: First, the replenishment calculation model is singular and cannot adapt to the needs of multiple business scenarios, nor can it be flexibly adjusted according to actual operational scenarios; second, in procurement scenarios, the deduplication calculation of shared inventory is not implemented, resulting in a large deviation between recommended procurement quantities and actual demand, which can easily lead to stockouts or inventory backlogs; third, existing inventory models mostly use a single algorithm for demand forecasting, without fully integrating multi-source data (such as customs supervision data and competitor data), resulting in insufficient prediction accuracy, and there is a lack of effective collaboration between sub-models, making it impossible to achieve end-to-end inventory optimization; fourth, after the model is built, there is a lack of a dynamic calibration mechanism, making it difficult to adapt to dynamic changes in the cross-border e-commerce supply chain, such as changes in supplier capacity, fluctuations in logistics timeliness, and adjustments in customs supervision policies, leading to a decline in model accuracy after long-term use and an inability to continuously provide reliable support for inventory management.
[0004] Crucially, in existing inventory optimization technologies, sales forecasting and demand forecasting are often singularly related or independent, lacking a coordinated mechanism. Some technologies do not have a separate sales forecasting module, relying solely on simple historical sales statistics to deduce demand, failing to accurately capture sales fluctuation patterns across different cross-border platforms and timeframes. While some technologies do have sales forecasting functions, they employ a single time-series algorithm without integrating multi-dimensional influencing factors such as competitor data and market demand changes, resulting in low sales forecast accuracy. This fails to provide reliable basic data support for demand forecasting, inventory allocation, and replenishment scheduling, thus affecting the overall performance of the inventory optimization model and failing to address the core issue of mismatch between inventory, sales, and demand at its root.
[0005] Furthermore, existing sales forecasting functions are mostly static forecasts that do not take into account the dynamic changes in the cross-border e-commerce supply chain and are not adjusted in real time. When there are situations such as competitors adjusting prices, platform policy changes, or fluctuations in logistics timeliness, the sales forecast data cannot be updated in a timely manner, which further exacerbates the deviation in inventory optimization and leads to the dual risks of stockouts and loss of customers, as well as inventory backlog tying up capital.
[0006] Therefore, proposing a method for constructing an inventory optimization model to support the cross-border e-commerce supply chain, in order to overcome the difficulties of existing technologies, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0007] In view of this, the present invention provides a method for constructing an inventory optimization model to support cross-border e-commerce supply chain, which is used to solve the technical problems existing in the prior art.
[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for constructing an inventory optimization model to support cross-border e-commerce supply chain includes the following steps: S1. Build a multi-source data acquisition module for the cross-border e-commerce supply chain. The multi-source data acquisition module is used to collect raw data from each link of the cross-border e-commerce supply chain. The raw data includes supplier data, warehousing data, logistics data, sales data, customs supervision data, and competitor data. S2. Preprocess the collected raw data to obtain standardized data; S3. Construct the core framework of the inventory optimization model. The core framework includes a demand forecasting sub-model, an inventory allocation sub-model, a replenishment scheduling sub-model, a constraint configuration sub-model, and a sales forecasting sub-model. The sub-models interact bidirectionally through a data interface. S4. Input the obtained standardized data into the core framework and configure the initial parameters for each sub-model; S5. Perform collaborative training on each sub-model. Iteratively adjust the parameters of each sub-model to ensure that the output data of each sub-model meets the preset collaborative conditions, and obtain the trained inventory optimization model. S6. Connect the trained inventory optimization model with the cross-border e-commerce supply chain management system, and dynamically calibrate the inventory optimization model using real-time collected supply chain dynamic data to complete the construction of the inventory optimization model.
[0009] Optionally, the supplier data in S1 includes: supplier capacity data, delivery cycle data, delivery qualification rate data, and supply cost data; Warehouse data includes: inventory levels, inventory turnover rates, warehouse space occupancy, and warehouse cost data for all overseas and domestic warehouses. Logistics data includes: international transportation timeliness data, logistics cost data, logistics loss rate data, and logistics node status data; Sales data includes: order data, return data, average order value data, and sales fluctuation data from various cross-border platforms; Customs supervision data includes: customs clearance timeliness data, tariff data, and regulatory restriction data; Competitive data includes: competitor inventory data, competitor pricing data, and competitor sales data.
[0010] Optionally, preprocessing in S2 includes data cleaning, data deduplication, data completion, and data normalization, and the format of standardized data is unified to the preset data format; Data cleaning is used to remove outliers and invalid data from the original data; Outliers are identified using the 3σ criterion; invalid data includes data with missing key fields. Data completion uses interpolation to fill in missing non-critical field data; Data normalization uses the min-max normalization method to map the original data to the [0,1] interval.
[0011] Optionally, the demand forecasting sub-model in S3 adopts a hybrid algorithm that integrates deep learning and statistical analysis. The hybrid algorithm includes LSTM network and ARIMA algorithm. The demand forecasting sub-model is used to output demand forecast data for each SKU within a preset period based on the output data of the sales forecasting sub-model. The sales forecast sub-model adopts a fusion of time series analysis algorithm and machine learning algorithm. The time series analysis algorithm includes the Holt-Winters algorithm, and the machine learning algorithm includes the random forest algorithm. It is used to collect sales data, competitor data and market demand-related data from standardized data, and output sales forecast data for each SKU in a preset period and on different cross-border platforms. The inventory allocation sub-model is used to output inventory allocation data for each SKU at different warehouse nodes based on demand forecast data and inventory status at each warehouse node. The inventory allocation data includes the inventory transfer amount and inventory holding amount at each warehouse node. The replenishment scheduling sub-model is used to output replenishment timing and replenishment quantity data based on inventory allocation data, replenishment lead time and sales forecast data. The replenishment scheduling sub-model also supports the calculation of mixed flow replenishment schemes that combine sea and air freight. The constraint configuration sub-model is used to set the constraint parameters for inventory optimization. These constraints include inventory holding cost constraints, stockout constraints, storage capacity constraints, and customs supervision constraints.
[0012] Optionally, the constraint configuration sub-model is also used to adjust the range of constraint parameters in real time according to the dynamic changes in the cross-border e-commerce supply chain. These dynamic changes include changes in supplier capacity, logistics timeliness, customs supervision policies, and market demand.
[0013] Optionally, the initial parameters in S4 include: demand forecasting error threshold, safety stock threshold, replenishment batch threshold, and dynamic calibration cycle; The safety stock threshold is determined based on the demand fluctuations of each SKU, logistics lead time, and sales forecast data. The dynamic calibration cycle can be flexibly set according to the frequency of dynamic changes in the supply chain, with a range of 1-7 days. In the sales forecast parameters, the sales forecast period is consistent with the demand forecast period, the sales fluctuation coefficient ranges from 0.8 to 1.2, and the sales forecast error threshold ranges from 4% to 9%.
[0014] Optionally, the specific process of collaborative training in S5 is as follows: First, each sub-model is trained individually to ensure that the output error of each sub-model is lower than the corresponding preset error threshold. Specifically, the output error of the sales forecast sub-model is lower than the sales forecast error threshold, and the output errors of the other sub-models are lower than the preset error threshold. Then, the output data of the sales forecast sub-model is used as the input data of the demand forecast sub-model, the output data of the demand forecast sub-model is used as the input data of the inventory allocation sub-model, the output data of the inventory allocation sub-model is used as the input data of the replenishment scheduling sub-model, and the output data of the replenishment scheduling sub-model is fed back to the constraint configuration sub-model. The constraint configuration sub-model adjusts the constraint parameters according to the feedback data, and then inputs the adjusted constraint parameters into the other sub-models for multiple rounds of iterative training. Co-training continues until the deviation between the output data of each sub-model is less than the preset co-training deviation threshold. The preset error threshold ranges from 5% to 10%, and the preset coordination deviation threshold ranges from 3% to 5%.
[0015] Optionally, the specific process of dynamic calibration in S6 is as follows: every preset dynamic calibration cycle, real-time dynamic data of the cross-border e-commerce supply chain is collected, the real-time dynamic data is input into the inventory optimization model, the deviation between the current output data of the model and the actual supply chain data is calculated, and the parameters of each sub-model are adjusted according to the deviation value to achieve dynamic calibration of the model; among them, the parameters of the sales forecast sub-model are adjusted synchronously to ensure that the deviation between the sales forecast data and the real-time sales data is maintained within the sales forecast error threshold range; the real-time dynamic data includes real-time inventory data, real-time logistics timeliness data, real-time customs clearance data, real-time market demand data, and real-time sales data.
[0016] Optionally, the inventory optimization model also includes a data encryption submodule, which is used to encrypt the collected raw data, standardized data and model output data. The encryption process uses an asymmetric encryption algorithm to ensure the security of data transmission and storage.
[0017] Optionally, the core framework of S3 also includes a model validation sub-model. The model validation sub-model is used to validate the trained inventory optimization model using historical standardized data. After successful validation, it is then connected to the cross-border e-commerce supply chain management system. The validation standard is that the deviation between the model output data and the historical actual data is less than a preset validation deviation threshold. The preset validation deviation threshold ranges from 3% to 8%.
[0018] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for constructing an inventory optimization model to support the cross-border e-commerce supply chain, the beneficial effects of which are: 1) By building a multi-source data acquisition module, comprehensive data from suppliers, warehousing, logistics, sales, customs supervision, and competitors is collected. This solves the problem of single data dimensions and inability to fully reflect the supply chain status in existing technologies, providing comprehensive data support for inventory optimization. At the same time, through data preprocessing, the accuracy and standardization of the data are ensured, laying the foundation for model training. 2) The core framework of the constructed inventory optimization model includes four collaborative sub-models. The sub-models achieve bidirectional data interaction through data interfaces, which solves the defects of independent operation and poor collaboration of sub-models in the existing technology. Among them, the demand forecasting sub-model adopts a hybrid algorithm of LSTM and ARIMA to improve the accuracy of demand forecasting. The replenishment scheduling sub-model supports mixed flow replenishment schemes to adapt to business needs in multiple scenarios. The constraint configuration sub-model can dynamically adjust constraint parameters to adapt to dynamic changes in the supply chain. 3) Through sub-model collaborative training and dynamic model calibration mechanisms, the accuracy and stability of model output are ensured, solving the problems of insufficient model training and inability to adapt to dynamic changes in the supply chain in existing technologies. Among them, the sales forecast sub-model and the demand forecast sub-model are linked to improve the accuracy of demand forecasting and provide more realistic basic data for inventory allocation and replenishment scheduling. At the same time, the addition of a data encryption sub-module and a model verification sub-model improves the security and reliability of the model. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0020] Figure 1 The flowchart illustrates a method for constructing an inventory optimization model to support cross-border e-commerce supply chains, as provided by this invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] See Figure 1 As shown, this invention discloses a method for constructing an inventory optimization model to support cross-border e-commerce supply chains, comprising the following steps: S1. Build a multi-source data acquisition module for the cross-border e-commerce supply chain. The multi-source data acquisition module is used to collect raw data from each link of the cross-border e-commerce supply chain. The raw data includes supplier data, warehousing data, logistics data, sales data, customs supervision data, and competitor data. S2. Preprocess the collected raw data to obtain standardized data; S3. Construct the core framework of the inventory optimization model. The core framework includes a demand forecasting sub-model, an inventory allocation sub-model, a replenishment scheduling sub-model, a constraint configuration sub-model, and a sales forecasting sub-model. The sub-models interact bidirectionally through a data interface. S4. Input the obtained standardized data into the core framework and configure the initial parameters for each sub-model; S5. Perform collaborative training on each sub-model. Iteratively adjust the parameters of each sub-model to ensure that the output data of each sub-model meets the preset collaborative conditions, and obtain the trained inventory optimization model. S6. Connect the trained inventory optimization model with the cross-border e-commerce supply chain management system, and dynamically calibrate the inventory optimization model using real-time collected supply chain dynamic data to complete the construction of the inventory optimization model.
[0023] Furthermore, the supplier data in S1 includes: supplier capacity data, delivery cycle data, delivery qualification rate data, and supply cost data; Warehouse data includes: inventory levels, inventory turnover rates, warehouse space occupancy, and warehouse cost data for all overseas and domestic warehouses. Logistics data includes: international transportation timeliness data, logistics cost data, logistics loss rate data, and logistics node status data; Sales data includes: order data, return data, average order value data, and sales fluctuation data from various cross-border platforms; Customs supervision data includes: customs clearance timeliness data, tariff data, and regulatory restriction data; Competitive data includes: competitor inventory data, competitor pricing data, and competitor sales data.
[0024] Furthermore, preprocessing in S2 includes data cleaning, data deduplication, data completion, and data normalization, and the format of standardized data is unified to the preset data format; Data cleaning is used to remove outliers and invalid data from the original data; Outliers are identified using the 3σ criterion; invalid data includes data with missing key fields. Data completion uses interpolation to fill in missing non-critical field data; Data normalization uses the min-max normalization method to map the original data to the [0,1] interval.
[0025] Furthermore, the demand forecasting sub-model in S3 adopts a hybrid algorithm that integrates deep learning and statistical analysis. The hybrid algorithm includes LSTM network and ARIMA algorithm. The demand forecasting sub-model is used to output demand forecast data for each SKU within a preset period based on the output data of the sales forecasting sub-model. The sales forecast sub-model adopts a fusion of time series analysis algorithm and machine learning algorithm. The time series analysis algorithm includes the Holt-Winters algorithm, and the machine learning algorithm includes the random forest algorithm. It is used to collect sales data, competitor data and market demand-related data from standardized data, and output sales forecast data for each SKU in a preset period and on different cross-border platforms. The inventory allocation sub-model is used to output inventory allocation data for each SKU at different warehouse nodes based on demand forecast data and inventory status at each warehouse node. The inventory allocation data includes the inventory transfer amount and inventory holding amount at each warehouse node. The replenishment scheduling sub-model is used to output replenishment timing and replenishment quantity data based on inventory allocation data, replenishment lead time and sales forecast data. The replenishment scheduling sub-model also supports the calculation of mixed flow replenishment schemes that combine sea and air freight. The constraint configuration sub-model is used to set the constraint parameters for inventory optimization. These constraints include inventory holding cost constraints, stockout constraints, storage capacity constraints, and customs supervision constraints.
[0026] Furthermore, the constraint configuration sub-model is also used to adjust the range of constraint parameters in real time according to the dynamic changes in the cross-border e-commerce supply chain. These dynamic changes include changes in supplier capacity, logistics timeliness, customs supervision policies, and market demand.
[0027] Furthermore, the initial parameters in S4 include: demand forecasting error threshold, safety stock threshold, replenishment batch threshold, and dynamic calibration cycle; The safety stock threshold is determined based on the demand fluctuations of each SKU, logistics lead time, and sales forecast data. The dynamic calibration cycle can be flexibly set according to the frequency of dynamic changes in the supply chain, with a range of 1-7 days. In the sales forecast parameters, the sales forecast period is consistent with the demand forecast period, the sales fluctuation coefficient ranges from 0.8 to 1.2, and the sales forecast error threshold ranges from 4% to 9%.
[0028] Furthermore, the specific process of collaborative training in S5 is as follows: First, each sub-model is trained individually to ensure that the output error of each sub-model is lower than the corresponding preset error threshold. Specifically, the output error of the sales forecast sub-model is lower than the sales forecast error threshold, and the output errors of the other sub-models are lower than the preset error threshold. Then, the output data of the sales forecast sub-model is used as the input data of the demand forecast sub-model, the output data of the demand forecast sub-model is used as the input data of the inventory allocation sub-model, the output data of the inventory allocation sub-model is used as the input data of the replenishment scheduling sub-model, and the output data of the replenishment scheduling sub-model is fed back to the constraint configuration sub-model. The constraint configuration sub-model adjusts the constraint parameters according to the feedback data, and then inputs the adjusted constraint parameters into the other sub-models for multiple rounds of iterative training. Co-training continues until the deviation between the output data of each sub-model is less than the preset co-training deviation threshold. The preset error threshold ranges from 5% to 10%, and the preset coordination deviation threshold ranges from 3% to 5%.
[0029] Furthermore, the specific process of dynamic calibration in S6 is as follows: every preset dynamic calibration cycle, real-time dynamic data of the cross-border e-commerce supply chain is collected, and the real-time dynamic data is input into the inventory optimization model. The deviation between the current output data of the model and the actual supply chain data is calculated, and the parameters of each sub-model are adjusted according to the deviation value to achieve dynamic calibration of the model. Among them, the parameters of the sales forecast sub-model are adjusted synchronously to ensure that the deviation between the sales forecast data and the real-time sales data is maintained within the sales forecast error threshold range. The real-time dynamic data includes real-time inventory data, real-time logistics timeliness data, real-time customs clearance data, real-time market demand data, and real-time sales data.
[0030] Furthermore, the inventory optimization model also includes a data encryption submodule, which is used to encrypt the collected raw data, standardized data, and model output data. The encryption process uses an asymmetric encryption algorithm to ensure the security of data transmission and storage.
[0031] Furthermore, the core framework of S3 also includes a model validation sub-model. The model validation sub-model is used to validate the trained inventory optimization model using historical standardized data. After successful validation, it is then connected to the cross-border e-commerce supply chain management system. The validation standard is that the deviation between the model output data and the historical actual data is less than a preset validation deviation threshold. The preset validation deviation threshold ranges from 3% to 8%.
[0032] In a specific embodiment: S1: Build a multi-source data acquisition module for the cross-border e-commerce supply chain. The multi-source data acquisition module connects with cross-border e-commerce platforms, supplier systems, warehouse management systems, logistics service provider systems, customs supervision systems, and competitor data acquisition platforms through API interfaces to collect raw data from each link of the cross-border e-commerce supply chain in real time. Among them, supplier data includes supplier capacity data, supply cycle data, supply qualification rate data, and supply cost data. Warehouse data includes inventory levels (1200 units, 800 units, and 600 units respectively) for domestic warehouse A, overseas warehouse B, and overseas warehouse C; inventory turnover rates (1.2 times / month, 0.9 times / month, and 1.0 times / month respectively); and warehouse space occupancy data (300m²). 2 220m 2 180m 2 ) and warehousing cost data (2 yuan / piece / month, 5 yuan / piece / month, and 4.5 yuan / piece / month, respectively); Logistics data includes international sea freight transit time data (25-30 days), international air freight transit time data (3-5 days), logistics cost data (8 RMB / piece for sea freight, 35 RMB / piece for air freight), logistics loss rate data (0.3% for sea freight, 0.1% for air freight), and logistics node status data (such as the location of sea freight vessels and the status of air freight flights). Sales data includes order data from Amazon and AliExpress platforms (320 orders per day), return data (12 returns per day), average order value (150 yuan / order), and sales fluctuation data (weekly fluctuation range ±8%). Customs supervision data includes customs clearance time data (5-7 days), tariff data (10%), and regulatory restriction data (such as categories of goods prohibited from cross-border transportation). The competitor data includes inventory data for three core competitors (950, 780, and 650 units respectively), pricing data for competitors (145 yuan / unit, 155 yuan / unit, and 148 yuan / unit respectively), and sales data for competitors (average daily sales of 280, 250, and 220 units respectively).
[0033] S2: The raw data collected in S1 is preprocessed to obtain standardized data. Among them, data cleaning uses the 3σ criterion to identify and remove outliers, and remove invalid data such as missing supplier capacity and inventory. Data completion uses linear interpolation to supplement missing non-critical field data. Data normalization uses the min-max normalization method to map all raw data to the [0,1] interval. The standardized data is uniformly in JSON format to facilitate data interaction between sub-models.
[0034] S3: The core framework for building the inventory optimization model includes a demand forecasting sub-model, an inventory allocation sub-model, a replenishment scheduling sub-model, a constraint configuration sub-model, a data encryption sub-module, a model validation sub-model, and a sales forecasting sub-model. These sub-models interact bidirectionally via a RESTful data interface. The sales forecasting sub-model uses a fusion of Holt-Winters time-series analysis and random forest machine learning algorithms to collect standardized sales data, competitor data, and market demand-related data. It outputs daily and weekly sales forecasts for each SKU on Amazon and AliExpress platforms for the next 30 days, with a sales forecast error controlled within 8%. The demand forecasting sub-model uses a hybrid algorithm combining LSTM networks and ARIMA algorithms. Based on the output data from the sales forecasting sub-model, it outputs demand forecast data for each SKU for the next 30 days. The inventory allocation sub-model outputs inventory allocation data for each SKU in domestic warehouse A, overseas warehouse B, and overseas warehouse C, based on the demand forecast data and the inventory status of each warehouse node. The replenishment scheduling sub-model outputs replenishment timing and quantity data based on inventory allocation data, replenishment lead time, and sales forecast data. It also supports mixed logistics replenishment schemes combining sea and air freight. The constraint configuration sub-model sets constraint parameters for inventory optimization, including inventory holding cost constraints (monthly holding cost per SKU not exceeding 10 yuan), stockout constraints (stockout rate not exceeding 2%), warehouse capacity constraints (maximum capacity of 1500 units for domestic warehouse A, 1000 units for overseas warehouse B, and 800 units for overseas warehouse C), and customs supervision constraints (compliance with tariff and category supervision requirements). The data encryption sub-module uses the RSA asymmetric encryption algorithm to encrypt the collected raw data, standardized data, and model output data. The model validation sub-model validates the trained inventory optimization model and verifies the deviation between sales forecast data and historical actual sales data.
[0035] S4: Input the standardized data obtained in S2 into the core framework built in S3 to configure initial parameters for each sub-model; among them, the demand forecasting cycle is 30 days, the inventory allocation threshold is 80% of the maximum capacity of each warehouse node (domestic warehouse A 1200 units, overseas warehouse B 800 units, overseas warehouse C 640 units), the replenishment lead time is 30 days for sea freight and 5 days for air freight, the customs clearance cycle is 7 days, the demand forecasting error threshold is 8%, the inventory safety stock threshold is determined based on the demand fluctuation range of each SKU, the logistics lead time and sales forecast data, the replenishment batch threshold is 500 units / batch, and the dynamic calibration cycle is 3 days; the sales forecasting parameters include a sales forecasting cycle of 30 days, a sales fluctuation coefficient of 1.0, and a sales forecasting error threshold of 8%.
[0036] S5: Perform collaborative training on each sub-model. First, train each sub-model individually to ensure that the output error of the sales forecast sub-model is below 8%, and the output errors of other sub-models are below a preset error threshold of 8%. Then, use the sales forecast data output by the sales forecast sub-model as input data for the demand forecast sub-model, the output data of the demand forecast sub-model as input data for the inventory allocation sub-model, and the output data of the inventory allocation sub-model as input data for the replenishment scheduling sub-model. Feedback the output data of the replenishment scheduling sub-model to the constraint configuration sub-model. The constraint configuration sub-model adjusts the constraint parameters based on the feedback data, and then inputs the adjusted constraint parameters to other sub-models for multiple rounds of iterative training. This continues until the deviation between the output data of each sub-model is less than a preset collaborative deviation threshold of 4%, resulting in a trained inventory optimization model.
[0037] S6: Validate the trained inventory optimization model using historical standardized data (supply chain data from the past 6 months). Validation is considered successful when the deviation between the model output data and historical actual data is 6%, and the deviation between the sales forecast data and historical actual sales data is 7%, both less than the corresponding preset error thresholds. Connect the validated inventory optimization model to the cross-border e-commerce supply chain management system. Every 3 days (dynamic calibration cycle), collect real-time dynamic data from the cross-border e-commerce supply chain and input this data into the inventory optimization model. Calculate the deviation between the model's current output data and actual supply chain data, and the deviation between the sales forecast data and real-time sales data. Adjust the parameters of each sub-model and the sales forecast sub-model synchronously based on the deviation values to achieve dynamic calibration of the model and complete the construction of the inventory optimization model.
[0038] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0039] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for constructing an inventory optimization model supporting cross-border e-commerce supply chain, characterized in that, Includes the following steps: S1. Build a multi-source data acquisition module for the cross-border e-commerce supply chain. The multi-source data acquisition module is used to collect raw data from each link of the cross-border e-commerce supply chain. The raw data includes supplier data, warehousing data, logistics data, sales data, customs supervision data, and competitor data. S2. Preprocess the collected raw data to obtain standardized data; S3. Construct the core framework of the inventory optimization model. The core framework includes a demand forecasting sub-model, an inventory allocation sub-model, a replenishment scheduling sub-model, a constraint configuration sub-model, and a sales forecasting sub-model. The sub-models interact bidirectionally through a data interface. S4. Input the obtained standardized data into the core framework and configure the initial parameters for each sub-model; S5. Perform collaborative training on each sub-model. Iteratively adjust the parameters of each sub-model to ensure that the output data of each sub-model meets the preset collaborative conditions, and obtain the trained inventory optimization model. S6. Connect the trained inventory optimization model with the cross-border e-commerce supply chain management system, and dynamically calibrate the inventory optimization model using real-time collected supply chain dynamic data to complete the construction of the inventory optimization model.
2. The method for constructing an inventory optimization model supporting cross-border e-commerce supply chain according to claim 1, characterized in that, The supplier data in S1 includes: supplier capacity data, delivery cycle data, delivery qualification rate data, and supply cost data; Warehouse data includes: inventory levels, inventory turnover rates, warehouse space occupancy, and warehouse cost data for all overseas and domestic warehouses. Logistics data includes: international transportation timeliness data, logistics cost data, logistics loss rate data, and logistics node status data; Sales data includes: order data, return data, average order value data, and sales fluctuation data from various cross-border platforms; Customs supervision data includes: customs clearance timeliness data, tariff data, and regulatory restriction data; Competitive data includes: competitor inventory data, competitor pricing data, and competitor sales data.
3. The method for constructing an inventory optimization model supporting cross-border e-commerce supply chain according to claim 1, characterized in that, Preprocessing in S2 includes data cleaning, data deduplication, data completion, and data normalization. The format of standardized data is unified to the preset data format. Data cleaning is used to remove outliers and invalid data from the original data; Outliers are identified using the 3σ criterion; invalid data includes data with missing key fields. Data completion uses interpolation to fill in missing non-critical field data; Data normalization uses the min-max normalization method to map the original data to the [0,1] interval.
4. The method for constructing an inventory optimization model supporting cross-border e-commerce supply chain according to claim 1, characterized in that, The demand forecasting sub-model in S3 adopts a hybrid algorithm that integrates deep learning and statistical analysis. The hybrid algorithm includes LSTM network and ARIMA algorithm. The demand forecasting sub-model is used to output demand forecast data for each SKU within a preset period based on the output data of the sales forecasting sub-model. The sales forecast sub-model adopts a fusion of time series analysis algorithm and machine learning algorithm. The time series analysis algorithm includes the Holt-Winters algorithm, and the machine learning algorithm includes the random forest algorithm. It is used to collect sales data, competitor data and market demand-related data from standardized data, and output sales forecast data for each SKU in a preset period and on different cross-border platforms. The inventory allocation sub-model is used to output inventory allocation data for each SKU at different warehouse nodes based on demand forecast data and inventory status at each warehouse node. The inventory allocation data includes the inventory transfer amount and inventory holding amount at each warehouse node. The replenishment scheduling sub-model is used to output replenishment timing and replenishment quantity data based on inventory allocation data, replenishment lead time and sales forecast data. The replenishment scheduling sub-model also supports the calculation of mixed flow replenishment schemes that combine sea and air freight. The constraint configuration sub-model is used to set the constraint parameters for inventory optimization. These constraints include inventory holding cost constraints, stockout constraints, storage capacity constraints, and customs supervision constraints.
5. The method for constructing an inventory optimization model supporting cross-border e-commerce supply chain according to claim 4, characterized in that, The constraint configuration sub-model is also used to adjust the range of constraint parameters in real time according to the dynamic changes in the cross-border e-commerce supply chain. These dynamic changes include changes in supplier capacity, logistics timeliness, customs supervision policies, and market demand.
6. The method for constructing an inventory optimization model supporting cross-border e-commerce supply chain according to claim 1, characterized in that, The initial parameters in S4 include: demand forecasting error threshold, safety stock threshold, replenishment batch threshold, and dynamic calibration cycle. The safety stock threshold is determined based on the demand fluctuations of each SKU, logistics lead time, and sales forecast data. The dynamic calibration cycle can be flexibly set according to the frequency of dynamic changes in the supply chain, with a range of 1-7 days. In the sales forecast parameters, the sales forecast period is consistent with the demand forecast period, the sales fluctuation coefficient ranges from 0.8 to 1.2, and the sales forecast error threshold ranges from 4% to 9%.
7. The method for constructing an inventory optimization model supporting cross-border e-commerce supply chain according to claim 1, characterized in that, The specific process of collaborative training in S5 is as follows: First, train each sub-model separately so that the output error of each sub-model is lower than the corresponding preset error threshold. Specifically, the output error of the sales prediction sub-model is lower than the sales prediction error threshold, and the output error of the other sub-models is lower than the preset error threshold. Then, the output data of the sales forecasting sub-model is used as the input data of the demand forecasting sub-model, the output data of the demand forecasting sub-model is used as the input data of the inventory allocation sub-model, the output data of the inventory allocation sub-model is used as the input data of the replenishment scheduling sub-model, and the output data of the replenishment scheduling sub-model is fed back to the constraint configuration sub-model. The constraint configuration sub-model adjusts the constraint parameters according to the feedback data, and then inputs the adjusted constraint parameters into other sub-models for multiple rounds of iterative training. Co-training continues until the deviation between the output data of each sub-model is less than the preset co-training deviation threshold. The preset error threshold ranges from 5% to 10%, and the preset coordination deviation threshold ranges from 3% to 5%.
8. The method for constructing an inventory optimization model supporting cross-border e-commerce supply chain according to claim 1, characterized in that, The specific process of dynamic calibration in S6 is as follows: Every preset dynamic calibration cycle, real-time dynamic data of the cross-border e-commerce supply chain is collected, and the real-time dynamic data is input into the inventory optimization model. The deviation between the current output data of the model and the actual supply chain data is calculated, and the parameters of each sub-model are adjusted according to the deviation value to achieve dynamic calibration of the model. Among them, the parameters of the sales forecast sub-model are adjusted synchronously to ensure that the deviation between the sales forecast data and the real-time sales data is maintained within the sales forecast error threshold. The real-time dynamic data includes real-time inventory data, real-time logistics timeliness data, real-time customs clearance data, real-time market demand data, and real-time sales data.
9. The method for constructing an inventory optimization model supporting cross-border e-commerce supply chain according to claim 1, characterized in that, The inventory optimization model also includes a data encryption submodule, which is used to encrypt the collected raw data, standardized data and model output data. The encryption process uses an asymmetric encryption algorithm to ensure the security of data transmission and storage.
10. The method for constructing an inventory optimization model supporting cross-border e-commerce supply chain according to claim 1, characterized in that, The core framework of S3 also includes a model validation sub-model. The model validation sub-model is used to validate the trained inventory optimization model using historical standardized data. After successful validation, it is then connected to the cross-border e-commerce supply chain management system. The validation standard is that the deviation between the model output data and the historical actual data is less than the preset validation deviation threshold. The preset validation deviation threshold ranges from 3% to 8%.