Cross-border e-commerce supply chain multi-platform collaboration and inventory dynamic prediction management system

Abstract

This invention relates to the field of inventory management technology and discloses a cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system. The system includes: a data acquisition unit that processes field formats, time granularity, and unit systems based on the heterogeneous data protocols of various cross-border e-commerce platforms, and indexes the processed data using timestamp numbers and platform codes; a structure fusion unit that resamples the data from each cross-border e-commerce platform based on the timestamp numbers in the acquired dataset, and merges records of the same product across multiple platforms based on the product code; a sequence parsing unit that segments the cumulative sales field according to the window span based on the fused structure matrix, and constructs a sequence feature chain based on the feature vectors of each segment; and a trend inference unit that performs difference calculations on the current inventory quantity based on the predicted sales quantity, and determines the replenishment quantity based on the result of the difference calculation. This invention ensures multi-platform collaboration and inventory adaptability to the market.

Description

Technical Field

[0001] This invention relates to the field of inventory management technology, and more specifically, to a cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system. Background Technology

[0002] With the rapid development of globalization and e-commerce, the cross-border e-commerce industry has experienced large-scale growth, and more and more merchants are choosing to operate on multiple cross-border e-commerce platforms to expand their market coverage. While the multi-platform operation model brings incremental revenue to merchants, it also places higher demands on supply chain management. At the multi-platform data collaboration level, different cross-border e-commerce platforms, based on their own technical architecture and business rules, adopt heterogeneous data protocols and formats. This results in differences in field definitions, time granularity, and unit formats for data such as order records and outbound records. Existing data integration often relies on simple field matching, which is not only inefficient but also prone to data distortion due to inaccurate field mapping, unit conversion errors, and inconsistent time granularity. This, in turn, affects collaborative analysis and decision-making across multiple platforms. At the same time, the cross-border e-commerce supply chain faces long logistics cycles, large fluctuations in market demand, and dispersed inventory across multiple platforms. It has high requirements for the accuracy and real-time performance of inventory forecasting. Current inventory forecasting is mostly based on historical sales averages and trend line fitting on a single platform, focusing only on the overall trend of sales changes and ignoring the fluctuation range of sales in different time periods and time disturbances. As a result, the forecast results cannot reflect the dynamic fluctuation patterns of market demand. Because the impact of different sales changes within a period on the inventory consumption rhythm is ignored, the replenishment volume of cross-border e-commerce cannot match the inventory demand.

[0003] Therefore, it is necessary to design a cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system to solve the problems existing in the current technology. Summary of the Invention

[0004] In view of this, the present invention proposes a cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system, which aims to solve the above problems.

[0005] This invention proposes a cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system, comprising:

[0006] The data acquisition unit is configured to process the field format, time granularity, and unit format of order records, outbound records, inbound records, and product status change records based on the heterogeneous data protocols of various cross-border e-commerce platforms, and to construct the acquisition dataset by indexing the processed data with timestamp numbers and platform codes.

[0007] The structural fusion unit is configured to resample the data from each cross-border e-commerce platform based on the timestamp number in the collected dataset, and merge the records of the same product from multiple platforms based on the product code to determine the fusion structure matrix.

[0008] The sequence parsing unit is configured to segment the cumulative sales field according to the window span based on the fusion structure matrix, and determine the slope of cumulative sales change, the width of fluctuation envelope and the progressive time perturbation amount in each segment based on the time gradient, and construct segment feature vectors, and construct a sequence feature chain based on the segment order of each segment feature vector.

[0009] The trend extrapolation unit is configured to extrapolate the slope change direction, fluctuation envelope extension and time disturbance accumulation segment by segment based on the sequence feature chain to determine the sales forecast, perform difference calculation on the current inventory based on the sales forecast, and determine the replenishment quantity based on the result of the difference calculation.

[0010] Furthermore, when processing the field formats, time granularity, and unit systems of order records, outbound records, inbound records, and product status change records based on the heterogeneous data protocols of various cross-border e-commerce platforms, and constructing the collected dataset using timestamp numbers and platform codes as indexes, the following steps are included:

[0011] The data acquisition unit maps the field names, field lengths, and field types of each cross-border e-commerce platform. It compares the fields of different cross-border e-commerce platforms one by one, redefines the types of fields that cannot be mapped, and truncates or fills the fields that cannot be mapped based on the differences in field length to determine the mapping result.

[0012] Time step conversion is performed on the time fields of various cross-border e-commerce platforms. The time fields at the minute, hour, or day level are quantized by step size. Based on the quantization results, the initial time values ​​are mapped to a unified time grid point to determine the time step conversion result.

[0013] The quantity field, amount field, and weight field are converted to units. A conversion factor is matched based on the corresponding initial unit label, and the conversion factor is calculated with the initial field value to determine the unit conversion result.

[0014] The mapping result, time step conversion result, and unit conversion result are written into the same record, and a timestamp number and platform code are written into the record. A sequential linked list is constructed based on the timestamp number in an ascending manner, and the link relationship between the predecessor node and the successor node is recorded in the linked list node to construct the collected dataset.

[0015] Furthermore, when resampling the timestamps of each cross-border e-commerce platform based on the timestamp numbers in the collected dataset, and merging the records of the same product across multiple platforms based on the product code to determine the fusion structure matrix, the process includes:

[0016] The structural fusion unit selects the resampling time step based on the timestamp number of the collected dataset, and divides the data records of each cross-border e-commerce platform into several time slices according to the resampling time step.

[0017] Within each time slice, each product record is filtered based on the product code. If a product record is missing within a time slice, cross-platform imputation is performed based on the same type of product records in adjacent time slices. The imputation value is determined by recursively extrapolating the sales change direction and magnitude in adjacent time slices.

[0018] The missing product records are imputed based on the imputed values. The imputed time slices and the time slices without missing product records are combined in chronological order to determine the fusion structure matrix.

[0019] Furthermore, when segmenting the cumulative sales field according to the window span based on the fusion structure matrix, and determining the slope of cumulative sales change, the width of the fluctuation envelope, and the progressive time perturbation within each segment based on the time gradient, and constructing the segment feature vector, the process includes:

[0020] The sequence parsing unit determines the window span based on the fusion structure matrix, divides the cumulative sales field into several continuous segments based on the window span, performs a difference operation on the cumulative sales field in each continuous segment, and determines the slope of the cumulative sales change based on the difference result and the time gradient.

[0021] Furthermore, when segmenting the cumulative sales field according to the window span based on the fusion structure matrix, and determining the slope of cumulative sales change, the width of the fluctuation envelope, and the progressive time perturbation within each segment based on the time gradient, and constructing the segment feature vector, the method further includes:

[0022] The maximum and minimum cumulative sales are determined based on each continuous segment. The fluctuation envelope width is determined based on the maximum and minimum cumulative sales. The progressive time disturbance is determined based on the interval change between adjacent time indices within each continuous segment. The slope of the cumulative sales change, the fluctuation envelope width, and the progressive time disturbance are normalized to determine the segment feature vector.

[0023] Furthermore, when constructing a sequence feature chain from the feature vectors of each segment based on the segment order, it includes:

[0024] The sequence parsing unit determines the segment order based on the time index of each consecutive segment, and arranges the feature vectors of each segment based on the segment order to determine the initial segment feature sequence.

[0025] Based on the start time index and end time index of each continuous segment, the sorted segment feature vectors are labeled with segment identifiers, and an index association relationship is established between the segment identifiers and the corresponding segment feature vectors. Based on the index association relationship, the sorted segment feature vectors are sequentially written into the sequence buffer, and the forward position index and backward position index of adjacent segment feature vectors are recorded in the sequence buffer. The sequence feature chain is determined based on the arrangement result of the segment feature vectors in the sequence buffer and the position index relationship.

[0026] Furthermore, when determining the sales forecast based on the segment-by-segment extrapolation of the slope change direction, fluctuation envelope extension, and time disturbance accumulation according to the sequence feature chain, the process includes:

[0027] The trend inference unit determines the slope direction label and slope amplitude parameter of each continuous segment based on the cumulative sales change slope of the feature vector of each segment in the sequence feature chain, determines the fluctuation adjustment coefficient based on the fluctuation envelope width of the feature vector of each segment, and determines the rhythm adjustment coefficient based on the progressive time disturbance of the feature vector of each segment.

[0028] Furthermore, when determining the sales forecast based on the segment-by-segment extrapolation of the slope change direction, fluctuation envelope extension, and time disturbance accumulation according to the aforementioned sequence feature chain, the method also includes:

[0029] When performing segment-by-segment extrapolation, the extrapolation amount is calculated using the slope direction label and slope amplitude parameter, and the fluctuation adjustment coefficient and rhythm adjustment coefficient are used as correction amounts. The predicted sales increment of the extrapolation segment is calculated segment by segment. Based on the extrapolation results, the predicted sales increment of each extrapolation segment is superimposed with the current sales to determine the sales prediction sequence. The predicted sales amount is determined based on the sales prediction sequence.

[0030] Furthermore, when determining the sales forecast quantity based on the sales forecast sequence, the process includes:

[0031] The trend analysis unit selects initial sales forecast values ​​within the preset forecast period from the sales forecast sequence based on the start and end times of the preset forecast period.

[0032] The time weight is determined based on the time interval between each initial sales forecast value within the preset forecast period, and the sales forecast amount is determined by weighting the initial sales forecast values ​​within the preset forecast period.

[0033] Furthermore, when performing a difference calculation on the current inventory quantity based on the predicted sales volume, and determining the replenishment quantity based on the result of the difference calculation, the process includes:

[0034] The trend analysis unit determines the inventory difference based on the current inventory level and the sales forecast.

[0035] The inventory difference is compared with the inventory difference threshold. If the inventory difference is greater than or equal to the inventory difference threshold, no replenishment is performed. If the inventory difference is less than the inventory difference threshold, the replenishment quantity is determined based on the relationship between the inventory difference and the inventory difference threshold.

[0036] Compared with existing technologies, the beneficial effects of this invention are as follows: By unifying the field format, time granularity, and unit system of order records, outbound records, and other data according to the heterogeneous data protocols of various cross-border e-commerce platforms, and constructing a collected dataset using timestamp numbers and platform codes as indexes, this replaces simple field matching, improves data integration efficiency, and avoids data distortion caused by field mapping deviations and unit conversion errors, laying the foundation for collaborative analysis of multi-platform data. Through time resampling and product code merging, the dispersion of time and product dimensions in multi-platform data is eliminated, achieving standardized aggregation of cross-platform data. This allows users to accurately grasp the overall dynamics of the same product across multiple platforms, avoiding the limitations of only focusing on overall trends. This comprehensively captures the dynamic fluctuations and time disturbances in sales. Based on segment-by-segment extrapolation of sequence feature chains, the accuracy and reliability of sales forecasting are improved. Combining the current inventory quantity with the calculation of inventory difference to determine replenishment quantity achieves accurate matching between replenishment quantity and inventory demand, optimizing inventory forecasting and replenishment decisions, reducing the risk of stockouts and backlogs, and improving the overall responsiveness and reliability of the supply chain. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a functional block diagram of a cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system provided in an embodiment of the present invention. Detailed Implementation

[0039] 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.

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] See Figure 1 As shown in some embodiments of this application, a cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system includes:

[0042] The data acquisition unit is configured to process the field formats, time granularity, and unit formats of order records, outbound records, inbound records, and product status change records based on the heterogeneous data protocols of various cross-border e-commerce platforms, and to construct the acquisition dataset by indexing the processed data with timestamp numbers and platform codes.

[0043] The structural fusion unit is configured to resample the data from each cross-border e-commerce platform based on the timestamp number in the collected dataset, and merge the records of the same product from multiple platforms based on the product code to determine the fusion structure matrix.

[0044] The sequence parsing unit is configured to segment the cumulative sales field according to the window span based on the fusion structure matrix, and determine the slope of the cumulative sales change, the width of the fluctuation envelope, and the progressive time perturbation amount within each segment based on the time gradient, and construct segment feature vectors. Based on the segment order, the feature vectors of each segment are constructed into a sequence feature chain.

[0045] The trend extrapolation unit is configured to extrapolate the slope change direction, fluctuation envelope extension and time disturbance accumulation segment by segment based on the sequence feature chain to determine the sales forecast, perform difference calculation on the current inventory based on the sales forecast, and determine the replenishment quantity based on the result of the difference calculation.

[0046] Specifically, the data acquisition unit operates based on the heterogeneous data protocols of various cross-border e-commerce platforms. These heterogeneous data protocols refer to the different data transmission and storage rules adopted by different cross-border e-commerce platforms due to differences in their technical architecture and business rules (for example, Amazon's order records have a time granularity accurate to the second, with the field name "Order Time" and the currency unit being USD; AliExpress's time granularity is accurate to the minute, with the field name "Transaction Time" and the currency unit being RMB; these are examples of heterogeneous data protocols). For order records, outbound records, inbound records, and product status change records, the unit processes their field format, time granularity, and unit format. Field format processing unifies fields with the same meaning across different platforms to avoid data loss due to incompatibility. Time granularity processing converts times with different precisions to the same granularity (e.g., all converted to seconds). Unit format processing unifies units such as currency, weight, and quantity (e.g., converting RMB to USD and grams to kilograms). The processed data is indexed using timestamp numbers and platform codes to construct a collected dataset. The timestamp number is a unique numerical code that identifies a specific moment (e.g., 1740000000 represents January 1, 2026, 00:00:00, ensuring uniqueness and traceability in the time dimension). The platform code is a unique identifier assigned to each cross-border e-commerce platform (e.g., Amazon code "AMZ", AliExpress code "AE", Shopee code "SPE"). Using timestamp numbers and platform codes as indexes, the collected dataset can quickly locate data from different platforms and at different times. For example, the original data for an Amazon order is "Order Time: 2026-01-15 10:30:25, Transaction Amount: $100, Product Code: G001". The AliExpress order data for the same product is: Transaction Time: 2026-01-15 At 10:30, the amount was 690 RMB, and the product code was G001. The data collection unit standardized the time to a second-level timestamp: 1740033025 (Amazon) and 1740033000 (AliExpress). The amount was standardized to USD (690 RMB is equivalent to 96 USD at the exchange rate). The fields were standardized to order time, order amount, and product code. The data was then included in the collection dataset using the timestamp + platform code as an index (e.g., 1740033025-AMZ, 1740033000-AE). This avoided the barriers of heterogeneous data from different platforms and replaced simple field matching. It improved data processing efficiency and avoided data distortion caused by field mismatch, unit conversion deviation, and chaotic time granularity, laying the foundation for collaborative data processing across multiple platforms.The structure fusion unit resamples data from various cross-border e-commerce platforms based on timestamps in the collected dataset. Time resampling integrates discrete data from different times at a unified granularity (e.g., merging data within 5-minute or 1-hour granularities) to eliminate time fragmentation issues across different platforms. Then, it merges records of the same product across multiple platforms based on product codes, which are unique identifiers assigned to each product (e.g., G001 represents a sweatshirt). Finally, it determines the fusion structure matrix, a structured data carrier where cells contain data such as order volume, outbound volume, and inbound volume for the product across multiple platforms at the corresponding time granularity. For example, if time resampling is set to a 1-hour granularity, and the collected dataset shows 5 orders for product G001 on AMZ platform and 3 orders for product G001 on AE platform during the period from 10:00 to 11:00 on January 15th, the structure fusion unit will treat this period as a row, G001 as a column, and the cells will record the merged order volume of 8 items, while also associating corresponding outbound and inbound volume data to form the fusion structure matrix. Transforming scattered multi-platform data into a standardized structured matrix eliminates data fragmentation, ensures the consistency of the time dimension and the aggregation of the product dimension of multi-platform data, and improves the accuracy of collaborative analysis of multi-platform data, providing a unified data system for inventory forecasting.

[0047] It is understandable that the sequence parsing unit segments the cumulative sales field according to the window span based on the fusion structure matrix. The cumulative sales field is the cumulative sales quantity of a certain product within the corresponding time granularity recorded in the fusion structure matrix. The window span is the length of the time segment used to segment the cumulative sales field (e.g., 1 day, 3 days is a window span, which can be flexibly adjusted according to the product sales cycle). Cumulative sales represent the total sales volume of a product across all connected cross-border e-commerce platforms from the start date of the statistics to a specific point in time. Within each segment, the sequence parsing unit determines the slope of cumulative sales change, the width of the fluctuation envelope, and the progressive time perturbation based on the time gradient. The time gradient is the minimum interval of time progression within the segment (e.g., for a segment with a window span of 1 day, the time gradient can be set to 1 hour, i.e., sales changes are analyzed hourly). The slope of cumulative sales change is the rate at which cumulative sales change with the time gradient within the segment. The slope of cumulative sales change is the cumulative sales at the end of the segment minus the cumulative sales at the beginning of the segment) divided by the segment length. A positive slope indicates that the cumulative sales within the segment are generally on an upward trend, while a negative slope indicates that the cumulative sales are generally on a downward trend. The width of the fluctuation envelope reflects the amplitude of cumulative sales fluctuations, and the progressive time perturbation quantifies the impact of time deviation on the cumulative sales statistics. These three indicators are normalized to construct segment feature vectors. The segment feature vectors integrate the cumulative sales change slope, fluctuation envelope width, and progressive time disturbance characteristics of a single segment to form a multi-dimensional data carrier. Based on the segment order, the feature vectors of each segment are constructed into a sequence feature chain. The sequence feature chain is a continuous feature sequence formed by arranging the feature vectors of multiple segments in chronological order (e.g., according to the segment order of January 1, January 2, and January 3, the corresponding feature vectors are sequentially linked to reflect dynamic changes). This avoids the limitation of only focusing on the overall sales trend, thereby accurately extracting the dynamic change characteristics of different segments, and comprehensively capturing the upward and downward trends, fluctuation amplitude, and irregular disturbance factors, thus improving the accuracy of prediction.The trend extrapolation unit extrapolates segment by segment based on the sequence feature chain, considering the slope change direction, fluctuation envelope extension, and time disturbance accumulation. Segment by segment extrapolation follows the temporal order of the sequence feature chain, combining the characteristic change patterns of preceding segments to deduce the characteristic trends of subsequent segments. Slope change direction extrapolation predicts whether sales will rise, fall, or remain stable in subsequent segments; fluctuation envelope extension extrapolation predicts the change in the cumulative sales fluctuation amplitude in subsequent segments; and time disturbance accumulation extrapolation summarizes the impact of disturbance factors from preceding segments to predict potential disturbances in subsequent segments. Based on the extrapolation results, the sales forecast is determined, representing the potential sales quantity of goods in subsequent segments. The difference between the current inventory level and the sales forecast is then calculated. The current inventory level is the actual available inventory quantity of goods (excluding unsellable inventory such as those that have been shipped but not yet delivered or damaged). The replenishment quantity is the quantity of goods needed to make up for inventory gaps and ensure that the predicted sales volume is met. Based on multi-dimensional dynamic characteristics, the prediction results can accurately reflect the dynamic fluctuation pattern of market demand, overcome the limitations of prediction based on the average value and trend line fitting of a single platform, and avoid prediction deviations caused by ignoring sales fluctuations and disturbance factors. At the same time, the replenishment quantity is determined by calculating the difference between inventory and predicted sales volume, achieving a precise match between replenishment quantity and inventory demand. This avoids order loss due to stockouts and prevents inventory backlog caused by excessive replenishment, thereby optimizing the inventory turnover efficiency of the cross-border e-commerce supply chain.

[0048] In some embodiments of this application, when processing the field formats, time granularity, and unit systems of order records, outbound records, inbound records, and product status change records based on the heterogeneous data protocols of various cross-border e-commerce platforms, and constructing a collection dataset using timestamp numbers and platform codes as indexes, the process includes: the data collection unit mapping the field names, field lengths, and field types of each cross-border e-commerce platform; comparing the fields of different cross-border e-commerce platforms item by item; redefining the types of fields that cannot be mapped; truncating or padding fields that cannot be mapped based on differences in field length; determining the mapping result; and processing the time fields of each cross-border e-commerce platform. The process involves converting time steps, quantizing minute, hour, or day-level time fields, mapping initial time values ​​to a unified time grid based on the quantization results, determining the time step conversion result, performing unit conversion on quantity, amount, and weight fields, matching conversion coefficients based on the corresponding initial unit labels, and performing coefficient operations between the conversion coefficients and initial field values ​​to determine the unit conversion result. The mapping result, time step conversion result, and unit conversion result are written to the same record, along with a timestamp number and platform code. A sequential linked list is constructed based on the timestamp number in an ascending order, and the link relationship between predecessor and successor nodes is recorded in the linked list nodes to construct the collected dataset.

[0049] Specifically, the data acquisition unit compares the fields of order records, outbound records, inbound records, and product status change records from each platform item by item. For fields with different names but the same business meaning, they are directly mapped to unified standardized fields. For fields that exist only on some platforms and cannot be directly matched across all platforms, the field type is redefined according to the field type to ensure that the field structure of data records across all platforms is consistent. For fields with the same meaning but different lengths, the content of fields exceeding the length is truncated, while the content of shorter fields is padded with specific placeholder characters to meet the length requirements, and finally the mapping result is determined. For example, since the meaning of OrderID on Amazon and Order Number on AliExpress is the same, they are uniformly mapped to a standardized field. However, niche cross-border platforms do not have this type of order identifier field. The data collection unit redefines the field type at the corresponding position to be consistent with other platforms, leaving the content blank. If the system sets the order identifier field length to 32 characters, AliExpress's 20-digit order number will be padded to 32 characters, and a Southeast Asian platform's 35-digit order identifier field will be truncated to 32 characters, retaining only the first 32 core characters. This eliminates data heterogeneity across different platforms at the field attribute level, solving the problem of data mismatch caused by inconsistent field names, lengths, and types, thereby improving the efficiency and accuracy of field integration. The time step conversion of the time field standardizes and aligns the inconsistent time granularity across different platforms. The data collection unit quantizes the time fields of each platform according to a unified time step, converting all initial time fields at the minute, hour, or day level into a time format matching the unified time step. Then, based on the quantization results, the initial time values ​​of each platform are mapped to a unified time grid to determine the time step conversion result. For example, if the system sets a unified time step of 1 minute and the corresponding unified time grid is year-month-day hour:minute, then an hourly time field on a certain platform, such as 2026-01-15 10:00, is quantized as 2026-01-15 10:00 and mapped to the corresponding time grid. Similarly, a dayly time field on another platform, 2026-01-15, is quantized as 2026-01-15 00:00 and mapped to the corresponding time grid. A minute-level time field, 2026-01-15 10:30, is directly matched to the corresponding time grid. By unifying the time standard across all platforms, data misalignment during subsequent time resampling due to differences in time precision is avoided, providing a precise time benchmark for aligning the time dimensions of data across multiple platforms. The data acquisition unit matches the corresponding conversion coefficients based on the initial unit labels of the quantity, amount, and weight fields for each platform, according to preset unit conversion rules. Then, the conversion coefficients are calculated with the initial field values ​​to determine the final unit conversion result.The conversion factor comes from a standardized unit conversion rule library. This library is determined by combining common standards in the cross-border e-commerce industry and unit specifications published by various platforms. For example, the system sets the unified units to "piece," "USD," and "kg." An order's initial amount field value is 690 CNY, with the initial unit label "CNY." After matching the conversion factor, it is calculated to be approximately 95.83 USD. Similarly, a product's initial weight field value is 500g, with the initial unit label "g." After matching the conversion factor, it is calculated to be 0.5kg. By eliminating the unit format barriers between different platforms, it solves the data statistics errors caused by inconsistent units (e.g., directly summing RMB and USD amounts). The data acquisition unit converts the mapping results and time step conversion results. The unit conversion results are integrated and written into the same data record to ensure that a single record contains complete standardized data of a specific cross-border e-commerce platform at a specific moment. A corresponding timestamp number and platform code are added to this record. The timestamp number is a numerical code that can uniquely identify a specific moment, and its value corresponds to the chronological order of time. The larger the value, the later the time. The platform code is a unique character or number identifier assigned to each cross-border e-commerce platform. A sequential linked list is constructed based on the timestamp number in an ascending manner. The link relationship between the predecessor node and the successor node is recorded in each node of the linked list. The predecessor node is the node before a certain node. The successor node is the node after a certain node in the sequential linked list constructed in ascending order of timestamp number. For example, a record integrates order field mapping results, minute-level time step conversion results, and USD amount unit conversion results. A timestamp of 1740033000 and platform code AE ​​are added. The predecessor node of this node is the record node with timestamp 1740032940, and the successor node is the record node with timestamp 1740033060. All nodes are linked sequentially according to their timestamps, forming a complete dataset. Organizing data using a sequential linked list structure enables fast data traversal and querying, avoiding the low processing efficiency caused by unordered storage. The dual index of timestamp and platform code accurately locates the time and source platform of each data entry, providing a clear and standardized data index basis for time resampling and product record merging.

[0050] In some embodiments of this application, when performing time resampling on each cross-border e-commerce platform based on the timestamp number in the collected dataset, and merging multi-platform records of the same product based on the product code to determine the fusion structure matrix, the process includes: the structure fusion unit selecting the resampling time step based on the timestamp number in the collected dataset, dividing the data records of each cross-border e-commerce platform into several time slices according to the resampling time step, filtering each product record based on the product code within each time slice, and if there are missing product records within a time slice, cross-platform interpolation is performed based on the same type of product records in adjacent time slices, and the interpolation value is determined by recursion based on the sales change direction and change magnitude of adjacent time slices, and the missing product records are interpolated based on the interpolation value, and the interpolated time slices and time slices without missing product records are combined in chronological order to determine the fusion structure matrix.

[0051] Specifically, the structure fusion unit reads the timestamp numbers of all data records in the collected dataset, and combines the resampling time step selected by the business requirements (such as 1 hour or 1 day). Based on this time step, the time axis of the collected dataset is divided into several continuous and non-overlapping time slices. Each time slice corresponds to a fixed time interval, and all data records are assigned to the corresponding time slice according to their own timestamp numbers. For example, the timestamps of the collected dataset are at the second level, such as 1740033000 corresponding to 2026-01-15 10:00:00. The structure fusion unit, combined with the merchant's hourly inventory analysis needs, selects a resampling time step of 1 hour, and divides the time axis into continuous time slices such as 2026-01-15 10:00:00-10:59:59 and 2026-01-15 11:00:00-11:59:59. All records with second-level timestamps are assigned to the corresponding time slices according to their intervals, thereby integrating the discrete second-level timestamp data into structured time slice data, while adapting to different business analysis needs. For each time slice that has been divided, the structural fusion unit uses the product code as the filtering basis to extract and summarize the records of the same product code from all cross-border e-commerce platforms within the same time slice, forming a multi-platform merged record of a certain product within that time slice. For different product codes within the same time slice, they are independently filtered and summarized to ensure that the record of each product is independently classified within the time slice.For example, in the time slice from 10:00:00 to 10:59:59 on January 15, 2026, the structural fusion unit filters records with product code G001, extracting 5 order records from Amazon (AMZ) and 3 order records from AliExpress (AE). These two records are then aggregated into a multi-platform merged record for product G001 within this time slice (total order quantity 8). Simultaneously, records with product code G002 are filtered, extracting 2 order records from Shopee (SPE), forming an independent record for product G002. This achieves aggregated analysis of the same product data across multiple platforms based on product codes. Missing product records occur when a product code has no records on any cross-border e-commerce platform within a specific time slice (e.g., product G001 has no records on any cross-border e-commerce platform within January 15, 2026). (No orders or outbound records are available in the time slice from 11:00:00 to 11:59:59). Cross-platform imputation utilizes the multi-platform merged records of the product code in adjacent time slices to fill in missing data. The structural fusion unit traverses the product records of each time slice. If a missing record is detected for a certain product code, the multi-platform merged records of the adjacent time slices (i.e., the previous and next time slices) of that product code are retrieved. Then, the sales change direction of the adjacent time slices is analyzed (e.g., 8 units sold in the previous slice and 4 units sold in the next slice, with a decreasing trend). The sales change magnitude of the adjacent time slices is then calculated (e.g., 8 units sold in the previous slice and 12 units sold in the next slice, with a change magnitude of +4 units). Based on the change direction and magnitude, the imputation value of the missing time slice is determined and filled into the missing position to complete the product record imputation.For example, product code G001 has no records in the time slice of 11:00-11:59. Its combined sales volume in the preceding time slice of 10:00-10:59 was 8 units, and in the following time slice of 12:00-12:59 it was 12 units, showing an upward trend. The time interval between two adjacent time slices is two resampling time steps (2 hours), with a total change of +4 units. Therefore, the change in a single time step is +2 units. Based on this, the interpolated value for the 11:00-11:59 time slice is 10 units. The structural fusion unit fills the G001 product record in this time slice with 10 units, interpolating the missing product records based on the interpolated value, thus resolving the issue of missing records in certain time zones. To address the issue of missing product records due to lack of transactions or missed data collection, and to avoid the risk of failing to extract continuous sales characteristics due to data breakpoints, the interpolation value is recursively derived based on the sales change patterns of adjacent slices, ensuring the reliability of the interpolated data. The structural fusion unit arranges the interpolated time slices together with the original time slices without missing product records in ascending order of timestamp numbers. Then, using the time slices as rows of a matrix and the product codes as columns, the final fusion structure matrix is ​​determined. The fusion structure matrix provides support for multi-platform collaboration of the system, enabling supply chain decisions to formulate replenishment strategies based on global data rather than data from a single platform, reducing the risk of over-replenishment or stockouts caused by fragmented platform data.

[0052] In some embodiments of this application, when the cumulative sales field is segmented according to the window span based on the fusion structure matrix, and the cumulative sales change slope, fluctuation envelope width and progressive time perturbation amount are determined based on the time gradient in each segment, and the segment feature vector is constructed, the following steps are included: the sequence parsing unit determines the window span based on the fusion structure matrix, divides the cumulative sales field into several continuous segments based on the window span, performs a difference operation on the cumulative sales field in each continuous segment, and determines the cumulative sales change slope based on the difference result and the time gradient.

[0053] In some embodiments of this application, when dividing the cumulative sales field into segments according to the window span based on the fusion structure matrix, and determining the cumulative sales change slope, fluctuation envelope width, and progressive time perturbation amount based on the time gradient within each segment, and constructing the segment feature vector, the method further includes: determining the maximum and minimum cumulative sales values ​​based on each continuous segment, determining the fluctuation envelope width based on the maximum and minimum cumulative sales values, determining the progressive time perturbation amount based on the interval change between adjacent time indices within each continuous segment, normalizing the cumulative sales change slope, fluctuation envelope width, and progressive time perturbation amount, and determining the segment feature vector.

[0054] Specifically, the sequence parsing unit reads the cumulative sales field from the fusion structure matrix. This field represents the cumulative sales volume of a product across multiple platforms, recorded in the fusion structure matrix. Then, it groups and merges consecutive time slices in the fusion structure matrix according to a preset window span, dividing it into several consecutive, non-overlapping segments. The window span is the number of time slices used for merging, and its length is determined comprehensively based on business characteristics such as the product's sales cycle and the frequency of market demand fluctuations. For example, if the time slices in the fusion structure matrix are divided by hours, and the preset window span is the number of time slices corresponding to several consecutive hours, the sequence parsing unit merges these consecutive time slices into one segment. Each segment corresponds to a continuous duration, covering all cumulative sales data for a product within that duration. Processing the continuous time series data in stages avoids data loss caused by performing overall analysis of massive amounts of time slice data, allowing each segment to focus on the sales change characteristics within a specific duration. The maximum cumulative sales value is the highest cumulative sales value among all time slices within a single segment, and the minimum cumulative sales value is the lowest cumulative sales value among all time slices within a single segment. The fluctuation envelope width reflects the range of fluctuation in cumulative sales within a segment. For each segment, the sequence parsing unit iterates through the cumulative sales field of all time slices within the segment, filters out the maximum and minimum cumulative sales values, and then determines the fluctuation envelope width of the segment based on these two values. For example, if a segment covers several consecutive time slices, after iterating through the cumulative sales data of these time slices, the highest and lowest values ​​are filtered out. The fluctuation envelope width can intuitively reflect the range of sales fluctuation from the lowest to the highest value within that period. The fluctuation envelope width accurately quantifies the degree of sales fluctuation within each segment, making up for the deficiency of inventory forecasting, which only focuses on the overall average sales and ignores the fluctuation amplitude. A time index is an identifier assigned to each time slice within a segment, reflecting its sequential position within the segment. The interval variation between adjacent time indices represents a deviation from the standard pattern caused by unforeseen factors, where time slices originally arranged in a fixed order deviate from this pattern. The progressive time perturbation represents the degree of irregular impact of this interval variation on cumulative sales. The sequence analysis unit assigns consecutive time indices to time slices within each segment, determines the location and magnitude of interval variations, and, based on the impact of these variations on sales, determines the progressive time perturbation for that segment, thereby capturing the impact of irregular time factors on sales.For example, if the time index interval of time slices within a certain segment is in a fixed order, and the actual interval between two adjacent time slices deviates from the interval due to factors such as logistics delays or sudden platform activities, the sequence parsing unit will determine the corresponding progressive time disturbance based on the impact of this interval change on sales. This reflects the sales fluctuation caused by the sudden factor, thereby capturing the impact of irregular time disturbances such as order delays and sudden activities on sales. This transforms previously unquantifiable disturbance factors into structured features, improving the adaptability of subsequent sales forecasts to market dynamics. For each segment, the cumulative sales change slope is determined based on the time gradient. The time gradient is the minimum interval of time progression within the segment, and the cumulative sales change slope is the rate at which the cumulative sales within the segment change with the time gradient, reflecting the overall upward and downward trend of sales during that period. The cumulative sales change slope, fluctuation envelope width, and progressive time disturbance are normalized to eliminate the magnitude differences between different features, ensuring that the segment feature vector can evenly reflect the three types of characteristics of sales trends, fluctuations, and disturbances, providing a data foundation for accurate sales forecasting.

[0055] In some embodiments of this application, when constructing a sequence feature chain from the feature vectors of each segment based on the segment order, the process includes: a sequence parsing unit determining the segment order based on the time index of each continuous segment; arranging the feature vectors of each segment based on the segment order to determine an initial segment feature sequence; labeling the sorted segment feature vectors with segment identifiers based on the start and end time indices of each continuous segment; establishing an index association relationship between the segment identifiers and the corresponding segment feature vectors; writing the sorted segment feature vectors sequentially into a sequence buffer based on the index association relationship; recording the forward and backward position indices of adjacent segment feature vectors in the sequence buffer; and determining the sequence feature chain based on the arrangement result and position index relationship of the segment feature vectors in the sequence buffer.

[0056] Specifically, the sequence parsing unit retrieves all generated segment feature vectors and determines the order of segments based on the time index corresponding to each segment. The time index is an identifier that reflects the position of a segment on the overall time axis and directly corresponds to the time slice order of the fusion structure matrix. All segment feature vectors are arranged sequentially according to time to form the initial segment feature sequence. For example, if the segmentation of a product covers multiple consecutive time periods from morning to night, and each segment has a corresponding time index, the sequence parsing unit arranges the segment feature vectors of the morning period first, the noon period next, and the evening period last, in chronological order to form the initial segment feature sequence. This ensures the temporal continuity of the initial segment feature sequence, avoids temporal misalignment caused by disordered generation order of segments, and ensures that subsequent trend inference can be based on feature sequences that conform to actual time patterns. The start time index is the time index corresponding to the first time slice contained in a single segment, and the end time index is the time index corresponding to the last time slice contained in a single segment. The segment identifier is a unique marker generated by combining the start and end time indices of a segment, which can uniquely identify the time range of that segment. The sequence parsing unit extracts the corresponding start and end time indices for each segment feature vector in the sorted initial segment feature sequence. Based on these two indices, a unique segment identifier is marked for the segment feature vector, and the segment identifier is bound to the corresponding segment feature vector to establish a one-to-one index association relationship, allowing each segment feature vector to be traced back to its corresponding specific time range through the segment identifier. The segment identifier achieves precise binding between segment feature vectors and time ranges, avoids confusion of segment feature vectors from different segments, and improves the accuracy and efficiency of data retrieval. The sequence buffer is a temporary storage space for sorted and labeled segment feature vectors. The forward position index is the position identifier of the previous segment feature vector within the sequence buffer, and the backward position index is the position identifier of the next segment feature vector within the sequence buffer. The sequence parsing unit writes the segment feature vectors with established index relationships sequentially into the sequence buffer according to the initial segment feature sequence. During the writing process, the corresponding forward and backward position indices are recorded for each segment feature vector. For the first segment feature vector in the sequence, only the backward position index is recorded; for the last segment feature vector in the sequence, only the forward position index is recorded; and for the middle segment feature vectors, both forward and backward position indices are recorded. The sequence buffer achieves ordered storage of segment feature vectors, avoiding data loss or out-of-order processing. The forward and backward position indices allow independent feature vectors to form bidirectional relationships, avoiding the risk that discrete feature vectors cannot reflect the temporal relationship between preceding and following segments.The sequence parsing unit, based on the already arranged segment feature vectors in the sequence buffer, and combining the forward and backward position indices corresponding to each segment feature vector, concatenates all segment feature vectors into a continuous chain structure in chronological order. This chain structure, ordered by time and linked by position indices, is the sequence feature chain. Transforming independent segment feature vectors into a sequence feature chain with continuous temporal correlation completely preserves the trajectory of product sales characteristics changing over time. Simultaneously, the chain structure supports subsequent trend extrapolation units in performing segment-by-segment traversal and extrapolation analysis, improving the efficiency of trend extrapolation.

[0057] In some embodiments of this application, when extrapolating the slope change direction, fluctuation envelope extension, and time disturbance accumulation segment by segment based on the sequence feature chain to determine the sales forecast, the method includes: the trend extrapolation unit determines the slope direction label and slope amplitude parameter of each continuous segment based on the cumulative sales change slope of the feature vector of each segment in the sequence feature chain, determines the fluctuation adjustment coefficient based on the fluctuation envelope width of the feature vector of each segment, and determines the rhythm adjustment coefficient based on the progressive time disturbance of the feature vector of each segment.

[0058] In some embodiments of this application, when extrapolating the slope change direction, fluctuation envelope extension, and time disturbance accumulation segment by segment based on the sequence feature chain to determine the sales forecast, the method further includes: when performing segment-by-segment extrapolation, using the slope direction label and slope amplitude parameter for extrapolation, and using the fluctuation adjustment coefficient and rhythm adjustment coefficient as correction amounts, performing segment-by-segment calculations on the predicted sales increment of the extrapolation segment, superimposing the predicted sales increment of each extrapolation segment with the current sales based on the extrapolation results to determine the sales forecast sequence, and determining the sales forecast based on the sales forecast sequence.

[0059] Specifically, the slope direction label is a marker generated based on the cumulative sales change slope in the segment feature vector, used to identify the direction of sales change. The slope magnitude parameter is a parameter extracted based on the magnitude of the cumulative sales change slope in the segment feature vector, used to reflect the rate of sales change; the larger the absolute value of the slope, the more the corresponding parameter value reflects the rapid changes in sales. The fluctuation adjustment coefficient is a correction parameter determined based on the fluctuation envelope width in the segment feature vector, used to adapt to the impact of sales fluctuation amplitude on the prediction results; the larger the fluctuation envelope width, the more the corresponding adjustment coefficient reflects the degree of interference of fluctuation on sales. The rhythm adjustment coefficient is a correction parameter determined based on the progressive time perturbation in the segment feature vector, used to adapt to the impact of irregular time perturbations on sales; the larger the progressive time perturbation, the more the corresponding adjustment coefficient reflects the sales deviation caused by time perturbations. The trend inference unit traverses all segment feature vectors in the sequence feature chain, determines the direction of change of the cumulative sales change slope of each segment and generates a corresponding slope direction label, then extracts the slope magnitude feature and converts it into a slope magnitude parameter. Simultaneously, for the fluctuation envelope width of each segment, based on the degree of sales fluctuation it reflects, it is converted into a corresponding fluctuation adjustment coefficient. Similarly, for the progressive time disturbance of each segment, based on the degree of time disturbance it reflects, it is converted into a corresponding rhythm adjustment coefficient. In other words, the trend inference unit traverses the feature vector of each segment in the sequence feature chain, extracts the cumulative sales change slope, and then determines the positive, negative, or zero value of the cumulative sales change slope, generating slope direction labels (positive slope corresponds to an upward label, negative slope to a downward label, and zero value to a stable label). The absolute value of the cumulative sales change slope is mapped to a slope amplitude parameter; the larger the absolute value of the slope, the higher the corresponding slope amplitude parameter value, indicating a faster rate of sales change during that period. Fluctuations are categorized based on the sales fluctuation patterns of similar products within the industry. Finally, the fluctuation envelope width is mapped to a fluctuation adjustment coefficient; the larger the fluctuation envelope width, the higher the corresponding fluctuation adjustment coefficient value, indicating a stronger interference of sales fluctuations during that period with the prediction results. The progressive time disturbance reflects the irregular impact of sudden time factors (such as platform activities and logistics delays) on sales. Combining the direction (positive disturbance increases sales, negative disturbance decreases sales) and magnitude of the disturbance, the progressive time disturbance is mapped to a rhythm adjustment coefficient with directional attributes. Positive disturbances correspond to positive coefficients, and negative disturbances correspond to negative coefficients; the greater the disturbance, the higher the absolute value of the coefficient. The sales trend, fluctuation, and disturbance characteristics abstracted from the sequence feature chain are transformed into labels and coefficients that can be directly used for extrapolation calculations, providing clear parameter support for quantitative prediction and improving the accuracy and reliability of dynamic prediction. The extrapolation segment is the segment corresponding to the target period for sales prediction after the time range covered by the sequence feature chain. The predicted sales increment is the change in sales in the extrapolated segment compared to the final sales of the last segment in the sequence feature chain.The current sales volume is the cumulative sales volume of the product corresponding to the last moment of the last segment of the sequence feature chain. The sales forecast sequence is a continuous sales sequence formed by arranging the predicted sales volumes of multiple extrapolation segments in chronological order. In the segment-by-segment extrapolation process, the trend extrapolation unit first determines the direction of sales change in the extrapolation segment using the slope direction label, and determines the basic rate of sales change in that direction using the slope amplitude parameter. The two are combined to generate the basic extrapolation amount of the extrapolation segment. The fluctuation adjustment coefficient is used as the first-level correction to correct the basic extrapolation amount, offsetting the interference of sales fluctuation amplitude on the prediction result. Then, the rhythm adjustment coefficient is used as the second-level correction to further correct the extrapolation amount after fluctuation correction, offsetting the impact of irregular time disturbances on sales. After double correction, the predicted sales increment of each extrapolation segment is obtained. The hierarchical calculation mode of basic extrapolation amount and double correction not only retains the trend of sales change, but also specifically offsets the interference of fluctuation and time disturbance, solving the problem that traditional trend fitting prediction ignores dynamic interference factors, thereby improving the accuracy of the prediction result. The trend extrapolation unit superimposes the predicted sales increment of each extrapolated segment with the current sales corresponding to the sequence feature chain to obtain the predicted sales of a single extrapolated segment. All the predicted sales of extrapolated segments are arranged in chronological order to form a continuous sales prediction sequence. That is, if the current sales of the sequence feature chain is the cumulative sales at a certain moment, and the predicted sales increment of a certain extrapolated segment is the increment corresponding to the increase, the two are superimposed to obtain the predicted sales of that extrapolated segment. Multiple extrapolated segments are arranged chronologically to form a sales prediction sequence covering multiple future periods. The sales data for a specific period needed by the merchant is extracted from the sales prediction sequence, which is the predicted sales volume for that period. By superimposing current sales to generate the sales prediction sequence, the continuity between predicted and historical data is ensured, avoiding prediction gaps. At the same time, the sales prediction sequence can cover multiple future periods, meeting the inventory planning needs of merchants in different dimensions, such as short-term and long-term cycles. The sales prediction volume extracted based on the sequence can accurately reflect the market demand trend of the corresponding period, enabling the sales prediction volume to truly reflect the dynamic changes in market demand, thereby achieving a precise match between inventory demand and replenishment volume.

[0060] In some embodiments of this application, when determining the sales forecast amount based on the sales forecast sequence, the process includes: a trend inference unit selecting initial sales forecast values ​​within the preset forecast period from the sales forecast sequence based on the start and end times of the preset forecast period, determining time weights based on the time intervals of each initial sales forecast value within the preset forecast period, and performing a weighted calculation on each initial sales forecast value within the preset forecast period to determine the sales forecast amount.

[0061] In some embodiments of this application, when performing a difference calculation on the current inventory quantity based on the sales forecast and determining the replenishment quantity based on the result of the difference calculation, the process includes: a trend analysis unit determining the inventory difference based on the current inventory quantity and the sales forecast, comparing the inventory difference with an inventory difference threshold, and if the inventory difference is greater than or equal to the inventory difference threshold, then no replenishment is performed; if the inventory difference is less than the inventory difference threshold, then the replenishment quantity is determined based on the relationship between the inventory difference and the inventory difference threshold.

[0062] Specifically, by pre-setting a forecast period to define the analysis scope and combining it with time weights to optimize forecast accuracy, discrete forecast values ​​are transformed into precise sales forecasts that align with business needs. The pre-set forecast period is a sales forecast time range predetermined by the merchant based on the replenishment rhythm and inventory planning requirements of cross-border e-commerce. Its start and end times precisely match actual business actions (such as weekly replenishment and monthly inventory checks). Initial sales forecast values ​​are extracted from the sales forecast sequence, representing sales forecast data for each segment within this period. The time interval refers to the time distance between each segment within the period and between each segment and the end of the period. The time weight is the allocation ratio determined based on this interval; the shorter the time interval and the closer to the end of the period, the higher the weight. The weighted calculation involves assigning each initial sales forecast value to its corresponding time weight. After multiplication and summation, the trend analysis unit first reads the start and end times of the merchant's preset forecast period. Using this as a condition, it iterates through the sales forecast sequence, filtering initial sales forecast values ​​within the preset forecast period and excluding data outside this range. For example, if the merchant presets a one-week forecast period for weekly replenishment, it only filters data for the next week's segmented time periods, excluding data from the second week to avoid bias. Time weights are determined based on the time intervals of the initial sales forecast values ​​within the preset forecast period, assigning higher weights to periods closer to the end of the period. For example, weekends and periods near the weekend in a weekly period are assigned higher weights than weekdays. Time weights strengthen the reference for recent trends. Through weighted calculations, all initial sales forecast values ​​within the preset forecast period are summed to obtain the sales forecast for that period. Accurate replenishment decisions are achieved by comparing the inventory difference with the inventory difference threshold. The inventory difference is the difference between the current inventory level and the predicted sales volume, directly reflecting the inventory's ability to cover the predicted sales. The inventory difference threshold is a safety stock threshold preset based on logistics cycles and demand fluctuations. The trend analysis unit calculates the inventory difference based on the current inventory level and the sales forecast, and then compares it with the inventory difference threshold. If the inventory difference is greater than or equal to the inventory difference threshold, it means that the current inventory can meet the forecast sales and there is sufficient safety stock to cope with sudden fluctuations, so no replenishment is needed. If the inventory difference is less than the inventory difference threshold, it means that the inventory is insufficient to cover demand, and the replenishment quantity is determined based on the difference between the two to ensure that the inventory difference after replenishment can reach or exceed the inventory difference threshold. This avoids the risk of stockouts and prevents inventory backlog caused by blind replenishment, thus adapting to the long logistics cycle of cross-border e-commerce and improving the stability of the supply chain.

[0063] In summary, the beneficial effects of this invention are as follows: By unifying the field formats, time granularity, and unit systems of order records, outbound records, and other data according to the heterogeneous data protocols of various cross-border e-commerce platforms, and constructing a collected dataset using timestamp numbers and platform codes as indexes, this replaces simple field matching, improves data integration efficiency, and avoids data distortion caused by field mapping deviations and unit conversion errors, laying the foundation for collaborative analysis of multi-platform data. Through time resampling and product code merging, the dispersion of time and product dimensions in multi-platform data is eliminated, achieving standardized aggregation of cross-platform data. This allows users to accurately grasp the overall dynamics of the same product across multiple platforms, avoiding the limitations of only focusing on overall trends. This comprehensively captures the dynamic fluctuations and time disturbances in sales. Based on segment-by-segment extrapolation of sequence feature chains, the accuracy and reliability of sales forecasting are improved. Combining the current inventory quantity with the calculation of inventory difference to determine replenishment quantity achieves accurate matching between replenishment quantity and inventory demand, optimizing inventory forecasting and replenishment decisions, reducing the risk of stockouts and backlogs, and improving the overall responsiveness and reliability of the supply chain.

[0064] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A cross-border e-commerce supply chain multi-platform collaboration and inventory dynamic prediction management system, characterized in that, include: The data acquisition unit is configured to process the field format, time granularity, and unit format of order records, outbound records, inbound records, and product status change records based on the heterogeneous data protocols of various cross-border e-commerce platforms, and to construct the acquisition dataset by indexing the processed data with timestamp numbers and platform codes. The structural fusion unit is configured to resample the data from each cross-border e-commerce platform based on the timestamp number in the collected dataset, and merge the records of the same product from multiple platforms based on the product code to determine the fusion structure matrix. The sequence parsing unit is configured to segment the cumulative sales field according to the window span based on the fusion structure matrix, and determine the slope of cumulative sales change, the width of fluctuation envelope and the progressive time perturbation amount in each segment based on the time gradient, and construct segment feature vectors, and construct a sequence feature chain based on the segment order of each segment feature vector. The trend extrapolation unit is configured to extrapolate the slope change direction, fluctuation envelope extension and time disturbance accumulation segment by segment based on the sequence feature chain to determine the sales forecast, perform difference calculation on the current inventory based on the sales forecast, and determine the replenishment quantity based on the result of the difference calculation. 2.The cross-border e-commerce supply chain multi-platform collaboration and inventory dynamic prediction management system according to claim 1, characterized in that, When processing the field formats, time granularity, and unit systems of order records, outbound records, inbound records, and product status change records based on the heterogeneous data protocols of various cross-border e-commerce platforms, and constructing the collected dataset using timestamp numbers and platform codes as indexes, the following steps are included: The data acquisition unit maps the field names, field lengths, and field types of each cross-border e-commerce platform. It compares the fields of different cross-border e-commerce platforms one by one, redefines the types of fields that cannot be mapped, and truncates or fills the fields that cannot be mapped based on the differences in field length to determine the mapping result. Time step conversion is performed on the time fields of various cross-border e-commerce platforms. The time fields at the minute, hour, or day level are quantized by step size. Based on the quantization results, the initial time values ​​are mapped to a unified time grid point to determine the time step conversion result. The quantity field, amount field, and weight field are converted to units. A conversion factor is matched based on the corresponding initial unit label, and the conversion factor is calculated with the initial field value to determine the unit conversion result. The mapping result, time step conversion result, and unit conversion result are written into the same record, and a timestamp number and platform code are written into the record. A sequential linked list is constructed based on the timestamp number in an ascending manner, and the link relationship between the predecessor node and the successor node is recorded in the linked list node to construct the collected dataset. 3.The cross-border e-commerce supply chain multi-platform coordination and inventory dynamic prediction management system according to claim 2, characterized in that, When resampling the timestamps of each cross-border e-commerce platform based on the timestamps in the collected dataset, and merging records of the same product across multiple platforms based on the product code to determine the fusion structure matrix, the process includes: The structural fusion unit selects the resampling time step based on the timestamp number of the collected dataset, and divides the data records of each cross-border e-commerce platform into several time slices according to the resampling time step. Within each time slice, each product record is filtered based on the product code. If a product record is missing within a time slice, cross-platform imputation is performed based on the same type of product records in adjacent time slices. The imputation value is determined by recursively extrapolating the sales change direction and magnitude in adjacent time slices. The missing product records are imputed based on the imputed values. The imputed time slices and the time slices without missing product records are combined in chronological order to determine the fusion structure matrix. 4.The cross-border e-commerce supply chain multi-platform coordination and inventory dynamic prediction management system according to claim 3, characterized in that, When segmenting the cumulative sales field according to the window span based on the fusion structure matrix, and determining the slope of cumulative sales change, the width of the fluctuation envelope, and the progressive time perturbation within each segment based on the time gradient, and constructing the segment feature vector, the process includes: The sequence parsing unit determines the window span based on the fusion structure matrix, divides the cumulative sales field into several continuous segments based on the window span, performs a difference operation on the cumulative sales field in each continuous segment, and determines the slope of the cumulative sales change based on the difference result and the time gradient.

5. The cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system according to claim 4, characterized in that, When dividing the cumulative sales field into segments according to the window span based on the fusion structure matrix, and determining the slope of cumulative sales change, the width of the fluctuation envelope, and the progressive time perturbation within each segment based on the time gradient, and constructing the segment feature vector, the method further includes: The maximum and minimum cumulative sales are determined based on each continuous segment. The fluctuation envelope width is determined based on the maximum and minimum cumulative sales. The progressive time disturbance is determined based on the interval change between adjacent time indices within each continuous segment. The slope of the cumulative sales change, the fluctuation envelope width, and the progressive time disturbance are normalized to determine the segment feature vector.

6. The cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system according to claim 5, characterized in that, When constructing a sequence feature chain from the feature vectors of each segment based on the segment order, the following is included: The sequence parsing unit determines the segment order based on the time index of each consecutive segment, and arranges the feature vectors of each segment based on the segment order to determine the initial segment feature sequence. Based on the start time index and end time index of each continuous segment, the sorted segment feature vectors are labeled with segment identifiers, and an index association relationship is established between the segment identifiers and the corresponding segment feature vectors. Based on the index association relationship, the sorted segment feature vectors are sequentially written into the sequence buffer, and the forward position index and backward position index of adjacent segment feature vectors are recorded in the sequence buffer. The sequence feature chain is determined based on the arrangement result of the segment feature vectors in the sequence buffer and the position index relationship.

7. The cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system according to claim 6, characterized in that, When determining the sales forecast based on the segment-by-segment extrapolation of the slope change direction, fluctuation envelope extension, and time disturbance accumulation according to the sequence feature chain, the following steps are included: The trend inference unit determines the slope direction label and slope amplitude parameter of each continuous segment based on the cumulative sales change slope of the feature vector of each segment in the sequence feature chain, determines the fluctuation adjustment coefficient based on the fluctuation envelope width of the feature vector of each segment, and determines the rhythm adjustment coefficient based on the progressive time disturbance of the feature vector of each segment.

8. The cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system according to claim 7, characterized in that, When determining the sales forecast based on the segment-by-segment extrapolation of the slope change direction, fluctuation envelope extension, and time disturbance accumulation according to the sequence feature chain, the method further includes: When performing segment-by-segment extrapolation, the extrapolation amount is calculated using the slope direction label and slope amplitude parameter, and the fluctuation adjustment coefficient and rhythm adjustment coefficient are used as correction amounts. The predicted sales increment of the extrapolation segment is calculated segment by segment. Based on the extrapolation results, the predicted sales increment of each extrapolation segment is superimposed with the current sales to determine the sales prediction sequence. The predicted sales amount is determined based on the sales prediction sequence.

9. The cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system according to claim 8, characterized in that, Determining the sales forecast based on the sales forecast sequence includes: The trend analysis unit selects initial sales forecast values ​​within the preset forecast period from the sales forecast sequence based on the start and end times of the preset forecast period. The time weight is determined based on the time interval between each initial sales forecast value within the preset forecast period, and the sales forecast amount is determined by weighting the initial sales forecast values ​​within the preset forecast period.

10. The cross-border e-commerce supply chain multi-platform collaboration and dynamic inventory forecasting management system according to claim 9, characterized in that, When performing a difference calculation on the current inventory quantity based on the sales forecast, and determining the replenishment quantity based on the result of the difference calculation, the following steps are included: The trend analysis unit determines the inventory difference based on the current inventory level and the sales forecast. The inventory difference is compared with the inventory difference threshold. If the inventory difference is greater than or equal to the inventory difference threshold, no replenishment is performed. If the inventory difference is less than the inventory difference threshold, the replenishment quantity is determined based on the relationship between the inventory difference and the inventory difference threshold.

Images