Single item storage demand prediction method and system based on multi-source data

Abstract

The invention discloses a single-item storage demand prediction method and system based on multi-source data, and relates to the technical field related to intelligent storage management.The method comprises the steps that single-item associated data are extracted from a multi-source database, and semantic tags are extracted through an NLP and a knowledge graph; activating an explicit-implicit coupling factor model of the storage demand, the explicit-implicit coupling factor model comprising an explicit-implicit factor extraction layer and an explicit-implicit coupling prediction layer, and establishing a single product prediction result; and reading real-time warehouse inventory data, performing balance analysis of multi-target scheduling according to the real-time warehouse inventory data and the single-item prediction result, and establishing a single-item warehouse real demand by using a balance analysis result. The technical problems of insufficient utilization of multi-source heterogeneous data, lack of dominant and recessive factor coupling analysis and insufficient dynamic inventory scheduling capability in the prior art are solved, and the technical effects of enhancing the multi-target collaborative optimization capability, improving the warehousing demand prediction precision and realizing intelligent adaptive inventory management are achieved.

Description

Technical Field

[0001] The present application relates to the technical field related to intelligent warehouse management, and in particular to a method and system for predicting single-product warehouse demand based on multi-source data. Background Art

[0002] With the rapid development of e-commerce and supply chain management, accurately forecasting individual product warehousing demand has become crucial for companies to optimize inventory management, reduce costs, and improve operational efficiency. Traditional warehousing demand forecasting relies primarily on structured data such as historical sales data and seasonal trends. This makes it difficult to fully tap into the potential information contained in multi-source, heterogeneous data (such as product descriptions, user reviews, and marketing copy), resulting in limited forecasting accuracy. This is especially true in complex market environments such as promotional events and new product launches. Relying solely on explicit factors (such as sales volume and inventory turnover) struggles to accurately capture changes in market demand, leading to discrepancies between forecasts and actual warehousing demand. Furthermore, warehousing demand forecasting must dynamically coordinate with real-time inventory data to balance multiple constraints, such as inventory costs, out-of-stock risks, and logistics efficiency. Existing methods struggle to adapt to the rapid changes in warehousing demand in complex market environments, hindering the implementation of more accurate, adaptive, and efficient inventory management.

[0003] At the current stage, relevant technologies have technical problems such as insufficient utilization of multi-source heterogeneous data, lack of coupling analysis of explicit and implicit factors, and insufficient dynamic inventory scheduling capabilities. Summary of the Invention

[0004] This application solves the technical problems existing in the prior art, such as insufficient utilization of multi-source heterogeneous data, lack of analysis of coupling of explicit and implicit factors, and insufficient dynamic inventory scheduling capabilities, by providing a method and system for predicting single-product warehouse demand based on multi-source data. It achieves the technical effects of enhancing multi-objective collaborative optimization capabilities, improving the accuracy of warehouse demand prediction, and realizing intelligent adaptive inventory management.

[0005] The present application provides a single-product warehousing demand prediction method based on multi-source data, the method comprising: extracting single-product related data from a multi-source database, and extracting semantic tags using NLP and knowledge graphs, wherein the single-product related data includes single-product descriptions, user comments, and activity copy; activating an explicit and implicit coupling factor model of warehousing demand, wherein the explicit and implicit coupling factor model includes an explicit and implicit factor extraction layer and an explicit and implicit coupling prediction layer, extracting explicit factors and implicit factors using the explicit and implicit factor extraction layer, synchronizing the explicit factors, the implicit factors, and the semantic tags to the explicit and implicit coupling prediction layer, and establishing a single-product prediction result; reading real-time warehousing inventory data, performing a balance analysis of multi-objective scheduling based on the real-time warehousing inventory data and the single-product prediction result, and establishing the actual single-product warehousing demand using the balance analysis result.

[0006] In a possible implementation, the method for predicting single-product warehousing demand based on multi-source data also performs the following processing: when any data is input into the explicit and implicit factor extraction layer, the data is classified explicitly and implicitly through a multidimensional discriminator, and the discrimination dimensions of the multidimensional discriminator include data structure, data source, data observability, and data verifiability; the data identified as explicit classification is normalized and time-windowed, and encoded into a fixed-length vector; the data identified as implicit classification is processed through nonlinear mapping and behavioral sparse smoothing to establish an implicit factor vector; explicit factors are established based on the fixed-length vector, and implicit factors are established based on the implicit factor vector.

[0007] In a possible implementation, the method for predicting single-product warehousing demand based on multi-source data also performs the following processing: after the explicit factors and the implicit factors are synchronized to the explicit-implicit coupling prediction layer, factor attention analysis is performed under three-dimensional weight features based on the factor attention mechanism to establish factor weights, and the three-dimensional weight features include time weight, scene weight, and label weight, and the label weight is constructed through the semantic label; after performing weighted mapping of explicit factors and implicit factors according to the factor weights, the weighted results of the explicit factors, the weighted results of the implicit factors and the semantic labels are input into the prediction channel of the explicit-implicit coupling prediction layer to perform correction prediction and establish the single-product prediction result.

[0008] In a possible implementation, the method for predicting single-product warehousing demand based on multi-source data further performs the following processing: after parsing the semantic labels, the explicit path is initialized through context enhancement, and the implicit path is adjusted through semantic perception, the explicit path and the implicit path are calculation paths of the prediction channel; the weighted result of the explicit factor is input into the initialized explicit path to establish the main signal output; the weighted result of the implicit factor is input into the adjusted implicit path to establish a dynamic correction factor; after compensating the main signal output according to the dynamic correction factor, the single-product prediction result is established.

[0009] In a possible implementation, the method for predicting single-product warehousing demand based on multi-source data also performs the following processing: structurally transforming the real-time warehousing inventory data to generate inventory status data, wherein the inventory status data includes current inventory, available inventory, safety inventory, unit storage cost, and turnover speed; configuring optimization goals for multi-objective scheduling, wherein the optimization goals include inventory satisfaction goals, cost control goals, turnover efficiency goals, and fluctuation smoothness goals; utilizing the optimization goals to perform a balance analysis based on inventory status data and the single-product prediction results to establish the actual demand for single-product warehousing.

[0010] In a possible implementation, the method for predicting single-product warehousing demand based on multi-source data also performs the following processing: parsing the semantic labels and establishing a basic weight template based on the semantic labels; calling historical decision errors and using the historical decision errors to perform feedback adjustment on the basic weight template, and completing the weight configuration of the optimization target based on the feedback adjustment results.

[0011] In a possible implementation, the method for predicting single-product warehousing demand based on multi-source data also performs the following processing: activating a scene recognition mechanism, using the scene recognition mechanism to perform scene recognition, and establishing a scene recognition result; configuring a scene optimization template according to the scene recognition result, and using the scene optimization template to perform the basic weight template compensation.

[0012] The present application also provides a single product warehousing demand prediction system based on multi-source data, the system including: a semantic label extraction unit, used to extract single product related data from a multi-source database, and use NLP and knowledge graphs to extract semantic labels, wherein the single product related data includes single product descriptions, user comments, and activity copy; a single product prediction result establishment unit, used to activate the explicit and implicit coupling factor model of warehousing demand, the explicit and implicit coupling factor model includes an explicit and implicit factor extraction layer and an explicit and implicit coupling prediction layer, after using the explicit and implicit factor extraction layer to extract explicit factors and implicit factors, the explicit factors, the implicit factors, and the semantic labels are synchronized to the explicit and implicit coupling prediction layer to establish a single product prediction result; a single product warehousing real demand establishment unit, used to read real-time warehousing inventory data, perform a balance analysis of multi-objective scheduling based on the real-time warehousing inventory data and the single product prediction results, and use the balance analysis results to establish the single product warehousing real demand.

[0013] This application proposes a method and system for predicting single-product warehousing demand based on multi-source data. The system extracts single-product related data from a multi-source database and uses NLP and knowledge graphs to extract semantic tags. The system activates the explicit and implicit coupling factor model of warehousing demand, which includes an explicit and implicit factor extraction layer and an explicit and implicit coupling prediction layer, to establish single-product prediction results. The system reads real-time warehousing inventory data, performs a multi-objective scheduling balance analysis based on the real-time warehousing inventory data and the single-product prediction results, and uses the balance analysis results to establish the true demand for single-product warehousing. This solves the technical problems of insufficient utilization of multi-source heterogeneous data, lack of explicit and implicit factor coupling analysis, and insufficient dynamic inventory scheduling capabilities in the prior art, achieving the technical effects of enhancing multi-objective collaborative optimization capabilities, improving warehousing demand prediction accuracy, and realizing intelligent adaptive inventory management. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention are briefly introduced below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in precise order. Instead, various steps may be processed in reverse order or simultaneously as needed. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0015] Figure 1 A flowchart of a method for predicting single-product warehousing demand based on multi-source data provided in an embodiment of the present application; Figure 2 A schematic diagram of a process for extracting explicit and implicit factors in a single product warehousing demand forecasting method based on multi-source data provided in an embodiment of the present application.

[0016] Figure 3 A schematic diagram of the structure of a single product warehousing demand forecasting system based on multi-source data provided in an embodiment of the present application.

[0017] Description of the accompanying drawings: semantic label extraction unit 10, single product prediction result establishment unit 20, single product warehousing actual demand establishment unit 30. DETAILED DESCRIPTION

[0018] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0019] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0020] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict, and the terms “first\second” involved are merely to distinguish similar objects and do not represent a specific ordering of the objects. The terms “including” and “having” and any variations are intended to cover non-exclusive inclusions. For example, a process, method, product, or server that includes a series of steps is not necessarily limited to those steps that are clearly listed, but may include other steps that are not clearly listed or that are inherent to these processes, methods, systems, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. The terms used herein are for the purpose of describing the embodiments of this application only.

[0021] The present application embodiment provides a method for predicting the storage demand of a single product based on multi-source data, such as Figure 1 As shown, the method includes: Step S100: extract single product related data from a multi-source database and extract semantic tags using NLP and knowledge graph, wherein the single product related data includes single product descriptions, user comments, and activity copy.

[0022] Preferably, single product related data is collected from multi-source databases, where the multi-source databases may include product management systems, e-commerce platforms, social media, etc.; single product related data includes single product descriptions, such as product name, specifications, function descriptions, materials and applicable scenarios, etc. structured text, which provides basic characteristics of the product and is used to construct entity attributes in the knowledge graph; user comments, such as consumer feedback, rating content, emotional tendencies and other unstructured texts, use NLP to analyze emotional polarity (positive / negative) and keywords to explore hidden needs or potential problems; activity copywriting, such as promotional advertisements, marketing language, activity rules, etc., to identify the short-term impact of marketing activities on demand.

[0023] Preferably, natural language processing (NLP) is used for text cleaning and entity recognition, that is, to remove irrelevant symbols in the text and extract key attributes from the product description, and then sentiment analysis is performed based on user comments, that is, to quantify the sentiment score from the user comments, such as "poor battery life" corresponds to a negative label, and finally keyword extraction is performed based on the activity copy, including identifying the promotion strength from the activity copy and establishing high-dimensional semantic labels; then, using knowledge graph construction, the extracted entities (such as products, functions, user emotions) are associated as graph nodes and semantic relationships are established, and finally, these data are converted into quantifiable structured semantic labels, such as "high demand seasonality", "user preference portability", "high promotion sensitivity", etc.

[0024] Step S200: activating the explicit and implicit coupling factor model of warehousing demand, wherein the explicit and implicit coupling factor model includes an explicit and implicit factor extraction layer and an explicit and implicit coupling prediction layer. After extracting explicit factors and implicit factors using the explicit and implicit factor extraction layer, the explicit factors, the implicit factors, and the semantic labels are synchronized to the explicit and implicit coupling prediction layer to establish a single product prediction result.

[0025] Preferably, the explicit and implicit coupling factor model is a composite analysis model for warehouse demand forecasting. The core lies in simultaneously quantifying explicit factor indicators and potential implicit factor indicators, including a two-level structure of explicit and implicit factor extraction layer and explicit and implicit coupling prediction layer; then the explicit and implicit coupling factor model is activated, that is, the explicit and implicit factor extraction layer is used for extraction, and then the explicit and implicit coupling prediction layer is used to establish the single product prediction results. Specifically, explicit factors are objective quantitative indicators directly related to warehouse demand, including historical sales data, inventory turnover rate, seasonal coefficient, promotion intensity, real-time inventory level, etc. For example, the historical sales surge of a certain product during the "Double 11" period is an explicit factor.

[0026] Preferably, latent factors are potential influencing factors mined through semantic analysis of multi-source data or derived through semantic tags. These may include user review sentiment, such as negative reviews of product quality that may inhibit repeat purchases; market popularity trends, such as sudden hot discussions on social media; product feature popularity, such as the frequency of mentions of "waterproof" in reviews; competitive product dynamics, such as the substitution effect brought about by price reductions of similar products; and marketing copy urgency, such as the rush to buy caused by "limited quantity" announcements. The extracted explicit and implicit factors and semantic tags are then synchronized as input data to the explicit and implicit coupled prediction layer. Specifically, through an attention mechanism or Bayesian network, the contribution of explicit and implicit factors is automatically adjusted based on the relevance of the factors to the target product. A neural network is then used to model the complex interactive relationships between explicit and implicit factors, such as the positive feedback effect of explicit promotion intensity and implicit user emotions, and the possibility that large discounts may amplify the traffic-generating effect of user reviews. Finally, a warehouse demand forecast result, i.e., a single product forecast, is output based on the comprehensive explicit and implicit features. For example, it can be identified that although the current sales (explicit) of a certain product are stable, the accumulation of implicit negative reviews may cause a 20% drop in demand in the next two weeks. Through the explicit and implicit coupling factor model, demand fluctuation signals that cannot be detected by relying solely on explicit data can be identified earlier, such as predicting the risk of unsold goods through implicit negative reviews, which in turn helps with intelligent warehouse scheduling.

[0027] Further, such as Figure 2As shown, step S200 also includes step S210, when any data is input into the explicit and implicit factor extraction layer, the data is classified into explicit and implicit categories through a multidimensional discriminator, and the discrimination dimensions of the multidimensional discriminator include data structure, data source, data observability, and data verifiability; step S220, the data identified as explicit classification are normalized and time window processed, and encoded into a fixed-length vector; step S230, the data identified as implicit classification are processed through nonlinear mapping and behavioral sparse smoothing to establish a latent factor vector; step S240, explicit factors are established according to the fixed-length vector, and latent factors are established according to the latent factor vector.

[0028] Preferably, when any raw data, such as sales records, user comment logs, etc., is input into the explicit and implicit factor extraction layer, the data is classified into explicit and implicit categories through a multidimensional discriminator, wherein the multidimensional discriminator automatically divides data types based on a preset four-dimensional discrimination rule, and the discrimination dimensions include data structure, data source, data observability, and data verifiability. Specifically, with respect to data structure, numerical table data, fields, or structured data such as sales statistics in the ERP system are classified as explicit factors, and unstructured text, behavior flow data, or text log data such as click sequences are classified as implicit factors; with respect to data sources, clear indicators from internal business systems such as sales data and inventory databases are considered explicit, and external platform or embedded point data such as social media sentiment analysis results, click and comment data, and search keyword frequency are considered implicit; with respect to observability, indicators that can be directly monitored in real time, such as the order volume of the day, are explicit, and features that need to be extracted or inferred through the model, such as the probability of user purchase intention, are implicit; with respect to verifiability, data that supports historical backtesting, such as financial statements, are classified as explicit, and data with subjectivity or diversity, such as comment sentiment polarity, are classified as implicit. For example, the 30-day return rate of a certain product is identified as an explicit factor because it comes from a structured ERP system and is verifiable; while the frequency of mentions in the grass-roots posts on the Xiaohongshu platform is identified as an implicit factor because it needs to be extracted from unstructured text and is difficult to trace back.

[0029] Preferably, standardization processing is performed on the data marked as explicit classification, including normalization and time window alignment, that is, scaling the values of different units such as sales and inventory to the interval [0, 1] to eliminate dimensional differences, and then aggregating the data according to a unified period (such as a 7-day sliding window) to ensure time series consistency, such as rolling average of weekly sales into a monthly trend vector, and finally encoding it into a fixed-length vector, that is, converting the processed explicit data into a numerical vector with unified dimension, such as sales sequence, current branch warehouse inventory, sales turnover rate, return rate and average daily sales growth rate, etc. The three indicators of nearly 7 days of sales, inventory turnover rate, and promotion discount intensity are encoded as [0.82, 0.45, 0.93].

[0030] Preferably, data identified as latent classifications is processed through nonlinear mapping and behavioral sparse smoothing. Specifically, text, behavioral sequences, etc. are mapped to a low-dimensional semantic space through a deep learning model (such as Transformer), high-order patterns in the text / behavior are extracted, and the emotional word frequency is mapped to an emotional intensity scalar. Discrete behavioral data such as the number of add-to-carts are then Gaussian smoothed to eliminate noise interference and generate dense and stable latent factor vectors, such as search popularity index, comment sentiment score, add-to-cart conversion rate, and marketing copy strength score. Explicit factors are established based on fixed-length vectors, which are expressed as standardized vectors that can directly participate in numerical calculations and reflect objective business indicators. Latent factors are established based on latent factor vectors, encapsulating the quantitative expression of complex behaviors or emotions. They need to be decoded collaboratively with the explicit factors through a coupling layer, thereby achieving comparability and computability of heterogeneous data in a unified space.

[0031] Furthermore, step S200 also includes step S250, after the explicit factors and the implicit factors are synchronized to the explicit-implicit coupling prediction layer, factor attention analysis is performed under three-dimensional weight features based on the factor attention mechanism to establish factor weights, and the three-dimensional weight features include time weights, scene weights, and label weights, and the label weights are constructed through the semantic labels; step S260, after performing weighted mapping of explicit factors and implicit factors according to the factor weights, the explicit factor weighted results, the implicit factor weighted results and the semantic labels are input into the prediction channel of the explicit-implicit coupling prediction layer to perform correction prediction and establish the single product prediction result.

[0032] Preferably, the explicit and implicit coupling prediction layer deeply integrates explicit factors, implicit factors and semantic labels through dynamic weight allocation and multi-channel coordination mechanism, and finally outputs highly robust single product demand prediction results. Specifically, a dynamic weight system of three dimensions of time, scene and label is constructed through the factor attention mechanism, including time weight, scene weight and label weight. Among them, the time weight quantifies the effectiveness of the factor over time. For example, the weight of explicit sales data in the promotion period is higher than that in the normal period, while the emotional weight of implicit user comments in the early stage of new product launch is higher; the scene weight is to identify the importance difference of factors in different business scenarios. For example, the weight of the limited-time purchase semantic label of the activity copy in the 618 promotion scenario is increased to 2 to 3 times that of the ordinary scenario; the label weight constructs association rules based on semantic labels. For example, when a quality control risk label appears in the implicit factor, the confidence weight of the explicit historical sales is automatically reduced.

[0033] Preferably, three-dimensional weights are applied to explicit factors such as normalized inventory levels and implicit factors such as user add-to-cart conversion rates, generating explicit weighted results (e.g., peak season sales volume × time weight of 1.2) and implicit weighted results (e.g., negative review sentiment value × scenario weight of 0.8). The weighted results of the explicit and implicit factors, along with the semantic labels, are then input into the prediction channel of the explicit-implicit coupling prediction layer for corrective prediction. Specifically, semantic labels are used as regulators in the weighting process to adjust the weighted results of the explicit and implicit factors. The adjusted weighted explicit and implicit factors and semantic labels are then corrected and predicted using an LSTM temporal network and a Bayesian correction module. The explicit pathway primarily processes weighted numerical indicators and establishes a basic prediction curve. The implicit pathway identifies potential demand fluctuations through behavioral pattern analysis. The semantic correction channel uses label logic rules for final calibration and correction, outputting a single-product prediction result that simultaneously meets both numerical rigor (explicit dominance) and market sensitivity (implicit supplementation). For example, this can predict that demand for a product may decline in the later stages of a promotion due to the accumulation of implicit negative reviews.

[0034] Furthermore, step S260 also includes step S261, after parsing the semantic label, initializing the explicit path through context enhancement, and adjusting the implicit path through semantic perception, the explicit path and the implicit path are the calculation paths of the prediction channel; step S262, inputting the weighted result of the explicit factor into the initialized explicit path to establish the main signal output; step S263, inputting the weighted result of the implicit factor into the adjusted implicit path to establish a dynamic correction factor; step S264, after compensating the main signal output according to the dynamic correction factor, establishing the single product prediction result.

[0035] Preferably, the explicit and implicit coupling prediction layer integrates semantic labels, explicit factors and implicit factors through a dual-path collaborative computing mechanism, and finally generates an accurate single-product warehousing demand forecast result. Specifically, the semantic labels are parsed to extract structured business scenario features, such as promotion period, inventory clearance, etc., and the explicit path is initialized through context enhancement, that is, the context perception ability of the explicit path is enhanced according to the extracted features. For example, under the end-of-season clearance tag, the attenuation coefficient of historical sales data is automatically increased, and historical data for the same period is automatically loaded, so that the explicit path pays more attention to short-term fluctuations rather than long-term trends; the implicit path is dynamically adjusted through semantic perception, that is, the network weight is reconstructed using the emotional polarity and behavioral patterns in the tag, and the calculation logic of the implicit path is dynamically adjusted. For example, when a quality complaint tag is detected, the implicit path strengthens the sensitivity of the negative comment factor, so that the implicit path can flexibly select the optimal implicit signal parsing strategy for different semantic scenarios; among them, the explicit path and the implicit path are the calculation paths of the prediction channel.

[0036] Preferably, the initialized explicit pathway processes the weighted results of the explicit factors and outputs a primary signal, reflecting a baseline forecast based on historical patterns and current values. For example, a sales curve for the next 30 days is generated through a time series model. This is highly stable, but may overlook sudden market fluctuations. The adjusted implicit pathway processes the weighted results of the implicit factors and outputs a dynamic correction factor, which quantifies the offsetting effect of the implicit signal on the primary signal. For example, a demand attenuation coefficient of -0.15 caused by negative user reviews or a traffic gain of +0.2 caused by trending searches. The dynamic correction factor is updated in real time using lightweight models such as wavelet transforms to ensure sensitive response to subtle market changes. Finally, the primary signal output is compensated based on the dynamic correction factor. Specifically, when the implicit pathway detects a positive signal, positive compensation is applied, i.e., the primary forecast value is adjusted upward according to logical rules. When a risk signal appears, negative compensation is applied, i.e., the primary forecast value is gradually adjusted downward. Finally, a single product forecast is established, ensuring both business interpretability (explicit dominance) and the acuity of machine intelligence (implicit calibration).

[0037] Step S300 , read the real-time warehouse inventory data, perform a balance analysis of multi-objective scheduling based on the real-time warehouse inventory data and the single product forecast results, and use the balance analysis results to establish the actual demand for single product storage.

[0038] Preferably, the multi-objective balance analysis mechanism dynamically coordinates the contradictory constraints of real-time warehouse data and forecast results, conducts a balance analysis of the multi-objective scheduling of real-time warehouse inventory data and single product forecast results, and finally outputs the real demand for single product storage that conforms to the actual business logic. Specifically, it reads the real-time warehouse inventory data, such as the current available inventory of 3,000 pieces and 2,000 pieces in transit, compares the real-time inventory data with the single product forecast results, such as the demand forecast of 5,000 pieces in the next 7 days, identifies whether there is a gap, and if the forecast value exceeds the inventory safety threshold, triggers the dynamic adjustment mechanism; then, according to the difference between the real-time warehouse inventory data and the single product forecast results, constructs a dynamic adjustment mechanism. Quantitative conflict indicators are established as optimization goals, including inventory satisfaction goals, cost control goals, turnover efficiency goals, and fluctuation smoothing goals. Among them, the inventory satisfaction goal ensures that the spot rate is not lower than the threshold (such as 95%), and replenishment is given priority when the real-time inventory is too low; the cost control goal suppresses the risk of unsalable goods caused by excessive stocking, and constrains the maximum stocking quantity through the inventory holding cost coefficient; the turnover efficiency goal is used to optimize the inventory days, such as requiring a turnover of ≤5 days for food products, and initiating an inventory reduction strategy for products with near expiration dates; the fluctuation smoothing goal avoids drastic fluctuations in demand between adjacent cycles, such as a sudden increase of 200% in the forecast value this week, and reducing supply chain pressure through a trend smoothing algorithm.

[0039] Preferably, balance analysis refers to the use of the Pareto optimal principle for multi-objective decision-making, including the use of inventory level calibration, emergency scenario priority, volatility suppression, etc. to determine the balance analysis results. Among them, inventory level calibration means that when the forecast value requires 1,000 pieces of stock but the real-time inventory shows that there are still 300 pieces of available inventory, the actual demand is reduced to 700 pieces in combination with the turnover target and cost target; emergency scenario priority means that if the real-time inventory is in crisis and the satisfaction target is about to exceed the threshold, such as the spot rate has dropped to 92%, even if the forecast value has not reached the peak, the demand will still be increased according to the supply priority principle, and the cost overrun warning will be triggered at the same time; volatility suppression means that when the forecast value increases by 150% compared with last week, the demand will be released in stages through a smoothing algorithm to avoid warehouse throughput overload.

[0040] Preferably, the balance analysis results are corrected, including positive and negative corrections. Specifically, if a new product's social media popularity surges due to hidden channels, with a predicted demand of 500 units, but the real-time inventory is 0 and the supplier's minimum order quantity is 1,000 units, the balance analysis will accept a partial sacrifice in the cost target; the satisfaction target weight will be increased to 70%, generating a real demand of 1,000 units; and the product will be simultaneously marked as a high-risk new product, triggering turnover efficiency monitoring. If a slow-moving product has a predicted demand of 200 units, but the real-time inventory remains at 800 units, with a turnover period of 120 days, the balance analysis will force the real demand back to zero and initiate a clearance strategy; at the same time, the category weight will be deducted from the satisfaction target, and resources will be allocated to high-turnover products. This ensures that when the predicted value conflicts with the real-time inventory, the final real demand for the single product is output, thereby improving the accuracy of warehouse demand forecasting and dynamic scheduling capabilities.

[0041] Furthermore, step S300 also includes step S310, performing structural transformation on the real-time warehouse inventory data to generate inventory status data, wherein the inventory status data includes current inventory, available inventory, safety inventory, unit storage cost, and turnover speed; step S320, configuring optimization goals for multi-objective scheduling, wherein the optimization goals include inventory satisfaction goals, cost control goals, turnover efficiency goals, and fluctuation smoothness goals; step S330, utilizing the optimization goals to perform a balance analysis based on the inventory status data and the single product forecast results to establish the actual demand for single product storage.

[0042] Preferably, real-time warehouse inventory data is structurally transformed to obtain quantitative indicators with business semantics, i.e., inventory status data is generated. Specifically, these indicators include current inventory, available inventory, safety stock, unit storage cost, and turnover rate. Current inventory represents the actual number of items in stock, reflecting immediate supply capacity. Available inventory represents the available inventory after deducting locked orders. For example, inventory occupied by promotional pre-sales is not included in the current replenishment calculation. Safety stock refers to the minimum buffer threshold set based on historical fluctuations. For example, the safety stock for a certain category is the average daily sales volume × 7 days. Unit storage cost includes comprehensive costs such as storage occupancy, insurance, and loss. For example, the cost coefficient for fresh products is 3 times that of standard products. Turnover rate refers to the number of days the current inventory can support. Turnover days is the ratio of current inventory to the average sales volume over the past 7 days. For example, if real-time data for a certain product is converted into inventory status, the current inventory is 1,200 items, the available inventory is 800 items, the safety stock is 300 items, the unit storage cost is 0.5 yuan / day / item, and the turnover rate is 20 days.

[0043] Preferably, multi-objective scheduling optimization objectives are configured, including inventory fulfillment, cost control, turnover efficiency, and fluctuation smoothing. These optimization objectives are then used to perform a balance analysis based on inventory status data and individual product forecasts. This involves a multi-dimensional decision-making process based on inventory status and forecast values. For example, if forecast demand is 1,000 units, available inventory + in-transit inventory is 900 units, and safety stock is 300 units, and if the fulfillment objective is weighted heavily (e.g., during a major sales event), the actual demand is calculated as 1,000 units + 300 units - 900 units = 400 units. If the cost objective is dominant (e.g., during the end-of-quarter liquidation period), only the safety stock is replenished, resulting in a demand of 300 units. If turnover rate is detected to be less than the safety threshold, even if the forecast value does not reach the trigger level, the difference between the safety threshold and the current turnover is calculated and multiplied by the average daily sales volume to generate excess demand. For a forecast demand of 500 units for a high-storage-cost item, the maximum inventory limit is calculated based on the product of the forecasted weekly sales and the cost ratio, after being constrained by the cost objective, resulting in an actual demand of 300 units. Ultimately, the real demand for single product storage is obtained to ensure storage scheduling capabilities.

[0044] Preferably, step S320 also includes step S321, parsing the semantic label and establishing a basic weight template based on the semantic label; step S322, calling the historical decision error, using the historical decision error to perform feedback adjustment on the basic weight template, and completing the weight configuration of the optimization target according to the feedback adjustment result.

[0045] Preferably, semantic labels such as holiday promotions, new product launches, and negative review risks are obtained through analysis and classified and mapped to optimization target dimensions. Then, an initial weight distribution scheme is established based on the label combination, i.e., a basic weight template is established, as shown in Table 1: Table 1 Semantic tag basic weight data table Semantic Tags Inventory fulfillment weight Cost control weight Turnover efficiency weight Volatility Smoothing Weight Hot-selling items 0.6 0.1 0.2 0.1 Slow-selling warning 0.2 0.5 0.2 0.1 Slow turnover 0.2 0.3 0.4 0.1 new product 0.4 0.2 0.2 0.2 Preferably, historical decision errors are called upon and used to perform feedback adjustments on the basic weight template. Specifically, the deviation between historical forecasted demand and actual consumption is compared and attributed to weight configuration issues. For example, last week's underestimation of the impact of a sudden hit label led to insufficient satisfaction, necessitating an increase in the weight of the latent factor. A reinforcement learning mechanism is then used to adjust the basic weight template, adjusting the weight according to the direction of the error. For example, if historical errors indicate frequent stockouts, the new satisfaction weight is increased by the formula (new satisfaction weight = original weight × (1 + number of stockouts × 0.1). If unsold losses continue to exceed the standard, an exponential decay compensation is applied to the cost weight, such as +5% after each unsold item. Finally, the weight configuration of the optimization target is completed based on the feedback adjustment results. This includes prioritizing high-frequency error scenarios and retaining a minimum weight guarantee for strong semantic labels when label instructions conflict with error feedback. This ensures that reasonable weights are initialized through semantic associations, enabling intelligent warehouse scheduling decisions.

[0046] Furthermore, step S321 also includes step a, activating the scene recognition mechanism, using the scene recognition mechanism to perform scene recognition, and establishing a scene recognition result; step b, configuring a scene optimization template according to the scene recognition result, and using the scene optimization template to perform the basic weight template compensation.

[0047] Preferably, the scene recognition mechanism is used to dynamically perceive the business environment and intelligently classify it. Its core is to determine the current business scene type in real time through multi-source data fusion and pattern matching. When a change in timing signal, a sudden change in business indicators or a change in external market events is detected, the scene recognition mechanism is automatically activated. Specifically, the multi-source signals of the business environment are monitored in real time and analyzed to generate scene feature vectors, such as promotion calendar nodes (Double 11 warm-up period / explosion period / return period), a sudden increase in user purchase conversion rate, changes in search keyword popularity ranking, the promotion strength of competing products in the same category, supply chain warning signals and sudden changes in public opinion sentiment, and then the pre-defined scene types are matched through a pre-trained lightweight random forest classification model to output the scene classification label as the scene recognition result.

[0048] Preferably, multiple templates are called from the preset template library, that is, each scenario type is associated with a set of target weight bias rules, such as conventional scenarios, satisfaction 40%, cost 30%, turnover 20%, and smoothness 10%; big promotion scenarios, satisfaction 60%, cost 15%, turnover 15%, and smoothness 10%; clearance scenarios, satisfaction 20%, cost 10%, turnover 60%, and smoothness 10%; and then the scenario optimization template is configured according to the scenario recognition result, that is, the template parameters are fine-tuned according to real-time strength indicators such as the promotion intensity coefficient and public opinion sentiment score to obtain the scenario optimization template; finally, the basic weight template is compensated with the scenario optimization template, that is, the weight of the scene optimization template is used as the correction amount of the basic template. For example, the original weight satisfaction of the basic template is 50%, and the scenario optimization template requires satisfaction +10%, then the final compensated weight is 55%, thereby realizing intelligent decision-making and adaptive management of warehouse scheduling, improving the accuracy of warehouse demand forecasting, and reducing the error rate of warehouse decision-making.

[0049] In the above, refer to Figure 1 A method for predicting the storage demand of a single product based on multi-source data according to an embodiment of the present invention is described in detail. Figure 3 A single product warehousing demand forecasting system based on multi-source data according to an embodiment of the present invention is described.

[0050] According to an embodiment of the present invention, a single product warehouse demand forecasting system based on multi-source data is used to solve the technical problems existing in the prior art, such as insufficient utilization of multi-source heterogeneous data, lack of explicit and implicit factor coupling analysis, and insufficient dynamic inventory scheduling capabilities. It achieves the technical effects of enhancing multi-objective collaborative optimization capabilities, improving warehouse demand forecasting accuracy, and realizing intelligent adaptive inventory management. Figure 3 As shown, a single product warehousing demand prediction system based on multi-source data includes: a semantic label extraction unit 10, a single product prediction result establishment unit 20, and a single product warehousing real demand establishment unit 30.

[0051] The semantic tag extraction unit 10 is used to extract single product related data from a multi-source database and extract semantic tags using NLP and knowledge graphs, wherein the single product related data includes single product descriptions, user comments, and activity copy; the single product prediction result establishment unit 20 is used to activate the explicit and implicit coupling factor model of warehousing demand, wherein the explicit and implicit coupling factor model includes an explicit and implicit factor extraction layer and an explicit and implicit coupling prediction layer. After extracting explicit factors and implicit factors using the explicit and implicit factor extraction layer, the explicit factors, the implicit factors, and the semantic tags are synchronized to the explicit and implicit coupling prediction layer to establish single product prediction results; the single product warehousing real demand establishment unit 30 is used to read real-time warehousing inventory data, perform a balance analysis of multi-objective scheduling based on the real-time warehousing inventory data and the single product prediction results, and use the balance analysis results to establish single product warehousing real demand.

[0052] The specific configuration of the single product prediction result establishment unit 20 will be described in detail below. The single product prediction result establishment unit 20 further includes: when any data is input into the explicit and implicit factor extraction layer, the data is classified into explicit and implicit categories using a multidimensional discriminator, where the discriminant dimensions of the multidimensional discriminator include data structure, data source, data observability, and data verifiability; data identified as explicit categories are normalized and time-windowed, and encoded into fixed-length vectors; data identified as implicit categories are processed through nonlinear mapping and behavioral sparse smoothing to establish a latent factor vector; explicit factors are established based on the fixed-length vectors, and latent factors are established based on the latent factor vectors.

[0053] The specific configuration of the single product prediction result establishment unit 20 will be described in detail below. The single product prediction result establishment unit 20 further includes: after the explicit factors and the implicit factors are synchronized to the explicit-implicit coupling prediction layer, factor attention analysis is performed based on the factor attention mechanism under the three-dimensional weight features to establish factor weights. The three-dimensional weight features include time weight, scene weight, and label weight, and the label weight is constructed using the semantic label; after performing weighted mapping of the explicit factors and the implicit factors according to the factor weights, the explicit factor weighted result, the implicit factor weighted result, and the semantic label are input into the prediction channel of the explicit-implicit coupling prediction layer to perform correction prediction and establish the single product prediction result.

[0054] The specific configuration of the single product prediction result establishment unit 20 will be described in detail below. The single product prediction result establishment unit 20 further includes: after parsing the semantic labels, initializing the explicit pathway through contextual enhancement, and adjusting the implicit pathway through semantic perception, wherein the explicit pathway and the implicit pathway are the calculation pathways of the prediction channel; inputting the weighted results of the explicit factors into the initialized explicit pathway to establish the main signal output; inputting the weighted results of the implicit factors into the adjusted implicit pathway to establish a dynamic correction factor; and compensating the main signal output according to the dynamic correction factor to establish the single product prediction result.

[0055] The specific configuration of the unit 30 for establishing the actual demand for single-product warehousing will be described in detail below. The unit 30 further includes: performing structural transformation on the real-time warehouse inventory data to generate inventory status data, including current inventory, available inventory, safety inventory, unit storage cost, and turnover rate; configuring optimization objectives for multi-objective scheduling, including inventory satisfaction, cost control, turnover efficiency, and fluctuation smoothing; and utilizing the optimization objectives to perform a balance analysis based on the inventory status data and the single-product forecast results to establish the actual demand for single-product warehousing.

[0056] The specific configuration of the unit 30 for establishing the actual demand for single-product warehousing will be described in detail below. The unit 30 further includes: parsing the semantic tags and establishing a basic weight template based on the semantic tags; invoking historical decision errors, using the historical decision errors to perform feedback adjustments on the basic weight template, and completing the weight configuration of the optimization objective based on the feedback adjustment results.

[0057] The specific configuration of the unit 30 for establishing the actual demand for single-product warehousing will be described in detail below. The unit 30 further includes: activating a scene recognition mechanism, utilizing the scene recognition mechanism to perform scene recognition, and establishing a scene recognition result; configuring a scene optimization template based on the scene recognition result, and utilizing the scene optimization template to perform basic weight template compensation.

[0058] A single product warehousing demand forecasting system based on multi-source data provided by an embodiment of the present invention can execute a single product warehousing demand forecasting method based on multi-source data provided by any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.

[0059] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules may be used and run on the user terminal and / or server, and the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other and are not used to limit the scope of protection of the present invention.

[0060] The above specific embodiments do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the scope of protection of this application.

Claims

1. A method for predicting single product warehousing demand based on multi-source data, characterized in that: The method comprises: Extract product-related data from multi-source databases and extract semantic tags using NLP and knowledge graphs. The product-related data includes product descriptions, user reviews, and campaign copy. Activate the explicit and implicit coupling factor model of warehousing demand, which includes an explicit and implicit factor extraction layer and an explicit and implicit coupling prediction layer. After extracting explicit factors and implicit factors using the explicit and implicit factor extraction layer, synchronize the explicit factors, the implicit factors, and the semantic labels to the explicit and implicit coupling prediction layer to establish a single product prediction result. Read real-time warehouse inventory data, perform a balance analysis of multi-objective scheduling based on the real-time warehouse inventory data and the single product forecast results, and use the balance analysis results to establish the actual demand for single product warehousing.

2. The method for predicting single product warehousing demand based on multi-source data according to claim 1, characterized in that: The extracting of dominant factors and recessive factors by using the dominant and recessive factor extraction layer includes: When any data is input into the explicit and implicit factor extraction layer, the data is classified into explicit and implicit categories by a multidimensional discriminator, and the discriminant dimensions of the multidimensional discriminator include data structure, data source, data observability, and data verifiability; The data identified as dominant categories are normalized and time-windowed, and encoded into fixed-length vectors; The data identified as latent classification is processed through nonlinear mapping and behavioral sparse smoothing to establish a latent factor vector; An explicit factor is established according to the fixed-length vector, and a latent factor is established according to the latent factor vector.

3. The method for predicting single product warehousing demand based on multi-source data according to claim 2, characterized in that: The step of synchronizing the explicit factors, the implicit factors, and the semantic labels to the explicit-implicit coupling prediction layer to establish a single product prediction result includes: After the explicit factors and the implicit factors are synchronized to the explicit-implicit coupling prediction layer, factor attention analysis is performed under three-dimensional weight features based on the factor attention mechanism to establish factor weights. The three-dimensional weight features include time weight, scene weight, and label weight. The label weight is constructed using the semantic label. After performing weighted mapping of explicit factors and implicit factors according to the factor weights, the weighted results of the explicit factors, the weighted results of the implicit factors and the semantic labels are input into the prediction channel of the explicit-implicit coupling prediction layer to perform correction prediction and establish the single product prediction result.

4. The method for predicting single product warehousing demand based on multi-source data according to claim 3, characterized in that: The step of inputting the weighted result of the explicit factors, the weighted result of the implicit factors, and the semantic label into the prediction channel of the explicit-implicit coupling prediction layer to perform correction prediction includes: After parsing the semantic labels, the explicit pathway is initialized through context enhancement, and the implicit pathway is adjusted through semantic perception, wherein the explicit pathway and the implicit pathway are calculation pathways of the prediction channel; Inputting the weighted result of the dominant factor into the initialized dominant path to establish a main signal output; Inputting the weighted result of the latent factor into the adjusted latent pathway to establish a dynamic correction factor; After compensating the main signal output according to the dynamic correction factor, the single product prediction result is established.

5. The method for predicting single product warehousing demand based on multi-source data according to claim 1, characterized in that: The reading of real-time warehouse inventory data, performing a balance analysis of multi-objective scheduling based on the real-time warehouse inventory data and the single product forecast results, and establishing the actual single product warehousing demand using the balance analysis results include: Performing structural transformation on the real-time warehouse inventory data to generate inventory status data, wherein the inventory status data includes current inventory, available inventory, safety inventory, unit storage cost, and turnover speed; Configure the optimization objectives of multi-objective scheduling, including inventory satisfaction objectives, cost control objectives, turnover efficiency objectives, and fluctuation smoothing objectives; The optimization goal is used to perform a balance analysis based on the inventory status data and the single product forecast results to establish the actual demand for single product storage.

6. The method for predicting single product warehousing demand based on multi-source data according to claim 5, characterized in that: The optimization objectives of configuring multi-objective scheduling include: Parsing the semantic tags, and establishing a basic weight template according to the semantic tags; The historical decision error is called, and the historical decision error is used to perform feedback adjustment of the basic weight template, and the weight configuration of the optimization target is completed according to the feedback adjustment result.

7. The method for predicting single product warehousing demand based on multi-source data according to claim 6, characterized in that: The step of establishing a basic weight template according to the semantic tag includes: activating a scene recognition mechanism, performing scene recognition using the scene recognition mechanism, and establishing a scene recognition result; A scene optimization template is configured according to the scene recognition result, and the basic weight template is compensated using the scene optimization template.

8. A single product warehousing demand forecasting system based on multi-source data, characterized by: The system is used to implement the method for predicting single product warehousing demand based on multi-source data according to any one of claims 1 to 7, and the system includes: A semantic tag extraction unit is used to extract product-related data from a multi-source database and extract semantic tags using NLP and knowledge graphs, wherein the product-related data includes product descriptions, user reviews, and campaign copy; A single product prediction result establishment unit is used to activate the explicit and implicit coupling factor model of warehousing demand, the explicit and implicit coupling factor model including an explicit and implicit factor extraction layer and an explicit and implicit coupling prediction layer. After extracting explicit factors and implicit factors using the explicit and implicit factor extraction layer, the explicit factors, the implicit factors, and the semantic labels are synchronized to the explicit and implicit coupling prediction layer to establish a single product prediction result. The unit for establishing the actual demand for single product storage is used to read real-time storage inventory data, perform a balance analysis of multi-objective scheduling based on the real-time storage inventory data and the single product prediction results, and establish the actual demand for single product storage using the balance analysis results.

Images