A method, device, equipment and medium for commodity order forecasting based on large model

Abstract

The present application relates to a commodity order prediction method, apparatus, equipment and medium based on a large model, including determining target sales feature information from the sales guidance document of the commodity based on a pre-trained industry dictionary of the industry in which the commodity is located; obtaining special time description information, promotion time information of the ordering store, store address information and promotion terms information, and determining spatiotemporal correlation information based on the special time description information, promotion time information, store address information and promotion terms information; an acquisition module obtains weather warning information and historical store sales information; according to the target sales feature information, spatiotemporal correlation information, weather warning information and historical store sales information, dynamically adjusts weights through an attention weight allocation mechanism to determine the predicted commodity order information of the store; and visually displays the target display information through the risk warning module of the large model when the predicted commodity order information meets the preset conditions.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method, device, equipment and medium for commodity order forecasting based on a large model. Background Art

[0002] With the continuous development of the retail industry and the advancement of information technology, the sales and ordering models for fast-moving consumer goods (FMCG) are becoming increasingly complex. Traditionally, order quantities often rely on store managers' experience or simple historical statistical data, which can, to a certain extent, meet the needs of small, stable markets. However, with the continuous increase in the number of stores and product types, consumer preferences are also changing dynamically. Relying solely on experience or simple time series order forecasts can no longer accurately grasp order demand, often leading to inventory backlogs and stock-outs, which seriously affect sales efficiency and customer satisfaction.

[0003] Currently, order quantity forecasting methods based on statistics or traditional machine learning models, such as linear regression, ARIMA, and XGBoost, can be used. These methods can indeed improve forecasting performance when the data size is relatively limited or the scenario is relatively simple. However, when faced with multidimensional data in actual operations (such as historical order data, sales data, inventory data, and the brand, location, and industry attributes of each store), it is difficult to fully explore the correlations between complex features. Furthermore, during special periods such as holidays and promotional events, forecast results are prone to significant deviations, making it difficult to meet the higher requirements for real-time performance and accuracy.

[0004] The feasibility of using multi-layer neural networks and attention mechanisms (such as the Transformer architecture) for feature extraction and prediction. Large models offer significant advantages in processing high-dimensional, multimodal data. Furthermore, they can automatically learn underlying time series patterns and correlation features from massive amounts of historical data, providing more accurate and flexible forecasting capabilities for ordering behavior analysis. However, when applied to ordering scenarios, challenges remain: effectively utilizing multi-dimensional information from different stores, applying model results to actual business decisions, and updating forecasts in real time. Summary of the Invention

[0005] The purpose of the present invention is to provide a commodity order prediction method, device, equipment and medium based on a large model, aiming to solve the problems of how to effectively utilize the multi-dimensional information of different stores, how to apply the model results to actual business decisions, and how to update the prediction in real time when multi-layer neural networks and attention mechanisms are applied to commodity ordering scenarios.

[0006] To achieve the above objectives, a first aspect of an embodiment of the present disclosure provides a commodity order forecasting method based on a large model, the method comprising:

[0007] Through the policy analysis module in the large model, the target selling feature information is determined from the sales guidance document of the product according to the pre-trained industry dictionary of the industry in which the product is located; through the spatiotemporal association module in the large model, special time description information, promotion time information of the ordering store, store address information and promotion terms information are obtained, and the spatiotemporal association information of the store is determined based on the special time description information, the promotion time information, the store address information and the promotion terms information; through the multidimensional data acquisition module in the large model, weather warning information and store historical sales information are obtained; through the sales prediction module of the large model, according to the target selling feature information, the spatiotemporal association information, the weather warning information and the store historical sales information, the weight is dynamically adjusted through the attention weight allocation mechanism to determine the predicted product ordering information of the store; through the risk warning module of the large model, when the predicted product ordering information meets the preset conditions, the target display information is visualized.

[0008] In one possible implementation, the policy analysis module in the large model determines target selling feature information from the sales guidance document of the product based on a pre-trained industry dictionary for the industry in which the product is located, including:

[0009] Through the policy analysis module in the large model, the corresponding sales feature information is obtained according to the pre-trained industry dictionary of the industry in which the product is located, and the bidirectional recurrent neural network in the two-layer semantic parsing engine is used to determine the implicit restrictions from the sales feature information to obtain the actual sales feature information of the product in the sales guidance; the relationship extraction layer in the two-layer semantic parsing engine is used to establish the logical dependency relationship between policy clauses to obtain the sales constraint chain; based on the actual sales feature information and the actual sales feature information, the target sales feature information is determined.

[0010] In a possible implementation, the method further includes:

[0011] In response to obtaining a new sales guidance document, sensitive sales parameters are determined from the original sales guidance document; by protecting sensitive neurons, the adjustment range of the sensitive sales parameters is limited during incremental training to prevent the new sales guidance document from overwriting the sensitive sales parameters; according to the new sales guidance document, the policy analysis module is incrementally trained to obtain a new policy analysis module.

[0012] In a possible implementation, determining the spatiotemporal association information of the store based on the special time description information, the promotion time information, the store address information, and the promotion terms information includes:

[0013] The time description in the promotion time information is converted into a calendar, and the text semantics of the time description is bound to the date range to determine the promotion time window; during vector retrieval, the similarity between the special time description information and the promotion time window is synchronously constrained according to the time validity corresponding to the promotion time window, and the target promotion terms information is determined from the promotion terms information; according to the geographic location encoder, the store address information is vector-converted, and the store address is converted into a business district feature vector, and the business district feature vector includes at least one of the following: school density, transportation hub distance; according to the promotion time window, the target promotion terms information and the business district feature vector, they are associated to determine the spatiotemporal association information of the store.

[0014] In a possible implementation, the method further includes:

[0015] Obtain the actual product ordering information of the store for the predicted product ordering information, and determine whether the triggering update condition is met based on the actual product ordering information and the predicted product ordering information; if the triggering update condition is met, analyze the update policy information of the sales guidance document and / or monitor the sudden change information of the flow of people in the business district where the store address is located through text similarity; update the large model according to the updated policy information and / or the sudden change information of the flow of people to obtain a new large model.

[0016] In a possible implementation, the visual display of the target display information includes:

[0017] In a case where the predicted commodity ordering information output by the large model represents a suggestion to reduce the commodity order quantity of the store, the contribution of at least one of the target sales feature information, the spatiotemporal correlation information, the weather warning information and the store's historical sales information is displayed; the policy items that trigger the suggestion to reduce the commodity order quantity of the store are marked and displayed; the text segment that the large model focuses on is highlighted, and the target verification information corresponding to the text segment is determined from the target sales feature information, the spatiotemporal correlation information, the weather warning information and the store's historical sales information; the consistency index between the attention weight corresponding to the target verification information and the manual expert judgment is calculated and displayed.

[0018] In a possible implementation, the method further includes:

[0019] Obtain abnormal sales information from the store's historical sales information that represents the actual sales information of the store and meets preset elimination conditions, and eliminate the store's historical sales information corresponding to the abnormal sales information; determine whether there is new target information in the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information, and judge the deviation of the new target from the historical distribution through adversarial training, and if the deviation does not meet the preset deviation threshold, eliminate the new target information, and if the new target information represents the existence of the new target sales feature information, mark and isolate the target sales feature information before and after the addition; automatically check whether the predicted commodity ordering information complies with the preset industry rules, and if the predicted commodity ordering information does not comply with the preset industry rules, re-execute the prediction of the predicted commodity ordering information.

[0020] According to a second aspect of the embodiments of the present disclosure, there is provided a large-scale model-based commodity order forecasting device, the device comprising:

[0021] a sales information determination module configured to determine target sales feature information from the sales guidance document of the product using the policy analysis module in the large model and based on the industry dictionary pre-trained for the industry in which the product belongs;

[0022] a spatiotemporal association module configured to obtain, through the spatiotemporal association module in the large model, special time description information, promotion time information, store address information, and promotion terms information of the ordering store, and determine the spatiotemporal association information of the store based on the special time description information, the promotion time information, the store address information, and the promotion terms information;

[0023] An information acquisition module configured to acquire weather warning information and store historical sales information through the multidimensional data acquisition module in the large model;

[0024] a prediction module configured to determine predicted product order information for the store by dynamically adjusting weights using an attention weight allocation mechanism based on the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information using the sales prediction module of the large model;

[0025] The visualization display module is configured to visualize the target display information when the predicted commodity order information meets the preset conditions through the risk warning module of the large model.

[0026] According to a third aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the steps of any one of the methods described in the first aspect are implemented.

[0027] According to a fourth aspect of the present disclosure, an electronic device is provided, including:

[0028] A memory having a computer program stored thereon; a processor configured to execute the computer program in the memory to implement the steps of any one of the methods described in the first aspect.

[0029] The present invention provides a method, device, equipment, and medium for commodity order forecasting based on a large model. Compared with the prior art, it has the following advantages:

[0030] The policy analysis module and industry dictionary enable precise extraction of target sales characteristics from product sales guidance documents, ensuring analysis is based on accurate and targeted industry data, thereby improving the accuracy of subsequent analysis. The spatiotemporal correlation module integrates multi-dimensional information such as special time descriptions, promotional dates, store locations, and promotional terms to comprehensively capture store operating characteristics in different spatiotemporal contexts. This provides richer context for sales forecasts and enhances the spatiotemporal adaptability of forecasts.

[0031] The multidimensional data acquisition module captures weather warning information and historical store sales data. Combining external weather data with internal sales data provides more comprehensive data support for sales forecasts, helping to more accurately assess the various factors influencing sales. Using an attention weighting mechanism, the module dynamically adjusts the weights of various factors based on target sales characteristics, spatiotemporal correlations, weather warnings, and historical store sales data, enabling more flexible and accurate sales forecasts and effectively responding to market changes.

[0032] When the predicted product order information meets the preset conditions, the visual display is automatically triggered to provide users with risk warning information in a timely manner, helping users make decisions quickly, reduce potential losses, and improve the efficiency and intuitiveness of risk management.

[0033] In this way, through the full processing of multi-dimensional data, when the model is applied to the commodity ordering scenario, by integrating multi-source data, dynamically adjusting the prediction weight and providing a visual display of risk warnings, the accuracy and timeliness of commodity ordering forecasts are significantly improved. The multi-dimensional information of different stores is effectively utilized, the model results are applied to actual business decisions and the forecasts are updated in real time, which improves the forecast accuracy of commodity ordering and meets the differentiated needs of the market and stores.

[0034] Other features and advantages of the present disclosure will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings are used to provide a further understanding of the present disclosure and constitute a part of the specification. Together with the following detailed description, they are used to explain the present disclosure but do not constitute a limitation of the present disclosure. In the accompanying drawings:

[0036] Figure 1 The present invention is a flowchart showing a method for commodity order forecasting based on a large model according to an embodiment of the specification.

[0037] Figure 2 A block diagram of a commodity order forecasting device based on a large model is shown according to an embodiment of the specification.

[0038] Figure 3 It is a block diagram of another large model-based commodity order forecasting device according to an embodiment of the specification. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0040] This application provides a commodity order forecasting method based on a large model. Figure 1 This is a flowchart illustrating a method for commodity order forecasting based on a large model according to an embodiment. The method includes:

[0041] In step S11, the policy analysis module in the large model determines target selling feature information from the sales guidance document of the product according to the pre-trained industry dictionary of the industry in which the product is located;

[0042] The policy analysis module is a submodule within the larger model, specifically designed to parse and understand policy documents related to product sales. The industry dictionary is a pre-trained vocabulary set for a specific industry, encompassing terminology, regulations, and expressions unique to that industry. Target sales feature information is extracted from sales guidance documents and contains key information directly related to product sales, such as sales restrictions, promotion policies, and age restrictions.

[0043] In this disclosed embodiment, the policy analysis module uses pre-trained industry dictionaries to perform text parsing and semantic understanding of product sales guidance documents, identifying and extracting feature information directly related to product sales. This information forms the basis for subsequent sales forecasts and risk warnings.

[0044] In step S12, the spatiotemporal association module in the large model obtains special time description information, promotion time information of the ordering store, store address information, and promotion terms information, and determines the spatiotemporal association information of the store based on the special time description information, the promotion time information, the store address information, and the promotion terms information;

[0045] The spatiotemporal correlation module is a submodule within the larger model, used to integrate and analyze information related to time and space. Special time description information is information with temporal characteristics, such as holidays and special events. Promotional time information is the time of promotional activities set by the store. Store address information is the store's specific geographic location. Promotion terms information is the specific terms and conditions associated with the promotion. Spatiotemporal correlation information is information related to store sales, derived by comprehensively considering time, space, and promotional factors.

[0046] In the disclosed embodiment, the spatiotemporal correlation module obtains and integrates special time description information, promotion time information, store address information, and promotion terms information, and analyzes the correlation and influence between these information to determine the spatiotemporal correlation information of the store. This information helps to more accurately predict the store's sales under different time and space conditions.

[0047] In step S13, weather warning information and store historical sales information are obtained through the multidimensional data acquisition module in the large model;

[0048] The multidimensional data acquisition module is a submodule within the larger model, responsible for acquiring data related to product sales from multiple data sources. Weather warning information refers to weather-related warnings, such as heavy rain or high temperatures. Store historical sales information refers to a store's sales data over a period of time.

[0049] In this disclosed embodiment, the multidimensional data acquisition module connects to external data sources (such as the Meteorological Bureau and store sales systems) to obtain weather warning information and historical store sales data. This information provides more comprehensive data support for sales forecasts and helps to more accurately assess the various factors affecting sales.

[0050] In step S14, the sales forecast module of the large model dynamically adjusts the weights based on the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information through an attention weight allocation mechanism to determine the predicted product order information for the store;

[0051] The sales forecast module is a submodule within the larger model, used to predict product sales based on multiple factors. The attention weighting mechanism dynamically adjusts the weights of various factors, assigning different weights based on their impact on sales. Predicted product order information is derived from sales forecast results, recommending the quantity of goods that stores should order.

[0052] In the disclosed embodiment, the sales forecasting module uses an attention weighting mechanism to dynamically adjust the weights of various factors based on target sales characteristics, spatiotemporal correlation information, weather warning information, and historical store sales data, thereby determining predicted product order information for a store. This mechanism enables more flexible response to market changes and improves the accuracy of sales forecasts.

[0053] The above technical solution integrates multi-dimensional evidence reasoning, breaks through the traditional single data reasoning model, and constructs a three-layer evidence chain system: structured evidence: directly calling the store historical data in the database; policy evidence: real-time retrieval of the latest industry standards; environmental evidence: integration of external data (such as the assessment of the impact of meteorological warning information on logistics); through the attention weight distribution mechanism, dynamically adjust the influence ratio of the three types of evidence on the final decision.

[0054] In step S15, the risk warning module of the large model visualizes the target display information when the predicted commodity order information meets the preset conditions.

[0055] The risk warning module is a submodule within the larger model, responsible for issuing risk warnings when prediction results meet preset conditions. Preset conditions are pre-set thresholds or rules used to determine whether a risk warning is necessary. The target display information is the risk warning information that needs to be presented to users. Visualization presents risk warning information to users in the form of graphs, charts, and other visual formats.

[0056] In the disclosed embodiment, the differences between actual order execution data and model predictions are collected and annotated, especially the "reasonable order quantity" after manual correction, and regularly fed back into incremental fine-tuning training; a strategy replay mechanism for historical typical business scenarios (such as promotional activities, seasonal ordering strategies) is clearly set to protect core business knowledge in a weighted manner; special sampling weights are set for the learning of sensitive clauses in policies and regulations to avoid forgetting regulatory information and ensure the continued effectiveness of compliance knowledge.

[0057] The dynamically adaptive fine-tuning mechanism in the disclosed embodiment divides the large model into a policy understanding module, a sales forecast module, and a risk warning module, and dynamically activates modules related to the input data scenario through a gating mechanism; when the latest policy information is detected, the policy understanding module is preferentially called for local fine-tuning; in conventional scenarios, only the basic sales forecast module is activated to reduce resource consumption; special protection mechanisms are set up especially for regulatory-sensitive neurons (such as parameters related to monopoly licenses) to avoid overwriting or forgetting of regulatory-sensitive knowledge during incremental model training.

[0058] In the disclosed embodiment, the risk warning module visualizes the target display information when the predicted product order information meets preset conditions (such as predicted sales volume significantly lower or higher than the historical average). This display method helps users quickly understand the risk situation and make appropriate decisions.

[0059] In view of the characteristics of the tobacco industry, the parameter adjustment system with business scenario awareness divides the large model into independent functional modules such as policy understanding, sales forecast, and risk warning through modular fine-tuning components, and dynamically activates relevant modules through a gating mechanism; when it is detected that the input data involves the latest policy, the policy analysis module is called first; the basic forecast module is enabled during routine order forecasting; compliance protection mechanism: identifies neurons in the model parameters that are sensitive to regulations (such as decision nodes involving monopoly license verification), and limits the adjustment range of these parameters during incremental training to prevent new knowledge from overwriting key regulatory memories.

[0060] This technical solution leverages a policy analysis module and industry dictionary to accurately extract target sales characteristics from product sales guidance documents, ensuring that analysis is based on accurate and targeted industry data, thereby improving the accuracy of subsequent analysis. The spatiotemporal correlation module integrates multi-dimensional information such as special time descriptions, promotional dates, store locations, and promotional terms to comprehensively capture store operations in different spatiotemporal contexts. This provides richer context for sales forecasts and enhances the spatiotemporal adaptability of forecasts.

[0061] The multidimensional data acquisition module captures weather warning information and historical store sales data. Combining external weather data with internal sales data provides more comprehensive data support for sales forecasts, helping to more accurately assess the various factors influencing sales. Using an attention weighting mechanism, the module dynamically adjusts the weights of various factors based on target sales characteristics, spatiotemporal correlations, weather warnings, and historical store sales data, enabling more flexible and accurate sales forecasts and effectively responding to market changes.

[0062] When the predicted product order information meets the preset conditions, the visual display is automatically triggered to provide users with risk warning information in a timely manner, helping users make decisions quickly, reduce potential losses, and improve the efficiency and intuitiveness of risk management.

[0063] In this way, through the full processing of multi-dimensional data, when the model is applied to the commodity ordering scenario, by integrating multi-source data, dynamically adjusting the prediction weight and providing a visual display of risk warnings, the accuracy and timeliness of commodity ordering forecasts are significantly improved. The multi-dimensional information of different stores is effectively utilized, the model results are applied to actual business decisions and the forecasts are updated in real time, which improves the forecast accuracy of commodity ordering and meets the differentiated needs of the market and stores.

[0064] In one possible implementation, in step S11, the policy analysis module in the large model determines target selling feature information from the sales guidance document of the product based on a pre-trained industry dictionary for the industry in which the product is located, including:

[0065] In step S111, the policy analysis module in the large model obtains corresponding sales feature information based on the pre-trained industry dictionary for the industry in which the product is located, and uses the bidirectional recurrent neural network in the two-layer semantic parsing engine to determine implicit restrictions from the sales feature information to obtain the actual sales feature information of the product in the sales guide;

[0066] The two-layer semantic parsing engine includes a semantic parsing tool with two layers of processing logic. The first layer is responsible for basic semantic parsing, while the second layer performs deep semantic understanding and relationship extraction. The bidirectional recurrent neural network (BiRNN) is a neural network capable of processing sequential data. It improves parsing accuracy by considering the context of the sequence. Sales feature information is extracted from sales guidance documents and contains descriptions of features directly related to product sales, such as price, promotional methods, and sales targets. Implicit restrictions are restrictions that are not explicitly stated in the sales guidance documents but can be inferred through semantic parsing, such as sales time restrictions and sales area restrictions. Actual sales feature information is a complete and accurate description of the product's features in the sales guidance, obtained by combining the sales feature information and implicit restrictions.

[0067] In this disclosed embodiment, the policy analysis module first retrieves the corresponding sales feature information from the sales guidance document based on a pre-trained industry dictionary for the product's industry. Then, using a bidirectional recurrent neural network (BiRNN) within a two-layer semantic parsing engine, it performs deep semantic analysis on this sales feature information, identifying and extracting implicit restrictions. By combining this sales feature information with the implicit restrictions, the actual sales feature information for the product in the sales guidance is obtained.

[0068] In step S112, the relationship extraction layer in the two-layer semantic parsing engine is used to establish logical dependencies between policy clauses to obtain a sales constraint chain;

[0069] The relationship extraction layer, the second layer of the two-layer semantic parsing engine, is responsible for extracting relationships between entities from the text and establishing logical dependencies. Policy clauses: Specific provisions or requirements within sales guidance documents. Logical dependencies: The interdependencies or influences between policy clauses, such as the implementation of one clause may depend on the fulfillment of another. Sales constraint chains: A chain consisting of multiple policy clauses and their logical dependencies, used to describe various constraints and restrictions in the product sales process.

[0070] In this disclosed embodiment, the relationship extraction layer of a two-layer semantic parsing engine is used to analyze each policy clause in the sales guidance document, identifying and establishing logical dependencies between them. By connecting these relationships in series, a sales constraint chain is formed to comprehensively and systematically describe the various constraints and restrictions in the product sales process.

[0071] In step S113, the target selling characteristic information is determined according to the actual selling characteristic information and the actual selling characteristic information.

[0072] In this embodiment, the actual sales characteristics information obtained in step S111 and the sales constraint chain obtained in step S112 are further integrated and analyzed to determine target sales characteristics information that has a key impact on product sales forecasting and risk warning. This information will serve as important input for subsequent steps to improve the accuracy of sales forecasting and the effectiveness of risk warnings.

[0073] In the above solution, the terminology recognition layer uses pre-trained industry dictionaries to identify core concepts in policy documents (such as "exclusive license validity period" and "order quota"), and employs a bidirectional recurrent neural network to capture implicit restrictions in the context (such as the actual implementation standards of vague expressions such as "in principle"). The relationship extraction layer establishes logical dependencies between policy clauses. For example, when Policy A stipulates "no sales within 100 meters of a school," it is automatically linked to the "campus perimeter definition standards" in Policy B, forming a complete constraint chain.

[0074] In a possible implementation, the method further includes: in response to acquiring a new sales guidance document, determining sensitive sales parameters from an original sales guidance document;

[0075] New sales guidance documents: These are newly issued regulations or policy documents by the government or relevant departments that have a direct impact on product sales. Existing sales guidance documents: These are the collection of existing sales guidance documents upon which the current policy analysis module is based. Sensitive sales guidance parameters: These are key parameters or clauses within sales guidance that have a significant impact on product sales and require special protection to prevent arbitrary changes, such as age restrictions and sales bans in specific areas.

[0076] In the disclosed embodiment, when a new sales guidance document is obtained, it will first be compared and analyzed with the original sales guidance document. Through natural language processing technology and policy clause matching algorithm, the clauses in the original policy document that may change or require special attention in the new policy document are identified. The parameters corresponding to these clauses are sensitive sales guidance parameters. The purpose of determining these parameters is to ensure that these key parameters are not overwritten or changed by improper adjustments to the new policy document during the subsequent incremental training process.

[0077] By protecting sensitive neurons, the adjustment range of the sensitive sales parameters is limited during incremental training to prevent new sales guidance documents from overwriting the sensitive sales parameters;

[0078] Sensitive neuron protection: In deep learning models (such as the neural network in the policy analysis module), specific techniques (such as weight locking and regularization) are used to protect neurons that process sensitive sales guidance parameters, preventing them from being over-adjusted during training. Incremental training: Based on the existing model, training is performed using new data to update model parameters and improve the model's adaptability to new data. Adjustment amplitude: The degree of change in model parameters (including neuron weights) during training.

[0079] In the embodiment of the present disclosure, during incremental training, in order to protect sensitive sales guidance parameters from being overwritten by improper adjustments to the new policy document, a technology to protect sensitive neurons will be used. Specifically, the neurons that process these sensitive parameters will be specially marked or processed, such as locking their weights, applying larger regularization terms, etc., to limit their adjustment range during the training process. In this way, even if the new sales guidance document adjusts certain terms, it can ensure that the sensitive sales guidance parameters remain relatively stable and are not completely overwritten or changed.

[0080] In the policy analysis module's neural network, neurons processing sensitive sales guidance parameters are marked as sensitive. During incremental training, these neurons' weights are locked or regularized to limit their adjustment range. This ensures that even if new sales guidance documents slightly adjust age limits, the existing age limits are not completely overwritten, maintaining policy continuity and stability.

[0081] According to the new sales guidance document, the policy analysis module is incrementally trained to obtain a new policy analysis module.

[0082] In the disclosed embodiment, after determining the sensitive sales guidance parameters and protecting the sensitive neurons, the policy analysis module is incrementally trained based on the new sales guidance document. During the training process, the new policy document data is used to update the model parameters, but at the same time, the sensitive sales guidance parameters are ensured not to be over-adjusted. Through incremental training, the policy analysis module can learn new policy provisions and clauses while maintaining stable processing of the original sensitive parameters, thereby obtaining a new policy analysis module that is more adapted to the current sales guidance environment.

[0083] In a possible implementation, in step S12, determining the spatiotemporal association information of the store based on the special time description information, the promotion time information, the store address information, and the promotion terms information includes:

[0084] In step S121, the time description in the promotion time information is converted into a calendar, the text semantics of the time description is bound to the date range, and the promotion time window is determined;

[0085] Promotional time information: Information describing the specific schedule of product promotions, which may exist in text format, such as "during National Day" or "every Friday afternoon." Calendarization: The process of converting textual time descriptions into specific date ranges or time windows. Promotional time windows: The specific time range obtained after calendarization, used to indicate the specific execution period of promotional activities.

[0086] In the disclosed embodiment, when processing promotional time information, the textual description of the time must first be parsed and understood. Natural language processing techniques are used to identify key elements in the time description, such as holiday names, days of the week, and time periods. These elements are then matched and converted to calendar data to obtain a specific date range or time window, i.e., the promotional time window.

[0087] In step S122, during vector retrieval, based on the time validity corresponding to the promotion time window, the similarity between the special time description information and the promotion time window is synchronously constrained, and target promotion term information is determined from the promotion term information;

[0088] Special time description information: In addition to promotional time information, other time-related description information, such as holidays and special events, may have an impact on sales. Vector retrieval: Converts text information into vector form to facilitate similarity calculation and retrieval in vector space. Time validity: Describes whether specific time description information is valid or influential within a specific time window. Target promotion terms: Filters promotion terms from promotion terms information to identify those most relevant to the current time window and most likely to impact sales, taking time validity into account.

[0089] In this disclosed embodiment, vector retrieval not only considers the promotion time window itself but also combines it with special time description information to assess its temporal validity within the promotion time window. By calculating the similarity between the special time description information and the promotion time window, and considering temporal validity as a constraint, it is possible to more accurately filter out the promotion terms most likely to have an impact within the current time window, namely, the target promotion terms.

[0090] In step S123, the store address information is vector-converted according to the geographic location encoder, and the store address is converted into a business district feature vector, where the business district feature vector includes at least one of the following: school density and distance to a transportation hub;

[0091] Among them, a geocoder is a tool or model that converts geographic location information (such as an address) into a vector representation for processing and analysis in vector space. A business district feature vector is a multidimensional vector that describes the characteristics of the business district where a store is located. It may include multiple dimensions such as school density, distance to transportation hubs, population density, and the abundance of commercial facilities.

[0092] In the disclosed embodiment, a geolocation encoder is used to convert a store's address information into a vector form, namely a business district feature vector. During this process, the geolocation encoder considers multiple aspects of the address, such as geographic coordinates and surrounding environmental characteristics, and maps them into a multidimensional vector. Each dimension in the business district feature vector represents a specific business district characteristic, such as school density and proximity to transportation hubs. These characteristics help to more comprehensively describe the business district environment in which the store is located.

[0093] In step S124, the promotion time window, the target promotion term information, and the business district feature vector are associated to determine the spatiotemporal association information of the store.

[0094] In this disclosed embodiment, after obtaining the promotion time window, target promotion terms, and business district feature vectors, this information is correlated and integrated to form spatiotemporal correlation information for each store. This process considers the interactions and influences between time, promotion, and space, thereby providing a more comprehensive description of a store's sales characteristics and potential risks under different temporal and spatial conditions.

[0095] Compared to traditional knowledge bases that independently store text and numerical data, this technical solution uses spatiotemporal vector indexing technology: it converts time descriptions in promotional rules (such as "holiday period") into calendars, binds text semantics to specific date ranges, and simultaneously considers text similarity and time validity during vector retrieval. For example, only promotional terms available within the current time window are returned. A geolocation encoder is established to convert store addresses into business district feature vectors (such as school density, distance to transportation hubs, etc.), achieving intelligent matching in the spatial dimension.

[0096] In one possible implementation, the method further includes: obtaining actual product order information of the store for the predicted product order information, and determining whether a triggering update condition is met based on the actual product order information and the predicted product order information;

[0097] In this disclosed embodiment, actual product order information is first obtained from store orders and then compared with predicted product order information. The difference between the two (e.g., order quantity, order type, etc.) is calculated to determine whether the difference exceeds a preset threshold that triggers an update. If the difference exceeds the threshold, the current large model is deemed to be unable to accurately predict the store's order demand and is updated.

[0098] When the triggering update condition is met, analyzing the updated policy information of the sales guidance document and / or monitoring the sudden change information of the foot traffic in the business district where the store address is located through text similarity analysis;

[0099] Among them, updated policy information: Newly released or revised content in sales guidance documents may affect product sales and ordering. Sudden increases or decreases in foot traffic within the business district where the store is located may be caused by factors such as holidays and special events and have a direct impact on product sales. Text similarity analysis: By calculating the similarity between texts, we can determine the degree of difference between updated policy information and existing policies.

[0100] In the embodiment of the present disclosure, when the triggering update conditions are met, one of the two strategies or a combination of the two will be adopted to obtain update information. On the one hand, the latest sales guidance documents will be analyzed, and the differences between the updated policy information and the existing policies will be identified through text similarity analysis technology. These differences may involve sales restrictions, tax policies, promotional activities, etc. On the other hand, changes in the flow of people in the business district where the store address is located will be monitored, especially sudden changes in the flow of people, which may be caused by holidays, large-scale events, weather changes, etc. By analyzing this information, it is possible to understand the impact of changes in the external environment on product sales.

[0101] The large model is updated according to the update policy information and / or the pedestrian flow mutation information to obtain a new large model.

[0102] In the disclosed embodiments, after obtaining updated policy information and / or information on sudden changes in foot traffic, the large model is updated using this information as new training data or as a basis for adjusting parameters. This update process may involve adjusting model weights, adding new feature dimensions, modifying the model's loss function, and so on. Through updates, the large model can better adapt to changes in the external environment, improve the accuracy of forecasting product order demand, and enhance its ability to analyze information such as sales guidance and foot traffic in shopping districts.

[0103] Compared with traditional model updates that will forget old knowledge, a business strategy memory library is introduced to regularly store typical decision-making scenarios (such as Spring Festival stocking strategies and new store opening support plans). When updating the model, historical strategies are replayed according to business importance to ensure the stability of the core business logic, set the priority protection level of regulatory provisions, and set a sampling weight of 5 times that of conventional strategies for decision-making models involving legal compliance.

[0104] Establish a multi-signal linkage evaluation system to accurately determine the timing of model updates: Business indicator monitoring: trigger an early warning when the inventory deviation rate is greater than 20% and the forecast accuracy is less than 85% for three consecutive days;

[0105] Policy change detection: Identifying major updates to the policy database through text similarity analysis; environmental change perception: Monitoring sudden changes in pedestrian flow in a business district (e.g., changes in customer demographics due to the opening of a new subway station). Incremental learning is initiated when any of these conditions are met.

[0106] In a possible implementation, in step S15, visually displaying the target display information includes:

[0107] In step S151, when the predicted product order information output by the large model indicates a recommendation to reduce the product order quantity of the store, the contribution of at least one of the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information is displayed;

[0108] In the disclosed embodiments, interpretability techniques for machine learning models, such as SHAP (SHapley Additive exPlanations) values ​​and LIME (Local Interpretable Model-agnostic Explanations), are used to calculate the contribution of each input feature to the prediction result. Contribution ranking and display: Based on the calculated contribution, the target sales feature information (such as product price and brand preference), spatiotemporal correlation information (such as business district traffic and holidays), meteorological warning information (such as extreme weather warnings), and store historical sales information (such as sales trends and seasonal fluctuations) are ranked, and the factors with the highest contribution are selected for visual display.

[0109] In step S152 , the policy item that triggers the recommendation to reduce the store's merchandise order quantity is marked and displayed;

[0110] In this disclosed embodiment, text mining is performed on sales guidance documents to extract policy items related to product sales, order quantities, and other factors. Policy and Forecast Correlation: The extracted policy items are analyzed to determine which ones are directly related to the current forecast recommendation to reduce order quantities. Labeling and Display: These relevant policy items are labeled in a visual interface, and specific policy content or links may be included for further user understanding.

[0111] In step S153, the text segment focused on by the large model is highlighted, and target verification information corresponding to the text segment is determined from the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information;

[0112] In the disclosed embodiment, the text segments that the large model pays special attention to during the prediction process are identified, and the target verification information corresponding to these segments is determined from the relevant information. This is usually achieved by:

[0113] Text fragment identification: Leveraging natural language processing techniques such as keyword extraction and named entity recognition, we identify text fragments that the large model specifically focuses on during prediction. Target verification information determination: Based on the identified text fragments, we determine relevant target verification information from target sales feature information, spatiotemporal correlation information, weather warning information, and historical store sales information. Highlighting: These text fragments are highlighted in the visual interface, along with the corresponding target verification information, to help users understand how the model makes predictions based on this information.

[0114] In step S154, the consistency index between the attention weight corresponding to the target verification information and the manual expert judgment is calculated and displayed.

[0115] In the embodiment of the present disclosure, the consistency index between the attention weight corresponding to the target verification information and the judgment of the manual expert is calculated to evaluate the prediction performance of the model. Attention weight calculation: The explainability technology of the machine learning model is used to calculate the attention weight of the target verification information in the model prediction process. Manual expert judgment collection: Invite manual experts to judge the same situation and collect their judgment results. Consistency index calculation: Calculate the consistency index by comparing the attention weight and the judgment results of the manual expert. This can be achieved by calculating correlation coefficients (such as Pearson correlation coefficient), classification accuracy, etc. Display: The calculated consistency index is displayed to the user so that they can evaluate the prediction performance of the model.

[0116] The above technical solution develops a clause mapping interpreter to achieve the connection between technical explanation and business cognition: when the model recommends reducing the order quantity of a certain store, it not only displays the contribution of each feature, but also marks the specific policy items triggered; the text fragments that the model pays attention to (such as restrictive clauses in policy documents) are highlighted, and the consistency index of their attention weight and human expert judgment is calculated.

[0117] In a possible implementation, the method further includes:

[0118] Acquire abnormal sales information from the store's historical sales information, which indicates that the actual sales information of the store meets a preset elimination condition, and eliminate the store's historical sales information corresponding to the abnormal sales information;

[0119] Abnormal sales data refers to sales data within a store's historical sales data that significantly deviates from normal sales patterns. This may be caused by data entry errors, special events (such as unusually popular promotions), or other unusual factors. Pre-set exclusion criteria include standards or thresholds used to determine whether sales data is abnormal, such as sudden surges or decreases in sales volume or abnormal sales prices.

[0120] In this disclosed embodiment, the store's historical sales data is first traversed, and abnormal sales information is filtered out based on preset exclusion criteria. These criteria may be based on statistical methods (such as Z-score, IQR, etc.) or business rules (such as single-day sales exceeding three times the historical average). Once abnormal sales information is identified, it is removed from the historical sales data to ensure that the accuracy of subsequent analysis (such as predicting product order information) is not affected by these outliers.

[0121] Determining whether new target information exists in the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information, and determining the deviation of the new target from the historical distribution through adversarial training, and if the deviation does not meet a preset deviation threshold, removing the new target information, and if the new target information indicates that the target sales feature information has been newly added, marking and isolating the target sales feature information before and after the addition;

[0122] Among them, new target information: newly appeared or updated data in target sales feature information, spatiotemporal correlation information, weather warning information, and store historical sales information. Adversarial training: a machine learning technology that simulates an adversarial environment to evaluate the robustness of the model or judge the distribution difference between new data and historical data. Deviation: an indicator that measures the degree of difference between the new target information and the historical data distribution. Preset deviation threshold: a standard or threshold used to determine whether the new target information is acceptable. Labeling and isolation: marking the new target sales feature information and processing it separately from the original feature information for subsequent analysis and comparison.

[0123] In the embodiment of the present disclosure, the target sales feature information, spatiotemporal correlation information, weather warning information, and store historical sales information are regularly checked to see if there is any new target information. For these new information, the deviation from the historical data will be calculated through adversarial training or other distribution difference evaluation methods. If the deviation exceeds the preset threshold, it will be considered that these new information may not be representative or abnormal, and will be eliminated. In particular, if the new target information involves the update of the target sales feature information, the feature information before and after the addition will be marked and isolated, so that the impact of these changes on the prediction results can be analyzed separately later.

[0124] Automatically check whether the predicted commodity order information complies with preset industry rules, and re-execute the prediction of the predicted commodity order information if the predicted commodity order information does not comply with the preset industry rules.

[0125] In the disclosed embodiments, the predicted product order information output by the large model undergoes a compliance check to ensure that it complies with pre-set industry rules. These rules may include the rationality of order quantities (e.g., not exceeding inventory capacity, meeting minimum order quantity requirements), and the compliance of order times (e.g., avoiding holidays, and complying with supplier delivery cycles). If the predicted product order information does not comply with these rules, the prediction process is rerun, potentially optimizing the prediction results by adjusting model parameters, adding new features, or adding constraints, until the generated predicted product order information complies with pre-set industry rules.

[0126] Outlier filtering: Eliminate abnormal records such as zero sales due to system failures; Distribution consistency detection: Determine the degree of deviation between new data and historical distribution through adversarial training; Business logic verification: Automatically check whether the data complies with industry regulations (such as the single order quantity does not exceed the license limit); Timeliness classification: Label and isolate data before and after policy changes to avoid confusion between new and old rules.

[0127] Combining the above embodiments, we will use an example to illustrate data and knowledge base construction, and unstructured text processing using two-layer semantic parsing. The terminology recognition layer uses a pre-trained recognition model based on industry-specific terminology to identify core concepts and implicit constraints (such as "monopoly license" and "order quota") in policy texts, and captures ambiguous expressions in the policy through a bidirectional recurrent neural network. The relationship extraction layer automatically analyzes the logical connections between policy clauses, constructs a policy constraint chain, and ensures the integrity and coherence of policy information.

[0128] Spatiotemporal knowledge base indexing: Temporal semantic indexing: Converts the time information in policy texts into specific calendar intervals to create a time-sensitive semantic vector index. Spatial location indexing: Leveraging geolocation encoding technology, store address characteristics (such as business district level and distance to transportation hubs) are converted into spatial feature vectors, enabling intelligent matching of policy information with geographic dimensions.

[0129] Structured data processing: Aggregate and process structured information such as historical order data, sales data, and inventory data to form a standardized time series database, providing basic data support for online forecasting.

[0130] Furthermore, a dynamic fine-tuning mechanism of the large model is adopted to fine-tune the construction and use of the dataset: the fine-tuning dataset includes historical order quantity, sales volume, inventory data and store characteristics (geographic location, gear level, etc.), as well as semantic parsing results related to policy texts.

[0131] The difference from the knowledge base: the knowledge base stores parsed policy knowledge and structured historical data, while the fine-tuning dataset is specific examples selected from the knowledge base for model training.

[0132] The specific steps for fine-tuning instructions are: Historical business data and policy text are obtained and generated into "instruction-context-expected answer" triples. For example, "Predict this week's order volume based on store A's data from the past four weeks" serves as the instruction, store information and policy conditions serve as the context, and the expected order volume serves as the answer. Triples are manually annotated or automatically generated to ensure data coverage of various business scenarios. Triples guide the model in learning prediction and decision-making logic for specific task scenarios, which differs from traditional supervised learning methods.

[0133] Unfreeze and fine-tune: Unfreeze specific layers of a large model (such as the latter few layers of a Transformer) and adjust their parameters to adapt to the business data; these parameters refer to the neural network weights and biases. Business data includes store sales data, policy-related data, and environmental data. Multiple rounds of training are used, with the data randomly shuffled in each round. Parameters are iteratively optimized until the evaluation metrics are optimized.

[0134] Perform evaluation and hyperparameter tuning: Evaluate model performance on the validation set (comparing real order data with model prediction results), using metrics such as mean squared error (MSE) and mean absolute error (MAE).

[0135] Hyperparameters such as the number of fine-tuning rounds, learning rate, and batch size are adjusted until the validation set performance reaches the optimal level. Compared with the existing technology, a special evaluation of the impact of policy text is added.

[0136] After fine-tuning, the large model uses LoRA or full fine-tuning (LoRA is the most important due to its high computational efficiency and significant results), enabling the large model to understand industry terms, predict order quantities, and explain key reasons. This explanatory ability comes from the built-in attention mechanism of the large model.

[0137] Online prediction process and multi-dimensional evidence fusion reasoning: Real-time retrieval and context building: For example, input: "Store A is located in a core business district. Sales of brand B cigarettes have increased weekly over the past four weeks. A new promotion policy was recently released." The corresponding policy text and historical sales data are retrieved to form a complete context. Three-layer evidence fusion reasoning: The large model uses a multi-head attention mechanism to calculate the attention weights of structured evidence, policy evidence, and environmental evidence, and then fuses the weights to calculate the final prediction.

[0138] Model prediction interpretability and business post-processing: Interpretable output mechanism: Provide natural language or visual explanations of prediction results, clearly labeling the triggered policy clauses, such as "According to regulations, the store's order quantity is reduced." Key policy text snippets in the decision-making process are highlighted, and auxiliary information such as feature contribution and attention weight is provided to help intuitively understand the prediction basis.

[0139] Flexible business post-processing rules: Develop business rules for forecast output, such as setting upper and lower limits for forecast values ​​and holiday ordering strategies, to ensure that forecast results are suitable for specific business scenarios.

[0140] Incremental learning in a closed-loop business feedback loop: Actual execution data reflow mechanism: Regularly collect actual order execution data and the "reasonable order quantity" adjusted by business personnel, mark the differences between predictions and actual executions, and reflow them for incremental training.

[0141] Policy memory and replay mechanism: This mechanism stores and labels historical typical decision-making scenarios (such as promotions and new store openings) and performs weighted policy replay to ensure the continuity and stability of key business strategies. Highly weighted sampling is set for regulatory-sensitive decision clauses to prevent incremental learning from causing the forgetting of regulatory-sensitive knowledge.

[0142] Rapid expansion and adaptation mechanism: Rapid model adaptation; use lightweight fine-tuning technologies (such as LoRA, Prefix-Tuning, Adapter) to achieve rapid model adaptation for new stores, new product specifications, and new policies.

[0143] Modular knowledge base update: A modular knowledge base storage structure is adopted to enable new policy documents or business rules to be quickly integrated into the knowledge base, and quickly adapt to new changes through a small amount of training data.

[0144] Compared with existing order forecasting technologies based on a single data source or traditional statistical / machine learning methods, the introduction of a large model and knowledge base retrieval mechanism enables multi-dimensional in-depth learning and dynamic understanding of ordering behavior. This is specifically reflected in the following aspects:

[0145] Improved prediction accuracy: Based on the integration of multi-source data such as historical orders, sales, inventory, store locations, and industry policies, the large model can capture more complex time series characteristics and store differences through deep learning and fine-tuning, effectively reducing prediction errors and lowering the risks of "out-of-stock" and "backlog".

[0146] Enhanced business decision-making flexibility: Relying on knowledge base retrieval, key text information such as promotion rules and regulatory restrictions can be automatically retrieved before prediction. This further assists the large model in making reasonable assessments of specific stores and special situations (holidays, seasonal fluctuations, sudden events, etc.), providing management with diversified ordering plans that can be adjusted according to actual conditions.

[0147] Explainability and business trust: When providing prediction results, large models can explain the main influencing factors through natural language or visualization, improving transparency and trust in the prediction process, and helping business personnel make quick decisions and adjust business strategies.

[0148] Continuous iteration capability: By collecting actual order and sales feedback after each business execution, incremental fine-tuning or retraining can be performed to allow the model to continuously adapt to market changes and individual store needs, forming a closed-loop management model and maintaining continuous optimization of predictive performance.

[0149] Deployment and scalability efficiency: Fine-tuning based on pre-trained models significantly reduces reliance on large-scale training data. When new stores, brands, or policies are added, the existing system can be quickly expanded, enabling flexible adaptation to retail networks of varying sizes.

[0150] To sum up, it has significant advantages in high-precision prediction, explainability, business agility and continuous iterative upgrades, and can effectively improve order management efficiency and the overall competitiveness of the supply chain.

[0151] In one embodiment, Figure 2 As shown, a commodity order forecasting device based on a large model is provided, comprising:

[0152] The sales information determination module 210 is configured to determine target sales feature information from the sales guidance document of the product based on the pre-trained industry dictionary of the industry in which the product is located through the policy analysis module in the large model;

[0153] The spatiotemporal association module 220 is configured to obtain special time description information, promotion time information, store address information, and promotion terms information of the ordering store through the spatiotemporal association module in the large model, and determine the spatiotemporal association information of the store based on the special time description information, the promotion time information, the store address information, and the promotion terms information;

[0154] The information acquisition module 230 is configured to acquire weather warning information and store historical sales information through the multi-dimensional data acquisition module in the large model;

[0155] The prediction module 240 is configured to determine the predicted product order information of the store by dynamically adjusting the weights through an attention weight allocation mechanism based on the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information using the sales prediction module of the large model;

[0156] The visualization display module 250 is configured to visualize the target display information when the predicted commodity order information meets the preset conditions through the risk warning module of the large model.

[0157] In a possible implementation, the sales information determination module 210 is configured to:

[0158] The policy analysis module in the large model obtains corresponding sales feature information based on the pre-trained industry dictionary for the product's industry, and uses the bidirectional recurrent neural network in the two-layer semantic parsing engine to determine implicit restrictions from the sales feature information to obtain the actual sales feature information of the product in the sales guide;

[0159] Using the relationship extraction layer in the two-layer semantic parsing engine, logical dependencies between policy clauses are established to obtain a sales constraint chain;

[0160] The target selling characteristic information is determined according to the actual selling characteristic information and the actual selling characteristic information.

[0161] In a possible implementation, the sales information determination module 210 is configured to:

[0162] In response to obtaining a new sales guidance document, determining sensitive sales parameters from the original sales guidance document;

[0163] By protecting sensitive neurons, the adjustment range of the sensitive sales parameters is limited during incremental training to prevent new sales guidance documents from overwriting the sensitive sales parameters;

[0164] According to the new sales guidance document, the policy analysis module is incrementally trained to obtain a new policy analysis module.

[0165] In a possible implementation, the spatiotemporal correlation module 220 is configured to:

[0166] Performing calendar conversion on the time description in the promotion time information, binding the text semantics of the time description with the date range, and determining the promotion time window;

[0167] During vector retrieval, based on the time validity corresponding to the promotion time window, the similarity between the special time description information and the promotion time window is synchronously constrained, and target promotion term information is determined from the promotion term information;

[0168] Performing vector conversion on the store address information according to the geographic location encoder to convert the store address into a business district feature vector, wherein the business district feature vector includes at least one of the following: school density and distance to a transportation hub;

[0169] The spatiotemporal association information of the store is determined by associating the promotion time window, the target promotion term information, and the business district feature vector.

[0170] In a possible implementation, the apparatus further includes an updating module configured to:

[0171] Obtaining actual product order information of the store for the predicted product order information, and determining whether a triggering update condition is met based on the actual product order information and the predicted product order information;

[0172] When the triggering update condition is met, analyzing the updated policy information of the sales guidance document and / or monitoring the sudden change information of the foot traffic in the business district where the store address is located through text similarity analysis;

[0173] The large model is updated according to the update policy information and / or the pedestrian flow mutation information to obtain a new large model.

[0174] In a possible implementation, the visualization display module 250 is configured to:

[0175] If the predicted product order information output by the large model indicates a recommendation to reduce the product order quantity of the store, displaying the contribution of at least one of the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information;

[0176] Label and display the policy items that trigger the recommendation to reduce the order quantity of the product in question;

[0177] Highlight the text segment that the large model focuses on, and determine the target verification information corresponding to the text segment from the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information;

[0178] Calculate and display the consistency index between the attention weight corresponding to the target verification information and the human expert judgment.

[0179] In a possible implementation, the apparatus further includes: a data processing module configured to:

[0180] Acquire abnormal sales information from the store's historical sales information, which indicates that the actual sales information of the store meets a preset elimination condition, and eliminate the store's historical sales information corresponding to the abnormal sales information;

[0181] Determining whether new target information exists in the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information, and determining the deviation of the new target from the historical distribution through adversarial training, and if the deviation does not meet a preset deviation threshold, removing the new target information, and if the new target information indicates that the target sales feature information has been newly added, marking and isolating the target sales feature information before and after the addition;

[0182] Automatically check whether the predicted commodity order information complies with preset industry rules, and re-execute the prediction of the predicted commodity order information if the predicted commodity order information does not comply with the preset industry rules.

[0183] Regarding the specific definition of a large-scale model-based commodity order forecasting device, please refer to the definition of a large-scale model-based commodity order forecasting method above, which will not be repeated here. The various modules in the above-mentioned large-scale model-based commodity order forecasting device can be implemented in whole or in part through software, hardware, or a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the corresponding operations of the above modules.

[0184] An embodiment of the present disclosure further provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method described in any one of the aforementioned embodiments when the program is executed by a processor.

[0185] The present disclosure also provides an electronic device, including:

[0186] a memory having a computer program stored thereon;

[0187] A processor is used to execute the computer program in the memory to implement the steps of the method in any one of the aforementioned embodiments.

[0188] Figure 3 The large-scale model-based commodity order forecasting device 100 shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the large-scale model-based commodity order forecasting device 100 may further include a communication component, which can be used for data interaction between the device 100 and other devices, such as sending or receiving data. It should be noted that the communication component in actual scheduling is not limited to one, and the structure of the large-scale model-based commodity order forecasting device 100 does not constitute a limitation on the embodiments of the present application.

[0189] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, or a combination of a DSP and a microprocessor.

[0190] Bus 1002 may include a path for transmitting information between the above components. Bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. Bus 1002 may be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, Figure 3 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0191] The memory 1003 may be a ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, a RAM (Random Access Memory) or other type of dynamic storage device that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium that can be used to carry or store program code and can be read by a computer, without limitation herein.

[0192] The memory 1003 is used to store program codes for executing the embodiments of the present disclosure, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the program codes stored in the memory 1003 to implement the steps shown in the embodiment of the aforementioned large model-based commodity order forecasting method.

[0193] An embodiment of the present disclosure also provides a computer-readable storage medium having program code stored thereon. When the program code is executed by a processor, the steps and corresponding contents of the aforementioned embodiment of a commodity order forecasting method based on a large model can be implemented.

[0194] The preferred embodiments of the present disclosure are described in detail above in conjunction with the accompanying drawings. However, the present disclosure is not limited to the specific details in the above embodiments. Within the technical concept of the present disclosure, various changes, modifications, replacements and variations can be made to these embodiments, and these changes, modifications, replacements and variations all fall within the scope of protection of the present disclosure.

[0195] It should also be noted that the various specific technical features described in the above specific embodiments may be combined in any suitable manner, unless there is any contradiction, and these combinations shall also be considered as the contents disclosed in this disclosure. To avoid unnecessary repetition, this disclosure will not further describe various possible combinations. The technical scope of this application is not limited to the contents of the specification and must be determined based on the scope of the claims.

Claims

1. A commodity order forecasting method based on a large model, characterized in that: The method comprises: Determining target sales feature information from the product's sales guidance document using a policy analysis module within the large model based on a pre-trained industry dictionary for the product's industry. The target sales feature information is key information directly related to the product's sales extracted from the sales guidance document, including sales restrictions, promotion policies, and age restrictions. Obtaining special time description information, promotion time information of the ordering store, store address information, and promotion terms information through the spatiotemporal association module in the large model, and determining the spatiotemporal association information of the store based on the special time description information, the promotion time information, the store address information, and the promotion terms information; Acquire weather warning information and store historical sales information through the multidimensional data acquisition module in the large model; The sales forecast module of the large model dynamically adjusts the weights based on the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information through an attention weight allocation mechanism to determine the store's predicted product order information; When the predicted commodity order information meets the preset conditions, the target display information is visualized through the risk warning module of the large model.

2. The method according to claim 1, characterized in that The policy analysis module in the large model determines target sales feature information from the sales guidance document of the product based on the pre-trained industry dictionary of the industry in which the product is located, including: The policy analysis module in the large model obtains corresponding sales feature information based on the pre-trained industry dictionary for the product's industry, and uses the bidirectional recurrent neural network in the two-layer semantic parsing engine to determine implicit restrictions from the sales feature information to obtain the actual sales feature information of the product in the sales guide; Using the relationship extraction layer in the two-layer semantic parsing engine, logical dependencies between policy clauses are established to obtain a sales constraint chain; The target selling characteristic information is determined according to the actual selling characteristic information and the actual selling characteristic information.

3. The method according to claim 2, characterized in that The method further comprises: In response to obtaining a new sales guidance document, determining sensitive sales parameters from the original sales guidance document; By protecting sensitive neurons, the adjustment range of the sensitive sales parameters is limited during incremental training to prevent new sales guidance documents from overwriting the sensitive sales parameters; According to the new sales guidance document, the policy analysis module is incrementally trained to obtain a new policy analysis module.

4. The method according to claim 1, characterized in that The determining of the spatiotemporal association information of the store according to the special time description information, the promotion time information, the store address information, and the promotion terms information includes: Performing calendar conversion on the time description in the promotion time information, binding the text semantics of the time description with the date range, and determining the promotion time window; During vector retrieval, based on the time validity corresponding to the promotion time window, the similarity between the special time description information and the promotion time window is synchronously constrained, and target promotion term information is determined from the promotion term information; Performing vector conversion on the store address information according to the geographic location encoder to convert the store address into a business district feature vector, wherein the business district feature vector includes at least one of the following: school density and distance to a transportation hub; The spatiotemporal association information of the store is determined by associating the promotion time window, the target promotion term information, and the business district feature vector.

5. The method according to any one of claims 1 to 4, characterized in that The method further comprises: Obtaining actual product order information of the store for the predicted product order information, and determining whether a triggering update condition is met based on the actual product order information and the predicted product order information; When the triggering update condition is met, analyzing the updated policy information of the sales guidance document and / or monitoring the sudden change information of the foot traffic in the business district where the store address is located through text similarity analysis; The large model is updated according to the update policy information and / or the pedestrian flow mutation information to obtain a new large model.

6. The method according to any one of claims 1 to 4, characterized in that The visual display of the target display information includes: If the predicted product order information output by the large model indicates a recommendation to reduce the product order quantity of the store, displaying the contribution of at least one of the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information; Label and display the policy items that trigger the recommendation to reduce the order quantity of the product in question; Highlighting the text segment focused on by the large model, and determining target verification information corresponding to the text segment from the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information; Calculate and display the consistency index between the attention weight corresponding to the target verification information and the human expert judgment.

7. The method according to any one of claims 1 to 4, characterized in that The method further comprises: Acquire abnormal sales information from the store's historical sales information, which indicates that the actual sales information of the store meets a preset elimination condition, and eliminate the store's historical sales information corresponding to the abnormal sales information; Determining whether new target information exists in the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information, and determining the deviation of the new target from the historical distribution through adversarial training, and if the deviation does not meet a preset deviation threshold, removing the new target information, and if the new target information indicates that the target sales feature information has been newly added, marking and isolating the target sales feature information before and after the addition; Automatically check whether the predicted commodity order information complies with preset industry rules, and re-execute the prediction of the predicted commodity order information if the predicted commodity order information does not comply with the preset industry rules.

8. A commodity order forecasting device based on a large model, characterized in that: The device comprises: a sales information determination module configured to determine target sales feature information from the sales guidance document of the product using the policy analysis module in the large model and based on a pre-trained industry dictionary for the industry in which the product is located, wherein the target sales feature information is key information directly related to the sales of the product extracted from the sales guidance document, and the target sales feature information includes sales restrictions, promotion policies, and age restrictions; a spatiotemporal association module configured to obtain, through the spatiotemporal association module in the large model, special time description information, promotion time information, store address information, and promotion terms information of the ordering store, and determine the spatiotemporal association information of the store based on the special time description information, the promotion time information, the store address information, and the promotion terms information; An information acquisition module configured to acquire weather warning information and store historical sales information through the multidimensional data acquisition module in the large model; a prediction module configured to determine predicted product order information for the store by dynamically adjusting weights using an attention weight allocation mechanism based on the target sales feature information, the spatiotemporal correlation information, the weather warning information, and the store's historical sales information using the sales prediction module of the large model; The visualization display module is configured to visualize the target display information when the predicted commodity order information meets the preset conditions through the risk warning module of the large model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. An electronic device, characterized in that: include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the steps of the method according to any one of claims 1 to 7.

Images