Commodity ordering prediction method and device based on large model, equipment and medium

Abstract

The invention relates to a commodity ordering prediction method and device based on a large model, equipment and a medium, and the method comprises the steps: determining target selling feature information from a selling guidance file of a commodity according to a pre-trained industry dictionary of the industry where the commodity is located; acquiring special time description information, promotion time information of the ordering shop, shop address information and promotion clause information, and determining time-space association information according to the special time description information, the promotion time information, the shop address information and the promotion clause information; the acquisition module acquires weather early warning information and historical sales information of a shop; according to the target selling feature information, the time-space association information, the weather early warning information and the historical selling information of the shop, dynamically adjusting the weight through an attention weight distribution mechanism, and determining predicted commodity ordering information of the shop; and performing visual display on the target display information through a risk early warning module of the large model under the condition that the commodity ordering information is predicted to meet a preset condition.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and in particular, to a commodity order prediction method, device, equipment and medium based on a large model. Background Art

[0002] With the continuous development of the retail industry and the improvement of the informatization level, the sales and order placement models of fast-moving consumer goods have become increasingly complex. Traditional order quantities usually rely on the empirical judgment of store managers or decision-making based on simple historical statistical data, which can meet the needs in a small-scale and stable market environment to a certain extent. However, with the continuous increase in the number of stores and product specifications, and the dynamic changes in consumer preferences, it has become difficult to accurately grasp the order demand simply relying on experience or simple time series forecasting of orders, and problems such as inventory backlogs or out-of-stock shortages often occur, seriously affecting sales efficiency and customer satisfaction.

[0003] Currently, order quantity prediction methods based on statistics or traditional machine learning models can be adopted, such as linear regression, ARIMA, XGBoost, etc. These methods can indeed improve the prediction effect when the data scale is relatively limited or the scenario is relatively single. However, in the face of multi-dimensional data in actual operations (such as historical order data, sales data, inventory data, and the grade, location, industry attributes, etc. of each store), it is difficult to fully explore the associations between complex features. At the same time, when there are special periods such as holidays and promotional activities, the prediction results are prone to large deviations and it is 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 Transformer structures) for feature extraction and prediction. On the one hand, large models have obvious advantages in processing high-dimensional and multi-modal data; on the other hand, large models can automatically learn potential time series laws and associated features based on a large amount of historical data, providing more accurate and flexible prediction capabilities for order behavior analysis. However, when applied to the order scenario, there are still problems such as 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. 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 problems such as 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 applying multi-layer neural networks and attention mechanisms to the commodity order scenario.

[0006] To achieve the above purpose, in the first aspect of the embodiments of the present disclosure, a commodity order prediction method based on a large model is provided, and the method includes: Through the policy analysis module in the large model, according to the industry dictionary pre-trained for the industry where the commodity is located, determine the target selling characteristic information from the sales guidance document of the commodity; obtain the special time description information, the promotion time information of the ordering store, the store address information, and the promotion terms information through the spatio-temporal association module in the large model, and determine the spatio-temporal 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; obtain the meteorological warning information and the store historical sales information through the multi-dimensional data acquisition module in the large model; through the sales volume prediction module of the large model, according to the target selling characteristic information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information, dynamically adjust the weights through the attention weight allocation mechanism, and determine the predicted commodity ordering information of the store; through the risk warning module of the large model, when the predicted commodity ordering information meets the preset conditions, visually display the target display information.

[0007] In a possible implementation manner, the step of determining the target selling characteristic information from the sales guidance document of the commodity through the policy analysis module in the large model according to the industry dictionary pre-trained for the industry where the commodity is located includes: Through the policy analysis module in the large model, according to the industry dictionary pre-trained for the industry where the commodity is located, obtain the corresponding selling characteristic information, and use the bidirectional recurrent neural network in the double-layer semantic parsing engine to determine the implicit restriction conditions from the selling characteristic information to obtain the actual selling characteristic information of the commodity in the sales guidance; use the relationship extraction layer in the double-layer semantic parsing engine to establish the logical dependence relationship between policy clauses to obtain the sales constraint chain; according to the actual selling characteristic information and the actual selling characteristic information, determine the target selling characteristic information.

[0008] In a possible implementation manner, the method further includes: In response to obtaining a new sales guidance document, determine the sensitive sales parameters from the original sales guidance document; protect the sensitive neurons and limit the adjustment range of the sensitive sales parameters during incremental training to prevent the new sales guidance document from overwriting the sensitive sales parameters; perform incremental training on the policy analysis module according to the new sales guidance document to obtain a new policy analysis module.

[0009] In a possible implementation manner, the step of determining the spatio-temporal 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: Perform calendar conversion on the time description in the promotion time information, bind the text semantics of the time description to a date range, and determine the promotion time window; during vector retrieval, synchronously constrain the similarity between the special time description information and the promotion time window according to the time validity corresponding to the promotion time window, and determine the target promotion clause information from the promotion clause information; according to the geographical location encoder, perform vector conversion on the store address information, and convert the store address into a business district feature vector, where the business district feature vector includes at least one of the following: school density, distance to transportation hubs; correlate the promotion time window, the target promotion clause information, and the business district feature vector to determine the spatio-temporal association information of the store.

[0010] In a possible implementation manner, the method further includes: Obtain the actual commodity order information of the store for the predicted commodity order information, and determine whether the trigger update condition is met according to the actual commodity order information and the predicted commodity order information; in the case where the trigger update condition is met, analyze the update policy information of the sales guidance document through text similarity and / or monitor the sudden change information of the pedestrian flow in the business district where the store address is located; update the large model according to the update policy information and / or the pedestrian flow sudden change information to obtain a new large model.

[0011] In a possible implementation manner, the visual display of the target display information includes: In the case where the predicted commodity order information output by the large model recommends reducing the commodity order quantity of the store, display the contribution degree of at least one of the target sales feature information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information; mark and display the policy item that triggers the recommendation to reduce the commodity order quantity of the store; highlight the text segment concerned by the large model, and determine the target verification information corresponding to the text segment from the target sales feature information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information; calculate and display the consistency index between the attention weight corresponding to the target verification information and the judgment of artificial experts.

[0012] In a possible implementation manner, the method further includes: Obtain the abnormal sales information in the historical sales information of the store that indicates that the actual sales information of the store meets the preset exclusion conditions, and exclude the historical sales information of the store corresponding to the abnormal sales information; determine that there is new target information in the target sales characteristic information, the spatio-temporal association information, the meteorological warning information, and the historical sales information of the store, and judge the deviation degree between the new target information and the historical distribution through adversarial training. In the case where the deviation degree does not meet the preset deviation threshold, exclude the new target information, and in the case where the new target information indicates the addition of the target sales characteristic information, perform labeling isolation on the target sales characteristic information before and after the addition; automatically check whether the predicted commodity order information conforms to the preset industry rules, and in the case where the predicted commodity order information does not conform to the preset industry rules, re-perform the prediction of the predicted commodity order information.

[0013] In a second aspect of the embodiments of the present disclosure, there is provided a commodity order prediction device based on a large model, the device includes: A sales information determination module, configured to determine target sales characteristic information from the sales guidance document of the commodity through the policy analysis module in the large model according to the industry dictionary pre-trained for the industry to which the commodity belongs; A spatio-temporal association module, configured to obtain special time description information, promotion time information of the ordering store, store address information, and promotion clause information through the spatio-temporal association module in the large model, and determine the spatio-temporal association information of the store according to the special time description information, the promotion time information, the store address information, and the promotion clause information; An information acquisition module, configured to obtain meteorological warning information and historical sales information of the store through the multi-dimensional data acquisition module in the large model; A prediction module, configured to determine the predicted commodity order information of the store through the sales volume prediction module of the large model, dynamically adjust the weights through the attention weight distribution mechanism according to the target sales characteristic information, the spatio-temporal association information, the meteorological warning information, and the historical sales information of the store; A visualization display module, configured to perform visualization display on the target display information through the risk warning module of the large model when the predicted commodity order information meets the preset conditions.

[0014] In a third aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in any one of the first aspects are implemented.

[0015] In a fourth aspect of the embodiments of the present disclosure, there is provided an electronic device, including: A memory on which a computer program is stored; a processor for executing the computer program in the memory to implement the steps of the method according to any one of the first aspects.

[0016] The present invention provides a commodity order prediction method, device, equipment and medium based on a large model. Compared with the prior art, it has the following beneficial effects: By using the policy analysis module and the industry dictionary, it is possible to accurately extract the target sales feature information from the commodity sales guidance document, ensure that the analysis is based on accurate and targeted industry data, and improve the accuracy of subsequent analysis. The spatio-temporal correlation module comprehensively captures the operating characteristics of the store in different spatio-temporal backgrounds by integrating multi-dimensional information such as special time descriptions, promotion times, store addresses and promotion terms, providing richer context information for sales volume prediction and enhancing the spatio-temporal adaptability of the prediction.

[0017] The multi-dimensional data acquisition module can acquire meteorological warning information and store historical sales information, combine external weather with internal sales data, provide more comprehensive data support for sales volume prediction, and help to more accurately evaluate various factors affecting sales. By adopting the attention weight allocation mechanism, it is possible to dynamically adjust the weights of various factors according to the target sales feature information, spatio-temporal correlation information, meteorological warning information and store historical sales information, realize more flexible and accurate sales volume prediction, and effectively respond to market changes.

[0018] When the predicted commodity order information meets the preset conditions, the visual display is automatically triggered, and the risk warning information is provided to the user in a timely manner, helping the user make decisions quickly, reducing potential losses, and improving the efficiency and intuitiveness of risk management.

[0019] In this way, through the full processing of multi-dimensional data, when the model is applied to the commodity order scenario, by integrating multi-source data, dynamically adjusting the prediction weights and providing risk warning visual display, the accuracy and timeliness of commodity order prediction are significantly improved, the multi-dimensional information of different stores is effectively utilized, the model results are applied to actual business decisions and the prediction is updated in real time, the prediction accuracy of commodity order is improved, and the differentiated needs of the market and stores are met.

[0020] Other features and advantages of the present disclosure will be described in detail in the subsequent specific implementation part. Brief Description of the Drawings

[0021] The drawings are used to provide a further understanding of the present disclosure, and constitute a part of the specification, and are used to explain the present disclosure together with the following specific implementation, but do not constitute a limitation to the present disclosure. In the drawings: Figure 1 is a flowchart showing a commodity order prediction method based on a large model according to an embodiment of the specification.

[0022] Figure 2 Block diagram of a commodity order prediction device based on a large model shown according to the embodiments of the specification.

[0023] Figure 3 It is a block diagram of another commodity order prediction device based on a large model shown according to the embodiments of the specification. Detailed implementation manners

[0024] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0025] The present application provides a commodity order prediction method based on a large model. Figure 1 It is a flowchart of a commodity order prediction method based on a large model shown according to an embodiment. The method includes: In step S11, through the policy analysis module in the large model, according to the industry dictionary pre-trained for the industry to which the commodity belongs, determine the target selling feature information from the sales guidance document of the commodity; Among them, the policy analysis module is a sub-module in the large model, which is specifically used to parse and understand policy documents related to commodity sales. The industry dictionary is a vocabulary set pre-trained for a specific industry, including unique terms, regulations and expressions in that industry. The target selling feature information is key information directly related to commodity sales extracted from the sales guidance document, such as sales restrictions, promotion policies, age restrictions, etc.

[0026] In the embodiments of the present disclosure, the policy analysis module uses the pre-trained industry dictionary to perform text parsing and semantic understanding on the sales guidance document of the commodity, and identifies and extracts feature information directly related to commodity sales. These information are the basis for subsequent sales volume prediction and risk warning.

[0027] In step S12, obtain special time description information, promotion time information of the ordering store, store address information and promotion clause information through the spatio-temporal correlation module in the large model, and determine the spatio-temporal correlation information of the store according to the special time description information, the promotion time information, the store address information and the promotion clause information; Among them, the spatio-temporal correlation module is a sub-module in the large model, which is used to integrate and analyze information related to time and space. Special time description information is information with time characteristics such as holidays and special events. Promotion time information is the time of the promotion activities set by the store. Store address information is the specific geographical location of the store. Promotion clause information is the specific terms and conditions related to the promotion activities. Spatio-temporal correlation information is the spatio-temporal characteristic information related to the store sales obtained by comprehensively considering time, space and promotion factors.

[0028] In the embodiments of the present disclosure, the spatio-temporal correlation module determines the spatio-temporal correlation information of the store by obtaining and integrating special time description information, promotion time information, store address information and promotion clause information, and analyzing the correlation and influence between these information. These information helps to more accurately predict the sales situation of the store under different time and space conditions.

[0029] In step S13, the meteorological warning information and the store historical sales information are obtained through the multi-dimensional data acquisition module in the large model; Among them, the multi-dimensional data acquisition module is a sub-module in the large model, which is responsible for obtaining data related to commodity sales from multiple data sources. Meteorological warning information is warning information related to weather, such as heavy rain and high temperature. Store historical sales information is the sales data of the store in the past period of time.

[0030] In the embodiments of the present disclosure, the multi-dimensional data acquisition module obtains the meteorological warning information and the store historical sales information by connecting to external data sources (such as meteorological bureaus, store sales systems, etc.). These information provides more comprehensive data support for sales volume prediction, and helps to more accurately evaluate various factors affecting sales.

[0031] In step S14, through the sales volume prediction module of the large model, according to the target selling characteristic information, the spatio-temporal correlation information, the meteorological warning information and the store historical sales information, the weights are dynamically adjusted through the attention weight distribution mechanism, and the predicted commodity ordering information of the store is determined; Among them, the sales volume prediction module is a sub-module in the large model, which is used to predict the sales volume of commodities according to various factors. The attention weight distribution mechanism can dynamically adjust the weights of various factors and allocate different weights according to the influence degree of each factor on the sales volume. The predicted commodity ordering information is the quantity of commodities recommended for the store to order according to the sales volume prediction result.

[0032] In the embodiments of the present disclosure, the sales volume prediction module dynamically adjusts the weights of various factors by using the attention weight distribution mechanism according to the target selling characteristic information, the spatio-temporal correlation information, the meteorological warning information and the store historical sales information, so as to determine the predicted commodity ordering information of the store. This mechanism can more flexibly respond to market changes and improve the accuracy of sales volume prediction.

[0033] The above technical solution performs multi-dimensional evidence fusion reasoning, breaks through the traditional single-data reasoning mode, and constructs a three-layer evidence chain system: structured evidence: directly call the historical data of the store in the database; policy evidence: retrieve the latest industry norms in real time; environmental evidence: integrate external data (such as the impact assessment of meteorological warning information on logistics); through the attention weight allocation mechanism, dynamically adjust the influence ratio of the three types of evidence on the final decision.

[0034] In step S15, when the risk warning module of the large model determines that the predicted commodity order information meets the preset conditions, the target display information is visually displayed.

[0035] Among them, the risk warning module is a sub-module in the large model, which is used to issue a risk warning when the prediction result meets the preset conditions. The preset conditions are pre-set thresholds or rules for judging whether a risk warning needs to be issued. The target display information is the risk warning information that needs to be displayed to the user. Visual display is to visually display the risk warning information to the user in the form of graphs, charts, etc.

[0036] In the embodiments of the present disclosure, the differences between the actual order execution data and the model prediction are collected and labeled, especially the "reasonable order quantity" after manual correction, and regularly fed back to the incremental fine-tuning training; a strategy replay mechanism for historical typical business scenarios (such as promotional activities, seasonal ordering strategies) is clearly set up to protect core business knowledge in the form of weights; special sampling weights are set for the learning of sensitive clauses of policies and regulations to avoid forgetting of regulatory information and ensure the continuous validity of compliance knowledge.

[0037] In the embodiments of the present disclosure, the dynamically adapted fine-tuning mechanism divides the large model into a policy understanding module, a sales volume prediction module, and a risk warning module, and dynamically activates the 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 normal scenarios, only the basic sales volume prediction module is activated to reduce resource consumption; a special protection mechanism is set specifically for neurons sensitive to regulations (such as parameters related to monopoly licenses) to avoid overwriting or forgetting of regulatory-sensitive knowledge during model incremental training.

[0038] In the embodiments of the present disclosure, when the risk warning module determines that the predicted commodity order information meets the preset conditions (such as the predicted sales volume is much lower or higher than the historical average level), the target display information is visually displayed. This display method helps users quickly understand the risk situation and make corresponding decisions.

[0039] For the characteristics of the tobacco industry, the parameter adjustment system for business scenario perception divides the large model into independent functional modules such as policy understanding, sales volume prediction, and risk warning through modular fine-tuning components, and dynamically activates relevant modules through a gating mechanism; when it detects that the input data involves the latest policy, it preferentially calls the policy analysis module; when making a regular order prediction, it enables the basic prediction module; compliance protection mechanism: identify the neurons in the model parameters that are sensitive to regulations (such as decision nodes involving the verification of monopoly licenses), and limit the adjustment range of these parameters during incremental training to prevent new knowledge from overwriting the memory of key regulations.

[0040] The above technical solution uses the policy analysis module and the industry dictionary to accurately extract the target selling feature information from the commodity sales guidance document, ensuring that the analysis is based on accurate and targeted industry data, and improving the accuracy of subsequent analysis. The spatio-temporal association module comprehensively captures the operating characteristics of the store under different spatio-temporal backgrounds by integrating multi-dimensional information such as special time descriptions, promotion times, store addresses, and promotion terms, providing richer context information for sales volume prediction and enhancing the spatio-temporal adaptability of the prediction.

[0041] The multi-dimensional data acquisition module can obtain meteorological warning information and the historical sales information of the store, combine external weather with internal sales data, provide more comprehensive data support for sales volume prediction, and help to more accurately evaluate various factors affecting sales. By adopting the attention weight allocation mechanism, it can dynamically adjust the weights of various factors according to the target selling feature information, spatio-temporal association information, meteorological warning information, and the historical sales information of the store, realizing more flexible and accurate sales volume prediction and effectively coping with market changes.

[0042] When the predicted commodity order information meets the preset conditions, it automatically triggers a visual display, provides risk warning information to the user in a timely manner, helps the user make decisions quickly, reduces potential losses, and improves the efficiency and intuitiveness of risk management.

[0043] In this way, through the full processing of multi-dimensional data, when the model is applied to the commodity order scenario, by integrating multi-source data, dynamically adjusting the prediction weights, and providing a visual display of risk warnings, it significantly improves the accuracy and timeliness of commodity order prediction, effectively utilizes the multi-dimensional information of different stores, applies the model results to actual business decisions, and updates the prediction in real time, improving the prediction accuracy of commodity orders and meeting the differentiated needs of the market and stores.

[0044] In one possible implementation, in step S11, determining the target selling feature information from the commodity sales guidance document by the policy analysis module in the large model according to the industry dictionary pre-trained according to the industry to which the commodity belongs includes: In step S111, through the policy analysis module in the large model, according to the industry dictionary pre-trained for the industry where the commodity is located, obtain the corresponding sales feature information, and use the bidirectional recurrent neural network in the two-layer semantic parsing engine to determine the implicit restriction conditions from the sales feature information, so as to obtain the actual sales feature information of the commodity in terms of sales guidance; Among them, 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, and the second layer performs in-depth semantic understanding and relationship extraction. The bidirectional recurrent neural network (BiRNN) is a neural network that can process sequential data, and it improves the accuracy of parsing by considering the context information of the sequence. The sales feature information is the feature description directly related to the commodity sales extracted from the sales guidance document, such as price, promotion method, sales target, etc. The implicit restriction condition is the restriction condition that is not directly and clearly stated in the sales guidance document but can be inferred through semantic parsing, such as sales time limit, sales area limit, etc. The actual sales feature information is the complete and accurate feature description of the commodity in terms of sales guidance after combining the sales feature information and the implicit restriction conditions.

[0045] In the embodiment of the present disclosure, the policy analysis module first obtains the corresponding sales feature information from the sales guidance document according to the industry dictionary pre-trained for the industry where the commodity is located. Then, using the bidirectional recurrent neural network (BiRNN) in the two-layer semantic parsing engine, perform in-depth semantic parsing on these sales feature information, identify and extract the implicit restriction conditions therein. By combining the sales feature information and the implicit restriction conditions, obtain the actual sales feature information of the commodity in terms of sales guidance.

[0046] In step S112, use the relationship extraction layer in the two-layer semantic parsing engine to establish the logical dependency relationship between policy clauses, and obtain the sales constraint chain; Among them, the relationship extraction layer: the second layer of the two-layer semantic parsing engine, responsible for extracting the relationship between entities from the text and establishing the logical dependency relationship. Policy clause: the specific regulations or requirements in the sales guidance document. Logical dependency relationship: the interdependent or influential relationship existing between policy clauses, such as the implementation of one clause may depend on the satisfaction of another clause. Sales constraint chain: a chain composed of multiple policy clauses and their logical dependency relationships, used to describe various constraints and restrictions in the commodity sales process.

[0047] In the embodiment of the present disclosure, use the relationship extraction layer in the two-layer semantic parsing engine to analyze each policy clause in the sales guidance document one by one, identify and establish the logical dependency relationship between them. By connecting these relationships in series, form the sales constraint chain to comprehensively and systematically describe various constraints and restrictions in the commodity sales process.

[0048] In step S113, the target selling feature information is determined based on the actual selling feature information and the actual selling feature information.

[0049] In the embodiment of the present disclosure, based on the actual selling feature information obtained in step S111 and the sales constraint chain obtained in step S112, these information are further integrated and analyzed to determine the target selling feature information that has a key impact on the prediction of commodity sales volume and risk warning. These information will be used as an important input for subsequent steps to improve the accuracy of sales volume prediction and the effectiveness of risk warning.

[0050] In the above solution, the term recognition layer identifies the core concepts in the policy document (such as "validity period of the monopoly license", "ordering quota") through a pre-trained industry dictionary, and uses a bidirectional recurrent neural network to capture the implicit restrictive conditions in the context (such as the actual execution standard of fuzzy expressions such as "in principle"); the relationship extraction layer establishes the logical dependence relationship between policy clauses. For example, when policy A stipulates that "sales are prohibited within 100 meters around the school", it is automatically associated with the "definition standard of the surrounding area of the campus" in policy B to form a complete constraint chain.

[0051] In a possible implementation manner, the method further includes: in response to obtaining a new sales guidance document, determining sensitive sales parameters from the original sales guidance document; Wherein, the new sales guidance document: a regulation or policy document newly issued by the government or relevant departments that has a direct impact on commodity sales. The original sales guidance document: the existing set of sales guidance documents on which the current policy analysis module is based. Sensitive sales guidance parameters: key parameters or clauses in the sales guidance that have a significant impact on commodity sales and need to be specially protected to prevent being arbitrarily changed, such as sales age restrictions, sales bans in specific areas, etc.

[0052] In the embodiment of the present disclosure, 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 algorithms, the clauses in the original policy document that may change or need special attention in the new policy document are identified, and the parameters corresponding to these clauses are the 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 of the new policy document during the subsequent incremental training process.

[0053] By protecting sensitive neurons, the adjustment range of the sensitive sales parameters is restricted during incremental training to prevent the new sales guidance document from overwriting the sensitive sales parameters; Among them, protecting sensitive neurons: In a deep learning model (such as a neural network in a policy analysis module), specific technical means (such as weight locking, regularization, etc.) are used to protect the neurons that process sensitive sales guidance parameters to prevent them from being over-adjusted during training. Incremental training: Based on an existing model, new data is used for training to update the model parameters and improve the model's adaptability to new data. Adjustment range: The degree of change of the model parameters (including neuron weights) during training.

[0054] In the embodiments of the present disclosure, when performing incremental training, in order to protect sensitive sales guidance parameters from being overwritten by inappropriate adjustments in new policy documents, a technique for protecting sensitive neurons is adopted. Specifically, the neurons that process these sensitive parameters will be specially marked or processed, such as locking their weights, applying a large regularization term, etc., so as to limit their adjustment range during training. 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.

[0055] In the neural network of the policy analysis module, the neurons that process sensitive sales guidance parameters are marked as sensitive neurons. During incremental training, their adjustment range will be limited by locking the weights of these neurons or applying a regularization term. In this way, even if the new sales guidance document makes a fine adjustment to the age limit, it can ensure that the original age limit is not completely overwritten and the continuity and stability of the policy are maintained.

[0056] According to the new sales guidance document, incremental training is performed on the policy analysis module to obtain the new policy analysis module.

[0057] In the embodiments of the present disclosure, after determining the sensitive sales guidance parameters and protecting the sensitive neurons, incremental training is performed on the policy analysis module according to the new sales guidance document. During the training process, new policy document data is used to update the model parameters, but at the same time, it is ensured that the sensitive sales guidance parameters are not over-adjusted. Through incremental training, the policy analysis module can learn new policy regulations and terms, and at the same time maintain stable processing of the original sensitive parameters, so as to obtain a new policy analysis module that better adapts to the current sales guidance environment.

[0058] In a possible implementation manner, in step S12, the determining the spatio-temporal association information of the store according to the special time description information, the promotion time information, the store address information, and the promotion clause information includes: In step S121, the time description in the promotion time information is converted into a calendar format, and the text semantics of the time description is bound to a date range to determine the promotion time window; Among them, promotion time information: Information describing the specific time arrangement of a product promotion activity, which may exist in text form, such as "during the National Day", "every Friday afternoon", etc. Calendarization conversion: The process of converting a time description in text form into a specific date range or time window. Promotion time window: The clear time range obtained after calendarization conversion, used to represent the specific execution period of the promotion activity.

[0059] In the embodiments of the present disclosure, when processing promotion time information, it is first necessary to parse and understand the time description in text form. Through natural language processing technology, key elements in the time description, such as festival names, days of the week, time periods, etc., are identified and matched with the calendar for conversion, so as to obtain a specific date range or time window, that is, the promotion time window.

[0060] In step S122, during vector retrieval, according to the time validity corresponding to the promotion time window, synchronously constrain the similarity between the special time description information and the promotion time window, and determine the target promotion clause information from the promotion clause information; Among them, special time description information: Other time-related description information except promotion time information, such as holidays, special events, etc., which may have an impact on sales. Vector retrieval: Converting text information into vector form for similarity calculation and retrieval in the vector space. Time validity: The property describing whether a specific time description information is valid or influential within a specific time window. Target promotion clause information: The promotion clause information that is most relevant to the current time window and most likely to affect sales, screened from the promotion clause information under the premise of considering time validity.

[0061] In the embodiments of the present disclosure, during vector retrieval, not only the promotion time window itself is considered, but also the special time description information is combined to evaluate the time validity of this information within the promotion time window. By calculating the similarity between the special time description information and the promotion time window and considering time validity as a constraint condition, it is possible to more accurately screen out the promotion clause that is most likely to have an impact within the current time window from the promotion clause information, that is, the target promotion clause information.

[0062] In step S123, according to the geographical location encoder, perform vector conversion on the store address information, and convert the store address into a business district feature vector, where the business district feature vector includes at least one of the following: school density, distance to transportation hubs; Among them, geographical location encoder: A tool or model that converts geographical location information (such as an address) into a vector representation for processing and analysis in the vector space. Business district feature vector: A multi-dimensional vector describing the characteristics of the business district where the store is located, which may include multiple dimensions such as school density, distance to transportation hubs, population density, and richness of commercial facilities.

[0063] In the embodiments of the present disclosure, a geographical location encoder is used to convert the address information of a store into a vector form, i.e., a business district feature vector. In this process, the geographical location encoder will consider multiple aspects of the address, such as geographical location coordinates, surrounding environment features, etc., and map them into a multi-dimensional vector. Each dimension in the business district feature vector represents a specific business district feature, such as school density, distance to transportation hubs, etc. These features help to more comprehensively describe the business district environment where the store is located.

[0064] In step S124, the spatio-temporal association information of the store is determined by associating the promotion time window, the target promotion clause information, and the business district feature vector.

[0065] In the embodiments of the present disclosure, after obtaining the promotion time window, the target promotion clause information, and the business district feature vector, these information are associated and integrated to form the spatio-temporal association information of the store. In this process, the interaction and influence between time, promotion, and space factors are considered, so as to more comprehensively describe the sales characteristics and potential risks of the store under different time and space conditions.

[0066] Compared with the traditional knowledge base that stores text and numerical data independently, the above technical solution adopts a spatio-temporal vector indexing technology: calendarize the time description in the promotion rules (such as "during holidays"), bind the text semantics to a specific date range, and synchronously consider text similarity and time validity during vector retrieval. For example, only return the promotion clauses available within the current time window, establish a geographical location encoder, and convert the store address into a business district feature vector (such as school density, distance to transportation hubs, etc.) to achieve intelligent matching in the spatial dimension.

[0067] In a possible implementation manner, the method further includes: obtaining the actual commodity order information of the store for the predicted commodity order information, and determining whether the trigger update condition is satisfied according to the actual commodity order information and the predicted commodity order information; In the embodiments of the present disclosure, first obtain the actual commodity order information from the store's orders, and then compare it with the predicted commodity order information. By calculating the differences between the two (such as order quantity differences, order type differences, etc.), determine whether these differences exceed the preset trigger update condition threshold. If the differences exceed the threshold, it is considered that the current large model may not be able to accurately predict the store's order demand, and it is updated.

[0068] When the trigger update condition is satisfied, analyze the update policy information of the sales guidance document through text similarity and / or monitor the sudden change information of the pedestrian flow in the business district where the store address is located; Among them, the updated policy information: newly released or modified content in the sales guidance document may affect the sales and ordering of goods. The sudden change information of the pedestrian flow in the business district: the situation where the pedestrian flow in the business district where the store is located suddenly increases or decreases, which may be caused by factors such as holidays and special events, and has a direct impact on the sales of goods. Text similarity analysis: By calculating the similarity between texts, the degree of difference between the updated policy information and the existing policies is judged.

[0069] In the embodiments of the present disclosure, when the update trigger condition is met, one or both of two strategies will be adopted to obtain the update information. On the one hand, the latest sales guidance document will be analyzed, and through text similarity analysis technology, the differences between the updated policy information and the existing policies will be identified, and these differences may involve sales restrictions, tax policies, promotional activities, etc. On the other hand, the change in the pedestrian flow in the business district where the store address is located will be monitored, especially the sudden change information of the pedestrian flow, which may be caused by factors such as holidays, large-scale events, and weather changes. By analyzing this information, the impact of changes in the external environment on the sales of goods can be understood.

[0070] Update the large model according to the updated policy information and / or the sudden change information of the pedestrian flow to obtain a new large model.

[0071] In the embodiments of the present disclosure, after obtaining the updated policy information and / or the sudden change information of the pedestrian flow, these information will be used as the basis for new training data or adjusting parameters to update the large model. The update process may involve adjusting the weights of the model, adding new feature dimensions, modifying the loss function of the model, etc. Through the update, the large model can better adapt to the changes in the external environment, improve the prediction accuracy of the ordering demand of goods, and the analysis ability of information such as sales guidance and pedestrian flow in the business district.

[0072] Compared with the traditional model update that will forget old knowledge, a business strategy memory library is introduced to regularly store typical decision scenarios (such as the Spring Festival stocking strategy and the support plan for the opening of new stores). When the model is updated, the historical strategies are weighted and replayed according to the business importance to ensure the stability of the core business logic. Set the priority protection level of the regulations, and set the sampling weight 5 times that of the conventional strategy for the decision-making mode involving legal compliance.

[0073] Establish a multi-signal linkage evaluation system to accurately judge the timing of model update: Business indicator monitoring: When the inventory deviation rate > 20% for 3 consecutive days and the prediction accuracy < 85%, an early warning is triggered; Policy change detection: Identify major updates to the policy library through text similarity analysis; Environmental change perception: Monitor sudden changes in the pedestrian flow data in the business district (such as the opening of a new subway station leading to changes in the customer group). If any of the above is met, incremental learning is started.

[0074] In a possible implementation, in step S15, the visual display of the target display information includes: In step S151, when the predicted commodity order information output by the large model suggests reducing the commodity order quantity of the store, display the contribution degree of at least one of the target sales characteristic information, the spatio-temporal correlation information, the meteorological warning information, and the store historical sales information; In the embodiments of the present disclosure, interpretability techniques of machine learning models, such as SHAP (SHapley Additive exPlanations) values, LIME (Local Interpretable Model-agnostic Explanations), etc., are used to calculate the contribution degree of each input feature to the prediction result. Contribution degree sorting and display: According to the calculated contribution degree, sort each factor in the target sales characteristic information (such as commodity price, brand preference), spatio-temporal correlation information (such as business district pedestrian flow, holidays), meteorological warning information (such as extreme weather warning), and store historical sales information (such as sales trend, seasonal fluctuation), and select several factors with the highest contribution degree for visual display.

[0075] In step S152, label and display the policy items that trigger the suggestion to reduce the commodity order quantity of the store; In the embodiments of the present disclosure, text mining is performed on the sales guidance document to extract policy items related to commodity sales, order quantity, etc. Policy and prediction association: Analyze the extracted policy items to determine which items are directly related to the current prediction of reducing the order quantity. Labeling and display: Label these relevant policy items in the visual interface and possibly attach specific policy content or links for users to further understand.

[0076] In step S153, highlight the text fragments concerned by the large model, and determine the target verification information corresponding to the text fragments from the target sales characteristic information, the spatio-temporal correlation information, the meteorological warning information, and the store historical sales information; In the embodiments of the present disclosure, the text fragments that the large model particularly concerns during the prediction process are identified, and the target verification information corresponding to these fragments is determined from the relevant information. This is usually achieved through the following methods: Text segment recognition: Using natural language processing techniques, such as keyword extraction and named entity recognition, to identify the text segments that the large model pays special attention to during the prediction process. Target verification information determination: Based on the identified text segments, determine the relevant target verification information from the target sales feature information, spatio-temporal association information, meteorological warning information, and store historical sales information. Highlight display: Highlight these text segments in the visualization interface and attach the corresponding target verification information so that users can understand how the model makes predictions based on this information.

[0077] In step S154, calculate and display the consistency index between the attention weight corresponding to the target verification information and the judgment of human experts.

[0078] In the embodiments of the present disclosure, calculate the consistency index between the attention weight corresponding to the target verification information and the judgment of human experts to evaluate the prediction performance of the model. Attention weight calculation: Using the interpretability technology of machine learning models, calculate the attention weight of the target verification information during the model prediction process. Collection of human expert judgments: Invite human 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 human experts. This can be achieved by calculating correlation coefficients (such as Pearson correlation coefficients), classification accuracies, etc. Display: Display the calculated consistency index to users so that they can evaluate the prediction performance of the model.

[0079] Develop a clause mapping interpreter for the above technical solutions to achieve the docking of technical interpretation and business understanding: When the model recommends reducing the order quantity of a certain store, not only display the contribution degree of each feature, but also mark the specific policy items triggered; highlight the text segments concerned by the model (such as restrictive clauses in policy documents), and calculate the consistency index between its attention weight and the judgment of human experts.

[0080] In a possible implementation manner, the method further includes: Obtain the abnormal sales information in the store historical sales information that indicates that the actual sales information of the store meets the preset exclusion conditions, and exclude the store historical sales information corresponding to the abnormal sales information; Among them, abnormal sales information: In the store historical sales information, sales data that deviates significantly from the normal sales mode, which may be caused by data entry errors, special events (such as extremely popular promotion activities), or other abnormal factors. Preset exclusion conditions: The criteria or thresholds used to judge whether the sales information is abnormal, such as a sudden increase or decrease in sales volume, abnormal sales price, etc.

[0081] In the embodiments of the present disclosure, first, the historical sales information of the store is traversed, and the abnormal sales information is filtered out according to the preset elimination conditions. These conditions may be set based on statistical methods (such as Z-score, IQR, etc.) or business rules (such as the single-day sales volume exceeding 3 times the historical average). Once the abnormal sales information is identified, it will be eliminated from the historical sales information to ensure that the accuracy of subsequent analysis (such as predicting the commodity order information) is not affected by these outliers.

[0082] It is determined that there is new target information in the target selling characteristic information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information, and the deviation degree of the new target information from the historical distribution is judged through adversarial training. In the case where the deviation degree does not meet the preset deviation degree threshold, the new target information is eliminated. In the case where the new target information indicates the addition of the target selling characteristic information, the target selling characteristic information before and after the addition is marked and isolated. Among them, new target information: data that newly appears or is updated in the target selling characteristic information, spatio-temporal association information, meteorological warning information, and store historical sales information. Adversarial training: A machine learning technique that evaluates the robustness of a model or judges the distribution difference between new data and historical data by simulating an adversarial environment. Deviation degree: An index that measures the degree of difference between new target information and the historical data distribution. Preset deviation degree threshold: A standard or threshold used to judge whether new target information is acceptable. Marking and isolation: Marking the newly added target selling characteristic information and separating it from the original characteristic information for subsequent analysis and comparison.

[0083] In the embodiments of the present disclosure, it is regularly checked whether there is new target information in the target selling characteristic information, spatio-temporal association information, meteorological warning information, and store historical sales information. For this newly added information, its deviation degree from the historical data is calculated through adversarial training or other distribution difference evaluation methods. If the deviation degree exceeds the preset threshold, it is considered that this newly added information may not be representative or there is an abnormality, and thus it is eliminated. In particular, if the new target information involves the update of the target selling characteristic information, the characteristic information before and after the addition is marked and isolated so as to separately analyze the influence of these changes on the prediction result subsequently.

[0084] Automatically check whether the predicted commodity order information conforms to the preset industry rules, and re-execute the prediction of the predicted commodity order information in the case where the predicted commodity order information does not conform to the preset industry rules.

[0085] In the embodiments of the present disclosure, compliance checks are performed on the predicted commodity order information output by the large model to ensure that it complies with preset industry rules. These rules may involve the rationality of the order quantity (such as not exceeding the inventory capacity and meeting the minimum order quantity requirements), the compliance of the order time (such as avoiding holidays and conforming to the supplier's delivery cycle), etc. If the predicted commodity order information does not conform to these rules, the prediction process will be re-executed, and the prediction results may be optimized by adjusting model parameters, adding new features or constraints, etc., until the generated predicted commodity order information conforms to the preset industry rules.

[0086] Outlier filtering: Eliminate abnormal records such as zero sales volume caused by system failures; Distribution consistency detection: Judge the deviation degree between new data and historical distribution through adversarial training; Business logic verification: Automatically check whether the data conforms to industry rules (such as the single order quantity does not exceed the license limit); Timeliness grading: Mark and isolate data before and after policy changes to avoid confusion between old and new rules.

[0087] Combined with the above embodiments, an example is used for illustration: Data and knowledge base construction, Unstructured text processing of double-layer semantic parsing: Term recognition layer: Based on a pre-trained recognition model of industry-specific terms, identify the core concepts and implicit constraints (such as "monopoly license", "order quota") in the policy text, and capture the ambiguous expressions in the policy through a bidirectional recurrent neural network. Relationship extraction layer: Automatically analyze the logical associations between policy clauses, construct a policy constraint chain, and ensure the integrity and coherence of policy information.

[0088] Knowledge base index of spatio-temporal association: Time semantic index: Convert the time information involved in the policy text into specific calendar intervals, and establish a timeliness semantic vector index. Spatial location index: Use geographic location coding technology to convert store address features (such as business district level, distance from transportation hubs) into spatial feature vectors, and realize the intelligent matching of policy information with the geographical dimension.

[0089] Structured data processing: Summarize 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 prediction.

[0090] Furthermore, adopt the dynamic fine-tuning mechanism of the large model, and the construction and use of the fine-tuning dataset: The fine-tuning dataset includes historical order quantities, sales volumes, inventory data, and store features (geographical location, grade, etc.), as well as the semantic parsing results related to the policy text.

[0091] Difference from the knowledge base: The knowledge base stores parsed policy knowledge and structured historical data, while the fine-tuning dataset is specific instances selected from the knowledge base for model training.

[0092] Specific steps for instruction fine-tuning: Obtain historical business data and policy texts, and create triples of "instruction-context-expected answer", such as: "Predict the order quantity this week based on the data of Store A in the past four weeks" as the instruction, store information and policy conditions as the context, and the expected order quantity as the answer. The triples are generated through manual annotation or automatically to ensure that the data covers a variety of business scenarios. The triples are used to guide the model to learn the prediction and decision-making logic in specific task scenarios, which is different from traditional supervised learning methods.

[0093] Thaw and fine-tune: Thaw specific layers of the large model (such as the last few layers of Transformer), and adjust the parameters of these layers to adapt to the business data; the parameters are the weights and biases of the neural network. Business data includes store sales data, policy-related data, and environmental data. Adopt multiple rounds of training, randomly shuffle the data in each round, and iteratively optimize the parameters until the evaluation metrics reach the optimal.

[0094] Perform evaluation and hyperparameter tuning: Evaluate the model performance on the validation set (comparison between real order data and model prediction results), using metrics such as mean squared error (MSE), mean absolute error (MAE), etc.

[0095] Adjust hyperparameters such as the number of fine-tuning rounds, learning rate, batch size, etc., until the performance of the validation set reaches the optimal. Compared with the existing technology, a special evaluation for the impact of policy texts is added.

[0096] After fine-tuning, the large model adopts LoRA or full-scale fine-tuning (where LoRA is the most important because of its high computational efficiency and significant effect), so that the large model can understand industry terms, predict order quantities, and explain the key reasons. The interpretability comes from the built-in attention mechanism in the large model.

[0097] Online prediction process and multi-dimensional evidence fusion reasoning: Real-time retrieval and context construction: Input example: "Store A is located in the core business district, the sales volume of Brand B cigarettes has increased week by week in the past four weeks, and a new promotion policy has been recently released", retrieve the corresponding policy text and historical sales data to form a complete context. Three-layer evidence fusion reasoning: The large model adopts a multi-head attention mechanism to calculate the attention weights of structured evidence, policy evidence, and environmental evidence respectively, and fuses and calculates the final prediction according to the weights.

[0098] Interpretability of model prediction and business post-processing: Interpretability output mechanism: Provide natural language or visual explanations of the prediction results, clearly mark the triggered policy clauses, such as "According to the regulations, the order quantity of this store has decreased". Highlight the key policy text fragments in the decision-making process, and provide auxiliary information such as feature contribution degree and attention weight to help intuitively understand the prediction basis.

[0099] Flexible business post - processing rules: Formulate business rules for the predicted output, such as setting upper and lower limits of predicted values, ordering strategies for holidays, to ensure that the prediction results adapt to specific business scenarios.

[0100] Incremental learning of business feedback loop: Actual execution data feedback mechanism: Regularly collect actual order execution data and the "reasonable order quantity" adjusted by business personnel, mark the differences between prediction and actual execution, and feedback for incremental training.

[0101] Policy memory bank and replay mechanism: Store and mark historical typical decision scenarios (such as promotional activities, opening of new stores), conduct weighted policy replay to ensure the continuity and stability of key business policies. Set high - weight sampling for decision terms sensitive to regulations to avoid forgetting of regulation - sensitive knowledge caused by incremental learning.

[0102] Fast expansion and adaptation mechanism: Fast model adaptation; Adopt lightweight fine - tuning techniques (such as LoRA, Prefix - Tuning, Adapter) to achieve fast model adaptation for new stores, new product specifications, and new policies.

[0103] Modular knowledge base update: Adopt a modular knowledge base storage structure 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.

[0104] Compared with existing order prediction technologies based on a single data source or traditional statistical / machine learning methods, by introducing large models and knowledge base retrieval mechanisms, multi - dimensional deep learning and dynamic understanding of order behavior have been achieved. Specifically, it is reflected in the following aspects: Improved prediction accuracy: Based on the integration of multi - source data such as historical orders, sales, inventory, store grades, and industry policies, the large model can capture more complex time - series features and store differences through deep learning and fine - tuning, effectively reducing prediction errors and reducing the risks of "stock - out" and "overstock".

[0105] Enhanced business decision - making flexibility: Relying on knowledge base retrieval, key text information such as promotional rules and regulatory restrictions can be automatically retrieved before prediction, further assisting the large model to make reasonable evaluations for specific stores and special situations (holidays, seasonal fluctuations, sudden activities, etc.), and providing diversified ordering plans that can be adjusted according to the actual situation for management.

[0106] Interpretability and business trust: While giving prediction results, the large model can clarify the main influencing factors in natural language or visual form, improve the transparency and trust in the prediction process, and assist business personnel in making quick decisions and adjusting business strategies.

[0107] Continuous Iteration Ability: By collecting actual order and sales feedback after each business execution, incremental fine-tuning or retraining can be carried out to enable the model to continuously adapt to market changes and the individual needs of stores, forming a closed-loop management mode and maintaining the continuous optimization of prediction performance.

[0108] Deployment and Expansion Efficiency: Fine-tuning based on pre-trained models significantly reduces the dependence on large-scale training data. When new stores, brands, or policies are added, it can be quickly expanded on the original system to achieve flexible adaptation to retail networks of different scales.

[0109] In summary, it has significant advantages in aspects such as high-precision prediction, interpretability, business agility, and continuous iterative upgrade, and can effectively improve the order management efficiency and the overall competitiveness of the supply chain.

[0110] In one embodiment, as Figure 2 shown, a commodity order prediction device based on a large model is provided, including: A sales information determination module 210, configured to determine target sales feature information from the sales guidance document of the commodity through the policy analysis module in the large model according to the industry dictionary pre-trained for the industry where the commodity is located; A spatio-temporal association module 220, configured to obtain special time description information, promotion time information of the order-taking store, store address information, and promotion clause information through the spatio-temporal association module in the large model, and determine the spatio-temporal association information of the store according to the special time description information, the promotion time information, the store address information, and the promotion clause information; An information acquisition module 230, configured to obtain meteorological warning information and store historical sales information through the multi-dimensional data acquisition module in the large model; A prediction module 240, configured to determine the predicted commodity order information of the store by dynamically adjusting the weights through the attention weight distribution mechanism according to the target sales feature information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information through the sales volume prediction module in the large model; A visualization display module 250, configured to visually display the target display information through the risk warning module in the large model when the predicted commodity order information meets the preset conditions.

[0111] In a possible implementation manner, the sales information determination module 210 is configured to: Through the policy analysis module in the large model, obtain the corresponding sales feature information according to the industry dictionary pre-trained for the industry to which the commodity belongs, and use the bidirectional recurrent neural network in the double-layer semantic parsing engine to determine the implicit restrictive conditions from the sales feature information, so as to obtain the actual sales feature information of the commodity in terms of sales guidance; Adopt the relationship extraction layer in the double-layer semantic parsing engine to establish the logical dependency relationship between policy clauses, and obtain the sales constraint chain; Determine the target sales feature information according to the actual sales feature information and the actual sales feature information.

[0112] In a possible implementation manner, the sales information determination module 210 is configured to: In response to obtaining a new sales guidance document, determine sensitive sales parameters from the original sales guidance document; By protecting sensitive neurons, limit the adjustment range of the sensitive sales parameters during incremental training to prevent the new sales guidance document from overwriting the sensitive sales parameters; According to the new sales guidance document, perform incremental training on the policy analysis module to obtain a new policy analysis module.

[0113] In a possible implementation manner, the spatio-temporal association module 220 is configured to: Perform calendar conversion on the time description in the promotion time information, bind the text semantics of the time description to the date range, and determine the promotion time window; During vector retrieval, according to the time validity corresponding to the promotion time window, synchronously constrain the similarity between the special time description information and the promotion time window, and determine the target promotion clause information from the promotion clause information; According to the geographical location encoder, perform vector conversion on the store address information, and convert the store address into a business district feature vector, where the business district feature vector includes at least one of the following: school density, distance to transportation hub; Associate the promotion time window, the target promotion clause information, and the business district feature vector to determine the spatio-temporal association information of the store.

[0114] In a possible implementation manner, the device further includes an update module, which is configured to: Obtain the actual commodity order information of the store for the predicted commodity order information, and determine whether the trigger update condition is met according to the actual commodity order information and the predicted commodity order information; When the trigger update condition is met, analyze the update policy information of the sales guidance document through text similarity and / or monitor the sudden change information of the pedestrian flow in the business district where the store address is located; Update the large model according to the update policy information and / or the sudden change information of the pedestrian flow to obtain a new large model.

[0115] In a possible implementation manner, the visualization display module 250 is configured to: When the predicted commodity order information output by the large model suggests reducing the commodity order quantity of the store, display the contribution degree of at least one of the target sales characteristic information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information; Mark and display the policy items that trigger the suggestion to reduce the commodity order quantity of the store; Highlight the text fragments concerned by the large model, and determine the target verification information corresponding to the text fragments from the target sales characteristic information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information; Calculate and display the consistency index between the attention weight corresponding to the target verification information and the judgment of human experts.

[0116] In a possible implementation manner, the device further includes: a data processing module, configured to: Obtain the abnormal sales information in the store historical sales information that represents the actual sales information of the store meeting the preset exclusion condition, and exclude the store historical sales information corresponding to the abnormal sales information; Determine that there is new target information in the target sales characteristic information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information, and judge the deviation degree of the new target from the historical distribution through adversarial training. When the deviation degree does not meet the preset deviation threshold, exclude the new target information, and when the new target information represents the addition of the target sales characteristic information, mark and isolate the target sales characteristic information before and after the addition; Automatically check whether the predicted commodity order information conforms to the preset industry rules, and re-perform the prediction of the predicted commodity order information when the predicted commodity order information does not conform to the preset industry rules.

[0117] For the specific limitations of a commodity order prediction device based on a large model, reference may be made to the limitations of a commodity order prediction method based on a large model in the foregoing text, which will not be elaborated herein. Each module in the above-mentioned commodity order prediction device based on a large model can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in the processor in the computer device in hardware form or be independent of it, or stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to each of the above modules.

[0118] An embodiment of the present disclosure also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of the method described in any one of the foregoing embodiments are implemented.

[0119] An embodiment of the present disclosure also provides an electronic device, including: A memory, on which a computer program is stored; A processor, configured to execute the computer program in the memory to implement the steps of the method described in any one of the foregoing embodiments.

[0120] Figure 3 The commodity order prediction device 100 based on a large model shown includes: a processor 1001 and a memory 1003. Among them, the processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the commodity order prediction device 100 based on a large model may further include a communication component, and the communication component can be used for data interaction between the device 100 and other devices, such as sending or receiving data, etc. It should be noted that in actual scheduling, the communication component is not limited to one, and the structure of the commodity order prediction device 100 based on a large model does not constitute a limitation to the embodiments of the present application.

[0121] The processor 1001 can be a CPU (Central Processing Unit, central processor), a general-purpose processor, a DSP (Digital Signal Processor, digital signal processor), an ASIC (Application Specific Integrated Circuit, application-specific integrated circuit), an FPGA (Field Programmable Gate Array, field programmable gate array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute various exemplary logical blocks, modules, and circuits described in connection with the disclosure of the present application. The processor 1001 can also be a combination that realizes computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0122] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 1002 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0123] The memory 1003 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices 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 compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, 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.

[0124] 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 above-mentioned commodity order forecasting method based on a large model.

[0125] The 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.

[0126] 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, substitutions and variations may be made to these embodiments, and these changes, modifications, substitutions and variations all fall within the protection scope of the present disclosure.

[0127] In addition, it should be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and the same should be regarded as the content disclosed in the present disclosure. To avoid unnecessary repetition, the present disclosure will not separately describe various possible combination methods. The technical scope of this application is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A commodity order prediction method based on a large model, characterized in that The method includes: Through the policy analysis module in the large model, according to the industry dictionary pre-trained for the industry where the commodity is located, determine the target selling feature information from the sales guidance document of the commodity; Through the spatio-temporal correlation module in the large model, obtain special time description information, the promotion time information of the ordering store, the store address information, and the promotion clause information, and determine the spatio-temporal correlation information of the store according to the special time description information, the promotion time information, the store address information, and the promotion clause information; Through the multi-dimensional data acquisition module in the large model, obtain meteorological warning information and the historical sales information of the store; Through the sales volume prediction module of the large model, according to the target selling feature information, the spatio-temporal correlation information, the meteorological warning information, and the historical sales information of the store, dynamically adjust the weights through the attention weight distribution mechanism, and determine the predicted commodity ordering information of the store; Through the risk warning module of the large model, when the predicted commodity ordering information meets the preset conditions, visually display the target display information.

2. The method according to claim 1, wherein The step of determining the target selling feature information from the sales guidance document of the commodity through the policy analysis module in the large model according to the industry dictionary pre-trained for the industry where the commodity is located includes: Through the policy analysis module in the large model, according to the industry dictionary pre-trained for the industry where the commodity is located, obtain the corresponding selling feature information, and use the bidirectional recurrent neural network in the double-layer semantic parsing engine to determine the implicit restrictive conditions from the selling feature information, and obtain the actual selling feature information of the commodity in the sales guidance; Use the relationship extraction layer in the double-layer semantic parsing engine to establish the logical dependence relationship between policy clauses and obtain the sales constraint chain; Determine the target selling feature information according to the actual selling feature information and the actual selling feature information.

3. The method according to claim 2, wherein The method further includes: In response to obtaining a new sales guidance document, determine the sensitive sales parameters from the original sales guidance document; By protecting sensitive neurons, limit the adjustment range of the sensitive sales parameters during incremental training to prevent the new sales guidance document from overwriting the sensitive sales parameters; According to the new sales guidance document, perform incremental training on the policy analysis module to obtain a new policy analysis module.

4. The method according to claim 1, characterized in that, The step of determining the spatio-temporal correlation information of the store according to the special time description information, the promotion time information, the store address information, and the promotion clause information includes: Perform calendar conversion on the time description in the promotion time information, bind the text semantics of the time description to the date interval, and determine the promotion time window; During vector retrieval, according to the time validity corresponding to the promotion time window, synchronously constrain the similarity between the special time description information and the promotion time window, and determine the target promotion clause information from the promotion clause information; According to the geographical location encoder, perform vector conversion on the store address information, and convert the store address into a business district feature vector, where the business district feature vector includes at least one of the following: school density, distance to transportation hubs; Associate according to the promotion time window, the target promotion clause information, and the business district feature vector to determine the spatio-temporal association information of the store.

5. The method according to any one of claims 1-4, characterized in that, The method further includes: Obtain the actual product order information of the store for the predicted product order information, and determine whether the trigger update condition is met according to the actual product order information and the predicted product order information; When the trigger update condition is met, analyze the update policy information of the sales guidance document through text similarity and / or monitor the sudden change information of the pedestrian flow in the business district where the store address is located; Update the large model according to the update policy information and / or the pedestrian flow sudden change information to obtain a new large model.

6. The method according to any one of claims 1-4, characterized in that The visual display of the target display information includes: When the predicted product order information output by the large model recommends reducing the product order quantity of the store, display the contribution degree of at least one of the target sales feature information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information; Mark and display the policy items that trigger the recommendation to reduce the product order quantity of the store; Highlight the text fragments concerned by the large model, and determine the target verification information corresponding to the text fragments from the target sales feature information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information; Calculate and display the consistency index between the attention weight corresponding to the target verification information and the judgment of artificial experts.

7. The method according to any one of claims 1-4, characterized in that The method further includes: Obtain the abnormal sales information in the store historical sales information that represents the actual sales information of the store meeting the preset exclusion conditions, and exclude the store historical sales information corresponding to the abnormal sales information; Determine that there is new target information in the target sales feature information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information, and judge the deviation degree of the new target from the historical distribution through adversarial training. When the deviation degree does not meet the preset deviation threshold, exclude the new target information. When the new target information represents the addition of the target sales feature information, label and isolate the target sales feature information before and after the addition; Automatically check whether the predicted product order information complies with the preset industry rules, and when the predicted product order information does not comply with the preset industry rules, re-perform the prediction of the predicted product order information.

8. A commodity order prediction device based on a large model, characterized in that, The device includes: A sales information determination module, configured to determine target sales feature information from the sales guidance document of the product through the policy analysis module in the large model according to the industry dictionary pre-trained in the industry where the product is located; A spatio-temporal association module, configured to obtain special time description information, promotion time information of an ordering store, store address information, and promotion term information through the spatio-temporal association module in the large model, and determine the spatio-temporal association information of the store according to the special time description information, the promotion time information, the store address information, and the promotion term information; An information acquisition module, configured to obtain meteorological warning information and store historical sales information through the multi-dimensional data acquisition module in the large model; A prediction module, configured to determine the predicted commodity ordering information of the store through the sales volume prediction module of the large model, dynamically adjust weights through an attention weight distribution mechanism according to the target selling characteristic information, the spatio-temporal association information, the meteorological warning information, and the store historical sales information; A visualization display module, configured to perform visualization display on target display information through the risk warning module of the large model when the predicted commodity ordering information meets a preset condition.

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

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

Images