Multi-strategy dynamic prediction method, system and device based on seasonal commodities and storage medium

Abstract

The invention relates to the technical field of large model prediction, and discloses a multi-strategy dynamic prediction method, system and device based on seasonal commodities and a storage medium. The multi-strategy dynamic prediction method is applied to an electronic prediction device, and specifically comprises the following steps: S101, obtaining multi-dimensional data information published by a historical network, the multi-dimensional data information carrying regional business data, marketing activity data, date feature data, commodity attribute data and historical operation data, and expanded seasonal climate characteristic data; and S102, preprocessing the acquired multi-dimensional data information, acquiring stock and stockout condition characteristics of historical and year-on-year sales promotion periods, performing identification based on the stock and stockout condition characteristics, completing historical sales data compensation, and acquiring historical sales compensation data. The method breaks through the limitation of a Gregorian calendar fixed period through the coupling of a lunar calendar sliding window and a trigonometric function, and improves the prediction accuracy compared with a conventional LSTM model.

Description

Technical Field

[0001] The present invention relates to the technical field of large-scale model prediction, and in particular to a multi-strategy dynamic prediction method, system, device and storage medium based on seasonal commodities. Background Art

[0002] With the booming development of the internet and the rapid increase in the informatization of society, the emergence of new technologies has increased the information intake of various social groups before consumer behavior. With the rapid rise of e-commerce, offline store sales are subject to regional restrictions, express delivery area restrictions, and artificial and relatively fixed promotion models. Seasonal products are particularly sensitive to promotional prices, promotion timing, and inventory management.

[0003] Enabling dynamic intelligent forecasting and replenishment methods at the right time can better avoid problems such as overstocking or dumping of out-of-season goods and untimely replenishment of popular goods. It can adaptively reduce inventory backlogs and ensure sufficient supply to meet consumer demand during seasonal peaks. Summary of the Invention

[0004] The purpose of the present invention is to provide a multi-strategy dynamic forecasting method, system, device and storage medium based on seasonal commodities. The intelligent historical sales compensation mechanism significantly corrects the data deviation caused by supply chain disruptions through out-of-stock period identification and counterfactual modeling. Combined with the dynamic attenuation factor design, it not only avoids the forecast distortion caused by over-compensation, but also ensures that the compensation amount is dynamically adjusted with the inventory turnover rate, aiming to solve the problems in the existing technology.

[0005] The present invention is implemented by applying a multi-strategy dynamic forecasting method based on seasonal commodities to an electronic forecasting device, specifically comprising the following steps: S101: Acquire multi-dimensional data information published on historical networks, wherein the multi-dimensional data information carries regional business data, marketing activity data, date feature data, product attribute data, historical business data, and extended seasonal climate feature data; S102: Preprocess the acquired multi-dimensional data information to obtain historical and year-on-year inventory out-of-stock characteristics during promotional periods, identify the inventory out-of-stock characteristics, complete historical sales data compensation, and obtain historical sales compensation data; S103: Using an initial algorithm model set including a long short-term memory network and gradient boosting to complete training on the historical sales compensation data, the training obtains sales characteristics of seasonal products within a preset time period; S104: Based on the sales characteristics of seasonal products in a preset time period, the sales characteristics are placed in an initial algorithm model including a long short-term memory network and gradient boosting, and a seasonal product prediction model set is generated through continuous identification and prediction; S105: Based on the preset screening conditions, the current multi-dimensional data information is called, and the seasonal commodity forecasting model set performs dynamic forecasting on the current multi-dimensional data information to generate multi-strategy replenishment suggestions.

[0006] Furthermore, in S101, the seasonal climate characteristic data is enhanced by introducing periodic trigonometric functions, historical order demand area information, important holidays and lunar calendar information.

[0007] Furthermore, the enhancement of the seasonal climate characteristic data includes: Generate the values ​​of the benchmark periodic functions sin(2πx / 365) and cos(2πx / 365) to represent the fixed periodicity of the Gregorian calendar year. Build dynamic period adjustment parameters based on the benchmark periodic functions and introduce a sliding window mechanism to dynamically adjust the period length according to the month. Construct a feature transformation for the lunar calendar and holidays, assign dynamic weights based on holiday types, align historical data by lunar calendar date, and calculate average sales volatility before, during, and after holidays. Obtain regional-climate coupling analysis data clusters, establish independent climate response models for each data cluster, and extract features including: regional temperature-sales curves and precipitation impact thresholds; The periodic values ​​of the baseline periodic functions sin(2πx / 365) and cos(2πx / 365) are cross-multiplied with the lunar month to capture the impact of the lunar leap month on the seasonal cycle. The regional climate zone labels are then combined with the future temperature forecast to generate a regional climate stress index. The regional climate zone labels are combined with the future temperature forecast values ​​to generate a regionalized climate pressure index. The features are analyzed through SHAP values ​​to screen out features with contribution values ​​above the threshold.

[0008] Furthermore, in S102, the acquired multi-dimensional data information is preprocessed, and the preprocessing includes: Data cleaning and integration: We use the quantile threshold method to identify abnormal sales fluctuations, unify time zones for multi-source data, establish a joint primary key for product, climate, and inventory, build an entity relationship network through a graph database, and identify implicit association rules. Intelligent historical sales compensation processing: Based on the ERP system log, the inventory status of the historical period is identified. When sales suddenly drop but the sales of competing products increase during the same period, the manual review process is triggered to determine the maximum possible estimate of inventory constraints, introduce the time decay effect, and dynamically adjust the decay parameters.

[0009] Furthermore, in S103, the training of the historical sales compensation data is completed using an initial algorithm model set including a long short-term memory network and gradient boosting, including: Input historical sales compensation data to complete time series feature reconstruction, and determine whether the constructed feature grouping is a climate-sensitive product group or an inventory-dependent product group; Configure the LSTM time series base model through the model collaboration mechanism, complete the baseline prediction of the gradient boosting residual correction model, and complete dynamic weight allocation; A classification model is used to distinguish the original data from the compensated data, and the Kolmogorov-Smirnov test is applied to align the data distribution before and after compensation. If the accuracy of the classification model is less than 95%, the data distribution correction and repair is triggered until the classification model accuracy reaches 95% and the correction stops.

[0010] Furthermore, in S104, based on the sales characteristics of seasonal products in a preset time period, the sales characteristics are placed in an initial algorithm model including a long short-term memory network and gradient boosting, and a seasonal product prediction model set is generated through continuous identification and prediction, including: Perform periodic decomposition on the acquired seasonal product sales characteristics, complete multi-granularity period division and feature generation statistics, and generate interactive features; Construct a dual-channel LSTM architecture model. The main channel of the LSTM architecture model processes the original sales time series data, and the auxiliary channel learns the periodic signal generated by sin(2πx / 365) / cos(2πx / 365). Finally, it outputs the sales time series data integrated with the periodic signal. Obtain spatiotemporal cross-features, complete the climate elasticity coefficient through panel data regression calculation, adopt the dynamic switching strategy of feature importance, complete the hierarchical prediction of interactive features and LSTM architecture model, and complete the training of seasonal commodity prediction model set based on basic weighted fusion and dynamic adjustment and incremental training trigger conditions.

[0011] Compared with the existing technology, the multi-strategy dynamic forecasting method, system, device and storage medium based on seasonal products provided by the present invention have the following beneficial effects: 1. By coupling the lunar calendar sliding window with trigonometric functions, the system breaks through the constraints of the Gregorian calendar's fixed cycle, improving prediction accuracy compared to traditional LSTM models. Furthermore, the climate impact factor calculation unit establishes a nonlinear temperature-sales relationship model, successfully capturing the impact of sudden climate events. This shortens prediction response times in cold wave warning scenarios. Furthermore, the intelligent historical sales compensation mechanism significantly corrects data biases caused by supply chain disruptions through out-of-stock period identification and counterfactual modeling. Combined with a dynamic attenuation factor design, this mechanism avoids prediction distortion caused by overcompensation and ensures that the compensation amount adjusts dynamically with inventory turnover. 2. A multi-strategy dynamic generation mechanism and model fusion architecture achieve a precise balance between inventory costs and sales opportunities. The three-dimensional replenishment decision-making system optimizes the supply chain, including: a climate resilience strategy that dynamically adjusts stocking levels; a historical analogy strategy that uses the DTW algorithm to match historically similar seasons, combined with lunar calendar feature alignment technology to improve the accuracy of seasonal stocking timing for clothing products; a dynamic safety stock strategy that adjusts safety thresholds based on real-time inventory turnover; and a dynamic adjustment of model weights by continuously comparing predicted values ​​with actual sales data. The online learning mechanism shortens response time to sudden climate events, reducing prediction deviation compared to static models. 3. This technical solution focuses on a multi-strategy dynamic forecasting method based on seasonal characteristics, supply chain inventory status, and future climate conditions. It deeply integrates knowledge from multiple fields, such as trend quantification of seasonal products, historical sales reference, and optimal forecast formulation. It is suitable for promotion forecasting and replenishment management of large-scale batch seasonal products, effectively resolving the problem of seasonal product backlogs, while avoiding the dilemma of being unable to meet consumer demand due to inventory shortages, helping companies achieve efficient supply chain management and accurate market response.

[0012] A multi-strategy dynamic forecasting system based on seasonal commodities is used to implement the above-mentioned multi-strategy dynamic forecasting method, and the dynamic forecasting system includes: The data enhancement module is used to collect multi-dimensional data information published on the historical network and introduce periodic trigonometric functions, historical order and demand area information, important holidays and lunar calendar information to enhance the extended seasonal climate characteristic data; The intelligent compensation module completes the preprocessing of multi-dimensional data information and obtains the inventory out-of-stock characteristics of historical and year-on-year promotion periods, completes the compensation of historical sales data, and obtains historical sales compensation data; The model integration module is used to complete the training of historical sales compensation data using an initial algorithm model set including a long short-term memory network and gradient boosting. The sales characteristics of seasonal products in a preset time period are placed in the initial algorithm model including a long short-term memory network and gradient boosting, and a seasonal product prediction model set is generated through continuous identification and prediction; The strategy generation module is used to dynamically predict the current multi-dimensional data information using the seasonal commodity forecasting model set and generate multi-strategy replenishment recommendations.

[0013] Specifically, the data enhancement module includes: A lunar calendar feature conversion unit, used to convert a Gregorian calendar date into a lunar calendar date feature; The climate impact factor calculation unit is used to establish a nonlinear relationship model between temperature, precipitation and product sales.

[0014] A multi-strategy dynamic prediction device based on seasonal commodities includes a storage unit and a processor, wherein the storage unit is used to store a computer program, and the processor runs the computer program to enable the multi-strategy dynamic prediction device to perform the multi-strategy dynamic prediction method based on seasonal commodities.

[0015] A computer storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned multi-strategy dynamic forecasting method based on seasonal products. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic block diagram of the process of obtaining multi-dimensional data information published on the historical network in the multi-strategy dynamic prediction method based on seasonal products proposed by the present invention; Figure 2 A schematic block diagram of the process of completing historical sales data compensation and obtaining historical sales compensation data in the multi-strategy dynamic forecasting method based on seasonal products proposed by the present invention; Figure 3 This is a schematic block diagram of the process of generating multi-strategy replenishment suggestions by continuously identifying and predicting a seasonal commodity forecast model set in the multi-strategy dynamic forecasting method based on seasonal commodities proposed by the present invention; Figure 4 This is a schematic diagram of the structure of the multi-strategy dynamic forecasting system based on seasonal products proposed by the present invention; Figure 5 This is a schematic diagram of the structure of the multi-strategy dynamic forecasting device based on seasonal commodities proposed by the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention 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 invention and are not intended to limit the present invention.

[0018] The implementation of the present invention is described in detail below with reference to specific embodiments.

[0019] The same or similar numbers in the drawings of this embodiment correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "up", "down", "left", "right", etc. indicate directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. This is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0020] Reference Figure 1-3 As shown, the multi-strategy dynamic forecasting method based on seasonal commodities is applied to electronic forecasting equipment and specifically includes the following steps: S101: Acquire multi-dimensional data information published on the historical network, where the multi-dimensional data information carries regional business data, marketing activity data, date feature data, product attribute data, historical business data, and extended seasonal climate feature data; Among them, seasonal climate characteristic data is enhanced by introducing periodic trigonometric functions, historical order demand area information, important holidays and lunar calendar information, including: include: Generate the values ​​of the benchmark periodic functions sin(2πx / 365) and cos(2πx / 365) to represent the fixed periodicity of the Gregorian calendar year. Build dynamic period adjustment parameters based on the benchmark periodic functions and introduce a sliding window mechanism to dynamically adjust the period length according to the month. Construct a feature transformation for the lunar calendar and holidays, assign dynamic weights based on holiday types, align historical data by lunar calendar date, and calculate average sales volatility before, during, and after holidays. Obtain regional-climate coupling analysis data clusters, establish independent climate response models for each data cluster, and extract features including: regional temperature-sales curves and precipitation impact thresholds; Specifically, the regional temperature-sales curve fits the polynomial relationship between temperature and sales, such as the inflection point of the negative correlation between down jacket sales and temperature; the precipitation impact threshold is a piecewise function that sets the precipitation effect on sales, such as light rain promoting the sales of rain gear, and heavy rain suppressing travel-related products.

[0021] The periodic values ​​of the baseline periodic functions sin(2πx / 365) and cos(2πx / 365) are cross-multiplied with the lunar month to capture the impact of the lunar leap month on the seasonal cycle. The regional climate zone labels are then combined with the future temperature forecast to generate a regional climate stress index. Combine regional climate zone labels with future temperature forecasts to generate a regionalized climate stress index. Analyze features using SHAP values ​​and select features with contributions above a threshold. Specifically, the prediction algorithm first needs to obtain a variety of data, such as regional business information, marketing information, date characteristics, product information, and historical operating data, namely basic sales and promotion data. For seasonal products, seasonal climate and weather information is particularly important for product prediction. Therefore, a Database (Impala, Hive) resource algorithm has been added to complete automatic product tagging. If the product is new, the Benchmark value is obtained, and then the RefInfo information is obtained to determine whether the product has a special promotion. If so, the special promotion is used as reference information. If not, the RefInfo information of non-new products is used. S102: Preprocess the acquired multi-dimensional data information to obtain historical and year-on-year inventory out-of-stock characteristics during promotional periods, identify the inventory out-of-stock characteristics, complete historical sales data compensation, and obtain historical sales compensation data; The obtained multi-dimensional data information is preprocessed, and the preprocessing includes: Data cleaning and integration: We use the quantile threshold method to identify abnormal sales fluctuations, unify time zones for multi-source data, establish a joint primary key for product, climate, and inventory, build an entity relationship network through a graph database, and identify implicit association rules. Intelligent historical sales compensation processing: Based on ERP system logs, the inventory status of historical periods is identified. When sales suddenly drop while competing products increase during the same period, a manual review process is triggered to determine the maximum possible estimate of inventory constraints. The time decay effect is introduced and the decay parameters are dynamically adjusted. Specifically, after acquiring the data, the system processes, cleans, and integrates the data based on the characteristics of the selected model to meet the needs of modeling. For example, based on inventory status during recent and year-over-year promotional periods, intelligent historical sales compensation is performed for items with significant out-of-stock conditions during recent and year-over-year promotional periods. This is done by removing year-over-year promotions. When processing the acquired data, the system uses pre-set initial algorithm models, such as Long Short-Term Memory (LSTM) and gradient boosting algorithms (XGBoost and LightGBM), to train the model based on the historical operating data of the data set. This yields a prediction model set (pkg) for seasonal products, which can be used as Ref Item Info information. S103: Using an initial algorithm model set including a long short-term memory network and gradient boosting to complete training on the historical sales compensation data, the training obtains sales characteristics of seasonal products within a preset time period; The training of historical sales compensation data was completed using an initial algorithm model set including long short-term memory networks and gradient boosting, including: Input historical sales compensation data to complete time series feature reconstruction, and determine whether the constructed feature grouping is a climate-sensitive product group or an inventory-dependent product group; Configure the LSTM time series base model through the model collaboration mechanism, complete the baseline prediction of the gradient boosting residual correction model, and complete dynamic weight allocation; Use the classification model to distinguish the original data from the compensated data, and apply the Kolmogorov-Smirnov test to align the data distribution before and after compensation. If the classification model accuracy is less than 95%, the data distribution correction is triggered until the classification model accuracy reaches 95% and the correction stops. S104: Based on the sales characteristics of seasonal products in a preset time period, the sales characteristics are placed in an initial algorithm model including a long short-term memory network and gradient boosting, and a seasonal product prediction model set is generated through continuous identification and prediction; Among them, by placing the sales characteristics of seasonal products in a preset time period into an initial algorithm model containing long short-term memory networks and gradient boosting, a seasonal product prediction model set is generated through continuous identification and prediction, including: Perform periodic decomposition on the acquired seasonal product sales characteristics, complete multi-granularity period division and feature generation statistics, and generate interactive features; Build a dual-channel LSTM architecture model. The main channel of the LSTM architecture model processes the original sales time series data, and the auxiliary channel learns the periodic signal generated by sin(2πx / 365) / cos(2πx / 365). Finally, it outputs the sales time series data integrated with the periodic signal. Obtain spatiotemporal cross-features, complete the calculation of climate elasticity coefficients through panel data regression, adopt a dynamic switching strategy of feature importance, complete hierarchical prediction of interactive features and LSTM architecture models, and complete the training of seasonal commodity forecasting model sets based on basic weighted fusion and dynamic adjustment and incremental training trigger conditions; S105: Based on the preset screening conditions, the current multi-dimensional data information is called, and the seasonal commodity forecasting model set performs dynamic forecasting on the current multi-dimensional data information to generate multi-strategy replenishment suggestions; When model training is complete, the Ref reference is calculated based on preset screening conditions, such as daily store sales, store quantity change and inventory status, member day / holiday sales ratio, historical and future weather conditions, etc. If it is not a reference, the Ref reference is determined to be an invalid benchmark. If it is a reference, the Ref reference can be used as the latest judgment benchmark and model training is carried out based on the new judgment benchmark. To increase the amount of model training, we added the sin(2πx / 365) and cos(2πx / 365) functions, along with historical order regions, important holidays, and the lunar calendar. This provides a comprehensive foundation for predictive modeling. Because weather conditions correlate with annual or quarterly seasonality, we matched periodic seasonality and promotional trends with the sin and cos functions, manually defining the periodic patterns that the two models can learn. For the cosine function, we can use a sliding window for January and December as needed. For example, if we set the lunar month as an example, the year-end and the beginning of the year are both cold winters, and sales have significant similarities. By adding date attribute labels such as weather, annual holidays, and New Year's Eve to distinguish them, we can significantly optimize the model's learning of seasonal product patterns and its ability to predict future trends. Forecast sales of each seasonal product, process the forecast control output rules, use the seasonal product forecast model set to perform actual forecasts, and finally complete the forecast output; That is, based on preset screening conditions, dynamic forecast suggestions for multiple strategies are obtained, which can be applied to dynamic replenishment strategies; By focusing on a multi-strategy dynamic forecasting method based on seasonal characteristics, supply chain inventory status and future climate conditions, it deeply integrates knowledge from multiple fields such as trend quantification of seasonal products, historical sales reference, and optimal forecast formulation. It is suitable for promotion forecasting and replenishment management of large-scale batch seasonal products, effectively resolving the problem of seasonal product backlogs, while avoiding the dilemma of being unable to meet consumer demand due to inventory shortages, helping companies achieve efficient supply chain management and accurate market response.

[0022] In this embodiment, the multi-strategy dynamic generation mechanism and model fusion architecture achieve a precise balance between inventory costs and sales opportunities. The three-dimensional replenishment decision-making system optimizes the supply chain, including: a climate resilience strategy to dynamically adjust the stocking level; a historical analogy strategy using the DTW algorithm to match historically similar seasons, combined with lunar calendar feature alignment technology to improve the accuracy of seasonal stocking timing for clothing products; a dynamic safety stock strategy adjusts the safety threshold through real-time inventory turnover rate, and dynamically adjusts the model weight distribution by continuously comparing predicted values ​​with actual sales data. The online learning mechanism shortens the response time to sudden climate events and reduces the prediction deviation compared to the static model.

[0023] This technical solution breaks through the fixed cycle limitations of the Gregorian calendar by coupling the lunar calendar sliding window with trigonometric functions. Compared with the traditional LSTM model, the prediction accuracy is improved. At the same time, the climate impact factor calculation unit establishes a nonlinear relationship model between temperature and sales, successfully capturing the impact of sudden climate events. In the cold wave warning scenario, the prediction response speed is shortened. In addition, the intelligent historical sales compensation mechanism significantly corrects the data deviation caused by supply chain disruptions through out-of-stock period identification and counterfactual modeling. Combined with the dynamic attenuation factor design, it not only avoids prediction distortion caused by over-compensation, but also ensures that the compensation amount is dynamically adjusted with inventory turnover.

[0024] Reference Figure 4 As shown, a multi-strategy dynamic prediction system based on seasonal products is used to execute the above-mentioned multi-strategy dynamic prediction method. The dynamic prediction system includes: a data enhancement module, which is used to collect multi-dimensional data information published by the historical network, and introduce periodic trigonometric functions, historical order area information, important holidays and lunar calendar information to enhance the extended seasonal climate characteristic data; an intelligent compensation module, which completes the preprocessing of multi-dimensional data information, and obtains the inventory out-of-stock situation characteristics of the historical and year-on-year promotion periods, completes the historical sales data compensation, and obtains the historical sales compensation data; a model integration module, which is used to complete the training of the historical sales compensation data using an initial algorithm model set including a long short-term memory network and a gradient boosting, and sets the sales characteristics of seasonal products in a preset time period based on the sales characteristics of the seasonal products in the preset time period. In the initial algorithm model including long short-term memory network and gradient boosting, a seasonal commodity prediction model set is generated through continuous identification and prediction; the strategy generation module is used for the seasonal commodity prediction model set to dynamically predict the current multi-dimensional data information and generate multi-strategy replenishment recommendations. The three-dimensional replenishment decision-making system in optimizing the supply chain includes: climate resilience strategy to dynamically adjust the stocking level, historical analogy strategy using DTW algorithm to match historical similar seasons, combined with lunar calendar feature alignment technology to improve the accuracy of seasonal stocking timing judgment for clothing products, dynamic safety stock strategy adjusts the safety threshold through real-time inventory turnover rate, and dynamically adjusts the model weight distribution by continuously comparing the predicted value with the actual sales data. The online learning mechanism shortens the response time to sudden climate events and reduces the prediction deviation compared with the static model.

[0025] In this city's example, the data enhancement module includes: a lunar calendar feature conversion unit, which is used to convert Gregorian calendar dates into lunar calendar date features; a climate impact factor calculation unit, which is used to establish a nonlinear relationship model between temperature, precipitation and commodity sales, realize a multi-strategy dynamic generation mechanism and model fusion architecture, and achieve a precise balance between inventory costs and sales opportunities. By coupling the lunar calendar sliding window with trigonometric functions, it breaks through the fixed cycle limitation of the Gregorian calendar, and the prediction accuracy is improved compared with the traditional LSTM model. At the same time, the climate impact factor calculation unit establishes a nonlinear relationship model between temperature and sales, successfully capturing the impact of sudden climate events. In the cold wave warning scenario, the prediction response speed is shortened. In addition, the intelligent historical sales compensation mechanism significantly corrects the data deviation caused by supply chain disruptions through out-of-stock period identification and counterfactual modeling. Combined with the dynamic attenuation factor design, it avoids the prediction distortion caused by over-compensation and ensures that the compensation amount is dynamically adjusted with the inventory turnover rate. Reference Figure 5 As shown, the multi-strategy dynamic forecasting device based on seasonal commodities includes a storage unit and a processor. The storage unit is used to store a computer program. The processor runs the computer program to enable the multi-strategy dynamic forecasting device to perform the multi-strategy dynamic forecasting method based on seasonal commodities. The technical advantages of this technical solution include dynamic cycle adjustment: Solve the problem of "misalignment between the fixed Gregorian calendar cycle and natural climate seasons" through the lunar calendar sliding window (such as the sales peak drift caused by the Lunar New Year); Refined regional coupling: Regional climate models avoid a “one-size-fits-all” approach. For example, a 5°C temperature drop will have different impacts on sales in Harbin (normal low temperature) and Shanghai (sudden cold wave).

[0026] Enhanced interpretability: SHAP value verification proves that the lunar solar terms and regional climate factors have a significant impact on the prediction results (for example, the SHAP value of the winter solstice accounts for 12%).

[0027] A computer storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned multi-strategy dynamic forecasting method based on seasonal products. The multi-strategy dynamic forecasting method, which is based on seasonal characteristics, supply chain inventory status, and future climate conditions, deeply integrates knowledge from multiple fields such as trend quantification of seasonal products, historical sales reference, and optimal forecast formulation. The method is suitable for promotion forecasting and replenishment management of large-scale batch seasonal products, effectively resolving the problem of seasonal product backlogs, while avoiding the dilemma of being unable to meet consumer demand due to inventory shortages, and helping enterprises achieve efficient supply chain management and accurate market response.

[0028] In this embodiment, the entire operation process can be controlled by a computer to provide signal feedback to implement the steps in sequence. These are all conventional knowledge of current automated control and will not be described in detail in this embodiment.

[0029] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-strategy dynamic forecasting method based on seasonal commodities, characterized by: Applied to electronic forecasting equipment, specifically including the following steps: S101: Acquire multi-dimensional data information published on historical networks, wherein the multi-dimensional data information carries regional business data, marketing activity data, date feature data, product attribute data, historical business data, and extended seasonal climate feature data; S102: Preprocess the acquired multi-dimensional data information to obtain historical and year-on-year inventory out-of-stock characteristics during promotional periods, identify the inventory out-of-stock characteristics, complete historical sales data compensation, and obtain historical sales compensation data; S103: Using an initial algorithm model set including a long short-term memory network and gradient boosting to complete training on the historical sales compensation data, the training obtains sales characteristics of seasonal products within a preset time period; S104: Based on the sales characteristics of seasonal products in a preset time period, the sales characteristics are placed in an initial algorithm model including a long short-term memory network and gradient boosting, and a seasonal product prediction model set is generated through continuous identification and prediction; S105: Based on the preset screening conditions, the current multi-dimensional data information is called, and the seasonal commodity forecasting model set performs dynamic forecasting on the current multi-dimensional data information to generate multi-strategy replenishment suggestions.

2. The multi-strategy dynamic forecasting method based on seasonal commodities according to claim 1, characterized in that: In S101, the seasonal climate characteristic data is enhanced by introducing periodic trigonometric functions, historical order demand area information, important holidays and lunar calendar information.

3. The multi-strategy dynamic forecasting method based on seasonal commodities according to claim 2, characterized in that: The enhancements to the seasonal climate characteristics data include: Generate the values ​​of the benchmark periodic functions sin(2πx / 365) and cos(2πx / 365) to represent the fixed periodicity of the Gregorian calendar year. Build dynamic period adjustment parameters based on the benchmark periodic functions and introduce a sliding window mechanism to dynamically adjust the period length according to the month. Construct a feature transformation for the lunar calendar and holidays, assign dynamic weights based on holiday types, align historical data by lunar calendar date, and calculate average sales volatility before, during, and after holidays. Obtain regional-climate coupling analysis data clusters, establish independent climate response models for each data cluster, and extract features including: regional temperature-sales curves and precipitation impact thresholds; The periodic values ​​of the baseline periodic functions sin(2πx / 365) and cos(2πx / 365) are cross-multiplied with the lunar month to capture the impact of the lunar leap month on the seasonal cycle. The regional climate zone labels are then combined with the future temperature forecast to generate a regional climate stress index. The regional climate zone labels are combined with the future temperature forecast values ​​to generate a regionalized climate pressure index. The features are analyzed through SHAP values ​​to screen out features with contribution values ​​above the threshold.

4. The multi-strategy dynamic forecasting method based on seasonal commodities according to claim 3, characterized in that: In S102, the acquired multi-dimensional data information is preprocessed, and the preprocessing includes: Data cleaning and integration: We use the quantile threshold method to identify abnormal sales fluctuations, unify time zones for multi-source data, establish a joint primary key for product, climate, and inventory, build an entity relationship network through a graph database, and identify implicit association rules. Intelligent historical sales compensation processing: Based on the ERP system log, the inventory status of the historical period is identified. When sales suddenly drop but the sales of competing products increase during the same period, the manual review process is triggered to determine the maximum possible estimate of inventory constraints, introduce the time decay effect, and dynamically adjust the decay parameters.

5. The multi-strategy dynamic forecasting method based on seasonal commodities according to claim 4, characterized in that: In S103, the training of the historical sales compensation data is completed using an initial algorithm model set including a long short-term memory network and gradient boosting, including: Input historical sales compensation data to complete time series feature reconstruction, and determine whether the constructed feature grouping is a climate-sensitive product group or an inventory-dependent product group; Configure the LSTM time series base model through the model collaboration mechanism, complete the baseline prediction of the gradient boosting residual correction model, and complete dynamic weight allocation; A classification model is used to distinguish the original data from the compensated data, and the Kolmogorov-Smirnov test is applied to align the data distribution before and after compensation. If the accuracy of the classification model is less than 95%, the data distribution correction and repair is triggered until the classification model accuracy reaches 95% and the correction stops.

6. The multi-strategy dynamic forecasting method based on seasonal products according to claim 5, characterized in that: In S104, based on the sales characteristics of seasonal products in a preset time period, the model is placed in an initial algorithm model including a long short-term memory network and gradient boosting, and a seasonal product prediction model set is generated through continuous identification and prediction, including: Perform periodic decomposition on the acquired seasonal product sales characteristics, complete multi-granularity period division and feature generation statistics, and generate interactive features; Construct a dual-channel LSTM architecture model. The main channel of the LSTM architecture model processes the original sales time series data, and the auxiliary channel learns the periodic signal generated by sin(2πx / 365) / cos(2πx / 365). Finally, it outputs the sales time series data integrated with the periodic signal. Obtain spatiotemporal cross-features, complete the climate elasticity coefficient through panel data regression calculation, adopt the dynamic switching strategy of feature importance, complete the hierarchical prediction of interactive features and LSTM architecture model, and complete the training of seasonal commodity prediction model set based on basic weighted fusion and dynamic adjustment and incremental training trigger conditions.

7. A multi-strategy dynamic forecasting system based on seasonal commodities, characterized by: Used to execute the multi-strategy dynamic prediction method according to any one of claims 1 to 6, the dynamic prediction system comprises: The data enhancement module is used to collect multi-dimensional data information published on the historical network and introduce periodic trigonometric functions, historical order and demand area information, important holidays and lunar calendar information to enhance the extended seasonal climate characteristic data; The intelligent compensation module completes the preprocessing of multi-dimensional data information and obtains the inventory out-of-stock characteristics of historical and year-on-year promotion periods, completes the compensation of historical sales data, and obtains historical sales compensation data; The model integration module is used to complete the training of historical sales compensation data using an initial algorithm model set including a long short-term memory network and gradient boosting. The sales characteristics of seasonal products in a preset time period are placed in the initial algorithm model including a long short-term memory network and gradient boosting, and a seasonal product prediction model set is generated through continuous identification and prediction; The strategy generation module is used to dynamically predict the current multi-dimensional data information using the seasonal commodity forecasting model set and generate multi-strategy replenishment recommendations.

8. The multi-strategy dynamic forecasting system based on seasonal products according to claim 7, characterized in that: The data enhancement module includes: A lunar calendar feature conversion unit, used to convert a Gregorian calendar date into a lunar calendar date feature; The climate impact factor calculation unit is used to establish a nonlinear relationship model between temperature, precipitation and product sales.

9. A multi-strategy dynamic forecasting device based on seasonal commodities, characterized by: The device comprises a storage unit and a processor, wherein the storage unit is used to store a computer program, and the processor runs the computer program to enable the multi-strategy dynamic prediction device to perform the multi-strategy dynamic prediction method based on seasonal commodities according to any one of claims 1 to 6.

10. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the multi-strategy dynamic forecasting method based on seasonal commodities according to any one of claims 1 to 6 are implemented.

Images