Algorithm for predicting warehouse-out quantity under condition of frequent quality and regulation change based on time sequence

Abstract

The invention discloses a time sequence-based algorithm for predicting the warehouse-out quantity under the condition of frequent change of quality regulations, and the algorithm is divided into three stages: a first stage: predicting the warehouse-out quantity of quality regulations and non-new quality regulations by using a time sequence model, and forming the warehouse-out quantity prediction of all the quality regulations together with the warehouse-out quantity of the non-new quality regulations obtained in the first stage; according to the method, the time sequence models are used in groups, various ex-warehouse quantities are predicted, the proportional relation between the new product ex-warehouse quantity and the non-new product ex-warehouse quantity in each product specification classification is calculated, the ex-warehouse quantity of the new product specification is calculated by using the previous result, and the ex-warehouse quantity of the new product specification and the obtained non-new product specification ex-warehouse quantity form ex-warehouse quantity prediction of all product specifications together, so that the prediction accuracy of the ex-warehouse quantity of all product specifications is improved. On the premise that the training speed of the model is within an acceptable range, the problem that an effective model is established to predict the ex-warehouse quantity under the condition that the specifications of fast moving consumer goods are changed frequently is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of product specification outbound quantity prediction, and specifically to an algorithm based on time series for predicting outbound quantity when product specifications change frequently. Background Art

[0002] The existing time series-based product specification outbound quantity prediction methods are not adaptable enough to fast-moving consumer goods. The main problem is that the product specifications of fast-moving consumer goods change frequently, new product specifications are often added, and a large proportion of product specifications are withdrawn within a few years. The life cycle of fast-moving consumer goods is measured by the first outbound time and the last outbound time, and fast-moving consumer goods with a life cycle of less than one year are defined as short life cycle. According to data analysis, the proportion of fast-moving consumer goods with short life cycles (life cycle less than one year) in XX warehouses shows significant fluctuations: 32.0% in 2021, 1.08% in 2022, 1.53% in 2023, and 38.86% in 2024. This results in these models being unable to make good predictions on the outbound quantity of new products, and due to the rapid withdrawal of product specifications, the model may not be able to obtain sufficiently long outbound data to derive patterns. Summary of the invention

[0003] The purpose of the present invention is to provide an algorithm based on time series for predicting the inventory shipment under the condition of frequent changes in product specifications, so as to solve the problems raised in the above background technology.

[0004] To achieve the above purpose, the present invention provides the following technical solution: an algorithm based on time series to predict the outbound quantity when product specifications change frequently. The algorithm is divided into three stages: Phase 1: Use time series models to predict product category and non-new product shipments; The second stage: for each product category, calculate the ratio of the new product quantity shipped out to the non-new product quantity shipped out within this category; The third stage: By using the results of the first two stages, the shipment volume of new specifications is calculated, and together with the shipment volume of non-new specifications in the first stage, it constitutes the shipment volume forecast for all specifications.

[0005] Preferably, the first stage uses a time series model to predict the shipment volume of non-new product specifications and product specification categories, and uses the CatBoost ladder boosting tree time series model in the time series processing library darts to predict the shipment volume of non-new product specifications and product specification categories. CatBoost is a machine learning algorithm based on gradient boosting, which can directly process categorical variables and is suitable for large-scale, multi-classification feature data. It uses the values ​​of the past several days (180 days in this case) as features, adds time-related covariates such as year, month, and Sunday, generates multiple time series models for training, and predicts the shipment volume of non-new product specifications and product specification categories in the next several days (30 days in this case), and marks them according to specific time period characteristics (such as holidays and promotions) to improve the model. Accuracy. The algorithm automatically uses a sliding window method to train data based on the parameters of input and output days. First, data cleaning is performed. We need to obtain the outbound volume of each product specification in each period of time (generally a month). A short time is conducive to improving the accuracy of short-term predictions, but it will bring greater computing overhead; and there are certain requirements for the data density of outbound volume, so it can only be applied to products with frequent outbound shipments. Since logistics only uses months as the unit for predicting outbound volume, we choose months as the statistical time unit. For a single product specification that is shipped multiple times within a month, its outbound volume is calculated in aggregate. For the outbound volume of each product specification obtained, the characteristics of the time period are added, including festival characteristics and specific event characteristics related to the time period, to help the model learn the outbound characteristics of the Spring Festival, so that the model You can make a better judgment. Concept explanation: New product specifications: For product specifications that do not exceed n time periods after the first shipment (inclusive), we consider them to be new product specifications. The larger the value of n, the more complex the model establishment, which will not only increase the amount of calculation, but also have higher requirements for the density of shipment data. Although the shipment volume of fast-moving consumer goods has monthly correlation (it can be roughly understood that the goods have different sales in each different season, and this pattern repeats in years), the value of n, multiplied by the length of the time period (usually months), should be 12 or its integer multiple. However, considering the data density and calculation amount, and new product specifications are usually not put on the market until their corresponding peak sales season, n can take some smaller values. For example, the corresponding time can be set to 6 months or less. For each product category, including the lower-level categories, it is necessary to count the number of new products in the category in different outbound time periods. The statistical content includes the following items: the number of product categories with outbound quantities for the first time in the time period, the number of product categories in a time period after the first outbound quantity (that is, the product categories with outbound quantities for the first time in the previous time period, regardless of whether the product categories have outbound quantities in the current time period), the number of product categories in two time periods after the first outbound quantity, and so on, until the number of product categories in n-1 time periods after the first outbound quantity. The value of n is determined according to the definition of new products, usually set to 6 months, which can be shortened to 3 months or extended to 12 months under special circumstances. When calculating the outbound ratio of a product category, multiple different outbound ratios may be generated. In this case,Priority is given to selecting the product specification outbound ratio with a larger proportion of the historical outbound volume to the total outbound volume of the product specification category as the representative. If there is no obvious dominant product specification, the weighted average of the outbound ratios of all product specifications is used. The weight is determined by the proportion of the historical outbound volume of each product specification to the total outbound volume of the category. These statistical values ​​are attached to the corresponding product specification category as the characteristics of the product specification category. The emergence of new products usually has a positive impact on the total outbound volume of the product specification category. Therefore, these characteristic values ​​will also have a positive effect on the outbound volume prediction of the product specification category. For each product specification and product specification category, a weight value is set according to its representativeness. , the default value is 1, the value is greater than 0, and can be a decimal. This weight value will affect the weight of this product specification and product specification category in the model training process. For the outbound volume and characteristics of these product specifications, a time series model is established. Since the time series model needs to put each product specification into training, it can only be executed sequentially on one GPU. When the number of product specifications is large, the training time is too long, and the outbound data between different product specification categories may lack consistency. Therefore, these product specification categories are divided into several groups, and each group is trained separately. Different models are obtained from the collection. At the same time, the group training does not affect each other, so it can be run in parallel on multiple GPUs. In order to reduce the time required for training, a basic number of training times (for example, 1000 times) is set. The number of training times for each specification is set to the product of its weight value. Concept explanation: Number of items in a single outbound package: A concept introduced to balance the difference in outbound quantities of different specifications under the same specification category due to different packaging. For example, two beer cans have similar volumes and are both outbound in boxes. Product A has 4 cans in a box and product B has 6 cans in a box. When calculating the relevant index, the outbound quantity of A needs to be multiplied by 4 and the outbound quantity of B needs to be multiplied by 6 to calculate the outbound situation under this specification category. In order to exclude its influence when the situation occurs, each specification should belong to multiple levels of specification categories. Taking XX Cola as an example, the specification categories can be divided into the first-level specification category of cola-type beverages, the second-level specification category of soft drinks, and the third-level specification category of beverages from fine to coarse. The higher-level specification categories can be used as a reference when the lower-level classification data is insufficient. When the classification data is not enough to meet the minimum sample size, the higher-level specification categories are used for prediction. For each specification category, the outbound ratio of each specification category is calculated respectively. The calculation formula for the outbound ratio of the specification is: , the outbound ratio represents the average outbound volume of non-new specifications under this specification category. Through this formula, the impact of different packaging specifications on the total outbound volume can be effectively balanced to avoid the increase in classification model errors caused by specification differences. After this stage is completed, multiple time series models will be obtained. Each non-new specification and specification category will exist in a model. Through model prediction and calculation, the outbound volume of non-new specifications in a specific future period of time and the outbound ratio i of the specification category in a specific future period of time can be obtained.

[0006] Preferably, the second stage calculates the ratio of the new product shipment volume to the non-new product shipment volume. Since the first stage model cannot accurately predict the new product shipment volume, this stage predicts the new product shipment volume by inferring the ratio of the new product shipment volume to the non-new product shipment volume in each product category. This stage needs to be performed separately for each first-level product category. When a certain category cannot be modeled due to insufficient quantity, its upper-level product category can be used for modeling. When modeling the upper-level product category, all of its included product categories should be recursively included, including all the products belonging to it. The historical records are processed from old to new according to each time period, and the product specifications that appear for the first time in this time period are found, and its two setting values ​​are read: A: The proportion of shipment time, which indicates the proportion of the shipment time of the new product specification in this time period. For the first shipment time period, A can be less than 1, indicating that the new product specification is shipped only during part of this time period. For subsequent time periods, A The default value is 1, which means that the new product specification is continuously shipped out during the entire time period. B: New product feature multiplier. The adjustment coefficient of the shipment quantity is determined according to factors such as the market demand and promotional activities of the product specification. Suppose the shipment quantity of this product specification in this time period is x, the total shipment quantity of this product specification category in this time period is s, and the number of products of this product specification category in this time period is q. Find: , this value is the ratio of the shipment volume of this product specification in the first time period to the average shipment volume of this product specification category in this time period. In the second to nth time period (the value of n is the same as the n defined for new products in the first phase), since A is 1, the above formula can be simplified to: , and set the ratio to 1 in the n+1th time period, use cubic spline interpolation to generate a smooth cubic polynomial curve for the ratios of all the above time periods, and fit the new product shipment curve Z for this product category.

[0007] Preferably, the third stage comprehensively calculates the shipment volume of new products and non-new products based on the results of the first two stages, generates shipment volume forecasts for all specifications, and directly uses the time series model generated in the first stage to predict the shipment volume of old specifications (specifications that have exceeded the n time periods defined by the new specifications). For new specifications with data of less than three time periods, for new specifications with historical data of less than three time periods (i.e., data of only 1 or 2 time periods), linear extrapolation method is used to predict shipment volume. Linear extrapolation method: When there is only data of one time period, directly use the shipment volume ratio of the first time period For reference, , when there are data from two time periods, read the ratio of the outbound volume in the first time period to the second time period and , and according to the changing trends of these two time periods, linear extrapolation is performed to predict the outbound volume ratio in the future time period t: ,in, and The first and second point values ​​of the new product specification outbound quantity curve, and The ratio of the outbound volume in the historical time period. The data is for new product specifications with three or more time periods. When the new product specification has historical data for three or more time periods, cubic spline interpolation is used to smooth the data to predict the future outbound volume. The cubic spline interpolation method is: read the outbound volume ratio of the previous n time periods from the historical data And the corresponding point value of the new product specification outbound quantity curve , using these data, a smooth curve is fitted through cubic spline interpolation to represent the change trend of the new product's outbound quantity in different time periods, and the value of the interpolation curve in the future time period t is calculated , this value is the predicted new product shipment ratio for this period of time: , calculate the predicted outbound volume of the new product specification. Whether the new product specification outbound volume ratio p is calculated by linear extrapolation or cubic spline interpolation, it can be combined with the product specification category outbound index i obtained in the first stage to finally calculate the predicted outbound volume of the new product specification in the future time period t: , where q is the number of new specifications in the specification category during the time period.

[0008] Compared with the prior art, the present invention has the following beneficial effects: The present invention uses a time series model in groups to predict various outbound quantities. For each specification classification, the proportional relationship between the outbound quantity of new products and the outbound quantity of non-new specifications in this classification is deduced. By using the previous results, the outbound quantity of new specifications is calculated, and together with the obtained outbound quantity of non-new specifications, it constitutes the outbound quantity prediction of all specifications. On the premise that the training speed of the model is within an acceptable range, the problem of establishing an effective model to predict the outbound quantity is solved when the specifications of fast-moving consumer goods change frequently. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0010] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0011] See also Figure 1 Based on time series, the algorithm predicts the outbound volume when product specifications change frequently. The algorithm is divided into three stages: Phase 1: Use time series models to predict product category and non-new product shipments; The second stage: for each product category, calculate the ratio of the new product quantity shipped out to the non-new product quantity shipped out within this category; The third stage: By using the results of the first two stages, the shipment volume of new specifications is calculated, and together with the shipment volume of non-new specifications in the first stage, it constitutes the shipment volume forecast for all specifications.

[0012] In the first stage, a time series model is used to predict the shipment volume of non-new products and product categories. The CatBoost ladder boosting tree time series model in the time series processing library darts is used to predict the shipment volume of non-new products and product categories. CatBoost is a machine learning algorithm based on gradient boosting that can directly process categorical variables and is suitable for large-scale, multi-classified feature data. It uses the values ​​of the past several days (180 days in this case) as features, adds time-related covariates such as year, month, and Sunday, generates multiple time series models for training, and predicts the shipment volume of non-new products and product categories in the next several days (30 days in this case), and marks them according to specific time period characteristics (such as holidays and promotions) to improve the accuracy of the model. The algorithm automatically uses a sliding window method to train data based on the parameters of input and output days. First, data cleaning is performed. We need to obtain the outbound volume of each product specification in each period of time (generally a month). A short time is conducive to improving the accuracy of short-term predictions, but it will bring greater computing overhead; and there are certain requirements for the data density of the outbound volume, so it can only be applied to products with frequent outbound shipments. Since logistics predicts outbound volume only in months, we choose months as the statistical time unit. For a single product specification that is shipped multiple times within a month, its outbound volume is calculated in aggregate. For the outbound volume of each product specification, the characteristics of the time period are added, including festival characteristics and specific event characteristics related to the time period, to help the model learn the outbound characteristics of the Spring Festival, so that the model can be better To make a better judgment, the concept is explained as follows: New product specifications: For product specifications that do not exceed n time periods after the first shipment (inclusive), we consider them to be new product specifications. The larger the value of n, the more complex the model establishment will be, which will not only increase the amount of calculation, but also have higher requirements for the density of shipment data. Although the shipment volume of fast-moving consumer goods has monthly correlation (it can be roughly understood that the goods have different sales in each different season, and this pattern repeats in years), the value of n, multiplied by the length of the time period (generally months), should be 12 or its integer multiple. However, considering the data density and calculation amount, and the new product specifications are usually not put on the market until their corresponding sales peak season, n can take some smaller values. For example, the corresponding time can be set to 6 months or less. For each There are a number of product categories, including lower-level categories. It is necessary to count the number of new products in this category in different outbound time periods. The statistical content includes the following items: the number of product categories with outbound quantities for the first time in this time period, the number of product categories in a time period after the first outbound quantity (that is, the product categories with outbound quantities for the first time in the previous time period, and regardless of whether the product categories have outbound quantities in this time period), the number of product categories in two time periods after the first outbound quantity, and so on, until the number of product categories in n-1 time periods after the first outbound quantity. The value of n is determined according to the definition of new products, usually set to 6 months, which can be shortened to 3 months or extended to 12 months under special circumstances. When calculating the outbound ratio of a product category, multiple different outbound ratios may be generated. In this case,Priority is given to selecting the product specification outbound ratio with a larger proportion of the historical outbound volume to the total outbound volume of the product specification category as the representative. If there is no obvious dominant product specification, the weighted average of the outbound ratios of all product specifications is used. The weight is determined by the proportion of the historical outbound volume of each product specification to the total outbound volume of the category. These statistical values ​​are attached to the corresponding product specification category as the characteristics of the product specification category. The emergence of new products usually has a positive impact on the total outbound volume of the product specification category. Therefore, these characteristic values ​​will also have a positive effect on the outbound volume prediction of the product specification category. For each product specification and product specification category, a weight value is set according to its representativeness. , the default value is 1, the value is greater than 0, and can be a decimal. This weight value will affect the weight of this product specification and product specification category in the model training process. For the outbound volume and characteristics of these product specifications, a time series model is established. Since the time series model needs to put each product specification into training, it can only be executed sequentially on one GPU. When the number of product specifications is large, the training time is too long, and the outbound data between different product specification categories may lack consistency. Therefore, these product specification categories are divided into several groups, and each group is trained separately. Different models are obtained from the collection. At the same time, the group training does not affect each other, so it can be run in parallel on multiple GPUs. In order to reduce the time required for training, a basic number of training times (for example, 1000 times) is set. The number of training times for each specification is set to the product of its weight value. Concept explanation: Number of items in a single outbound package: A concept introduced to balance the difference in outbound quantities of different specifications under the same specification category due to different packaging. For example, two beer cans have similar volumes and are both outbound in boxes. Product A has 4 cans in a box and product B has 6 cans in a box. When calculating the relevant index, the outbound quantity of A needs to be multiplied by 4 and the outbound quantity of B needs to be multiplied by 6 to calculate the outbound situation under this specification category. In order to exclude its influence when the situation occurs, each specification should belong to multiple levels of specification categories. Taking XX Cola as an example, the specification categories can be divided into the first-level specification category of cola-type beverages, the second-level specification category of soft drinks, and the third-level specification category of beverages from fine to coarse. The higher-level specification categories can be used as a reference when the lower-level classification data is insufficient. When the classification data is not enough to meet the minimum sample size, the higher-level specification categories are used for prediction. For each specification category, the outbound ratio of each specification category is calculated respectively. The calculation formula for the outbound ratio of the specification is: , the outbound ratio represents the average outbound volume of non-new specifications under this specification category. Through this formula, the impact of different packaging specifications on the total outbound volume can be effectively balanced to avoid the increase in classification model errors caused by specification differences. After this stage is completed, multiple time series models will be obtained. Each non-new specification and specification category will exist in a model. Through model prediction and calculation, the outbound volume of non-new specifications in a specific future period of time and the outbound ratio i of the specification category in a specific future period of time can be obtained.

[0013] In the second stage, the ratio of new product shipments to non-new product shipments is calculated. Since the model in the first stage cannot accurately predict the shipment of new products, this stage predicts the shipment of new products by estimating the ratio of new product shipments to non-new product shipments in each product category. This stage needs to be performed separately for each first-level product category. When a certain category cannot be modeled due to insufficient quantity, its upper-level product category can be used for modeling. When modeling the upper-level product category, all of its included product categories should be recursively included, including all the products belonging to it. The historical records are processed from old to new according to each time period, and the product specifications that appear for the first time in this time period are found. Its two setting values ​​are read: A: The proportion of shipment time, which indicates the proportion of the shipment time of the new product specification in this time period. For the first shipment time period, A can be less than 1, indicating that the new product specification is shipped only during part of this time period. For subsequent time periods, A The default value is 1, which means that the new product specification is continuously shipped out during the entire time period. B: New product feature multiplier. The adjustment coefficient of the shipment quantity is determined according to factors such as the market demand and promotional activities of the product specification. Suppose the shipment quantity of this product specification in this time period is x, the total shipment quantity of this product specification category in this time period is s, and the number of products of this product specification category in this time period is q. Find: , this value is the ratio of the shipment volume of this product specification in the first time period to the average shipment volume of this product specification category in this time period. In the second to nth time period (the value of n is the same as the n defined for new products in the first phase), since A is 1, the above formula can be simplified to: , and set the ratio to 1 in the n+1th time period, use cubic spline interpolation to generate a smooth cubic polynomial curve for the ratios of all the above time periods, and fit the new product shipment curve Z for this product category.

[0014] The third stage is based on the results of the first two stages. The outbound volume of new and non-new products is calculated comprehensively to generate outbound volume forecasts for all specifications. For old specifications (specifications that have exceeded the n time periods defined by the new specifications), the time series model generated in the first stage is directly used to predict their outbound volume. For new specifications with data of less than three time periods, for new specifications with historical data of less than three time periods (that is, only 1 or 2 time periods of data), the linear extrapolation method is used to predict the outbound volume. The linear extrapolation method: when there is only one time period of data, the outbound volume ratio of the first time period is directly used. For reference, , when there are data from two time periods, read the ratio of the outbound volume in the first time period to the second time period and , and according to the changing trends of these two time periods, linear extrapolation is performed to predict the outbound volume ratio in the future time period t: ,in, and The first and second point values ​​of the new product specification outbound quantity curve, and The ratio of the outbound volume in the historical time period. The data is for new product specifications with three or more time periods. When the new product specification has historical data for three or more time periods, cubic spline interpolation is used to smooth the data to predict the future outbound volume. The cubic spline interpolation method is: read the outbound volume ratio of the previous n time periods from the historical data And the corresponding point value of the new product specification outbound quantity curve , using these data, a smooth curve is fitted through cubic spline interpolation to represent the change trend of the new product's outbound quantity in different time periods, and the value of the interpolation curve in the future time period t is calculated , this value is the predicted new product shipment ratio for this period of time: , calculate the predicted outbound volume of the new product specification. Whether the new product specification outbound volume ratio p is calculated by linear extrapolation or cubic spline interpolation, it can be combined with the product specification category outbound index i obtained in the first stage to finally calculate the predicted outbound volume of the new product specification in the future time period t: , where q is the number of new specifications in the specification category during the time period.

[0015] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An algorithm based on time series to predict the outbound volume when product specifications change frequently, characterized by: The algorithm is divided into three stages: Phase 1: Use time series models to predict product category and non-new product shipments; The second stage: for each product category, calculate the ratio of the new product quantity shipped out to the non-new product quantity shipped out within this category; The third stage: By using the results of the first two stages, the shipment volume of new specifications is calculated, and together with the shipment volume of non-new specifications in the first stage, it constitutes the shipment volume forecast for all specifications.

2. The algorithm for predicting the outbound quantity based on time series under the condition of frequent changes in product specifications according to claim 1 is characterized by: In the first stage, the time series model is used to predict the shipment volume of non-new specifications and specification categories, and the CatBoost ladder boosting tree time series model in the time series processing library darts is used to predict the shipment volume of non-new specifications and specification categories. The values ​​of the past few days are used as features, and time-related covariates such as year, month, and Sunday are added to generate multiple time series models for training. The shipment volume of non-new specifications and specification categories in the next few days is predicted and marked according to the characteristics of a specific time period. For each specification category, including the lower-level classification, it is necessary to count the number of new products in the classification in different shipment time periods. The statistical content includes the following items: the number of specifications that have shipped out for the first time in this time period, and the number of specifications in a time period after the first shipment. , the number of specifications in two time periods after the first shipment, and so on, until the number of specifications in n-1 time periods after the first shipment. The value of n is determined according to the definition of new products, usually set to 6 months, and can be shortened to 3 months or extended to 12 months under special circumstances. For each specification and specification category, a weight value is set according to its representativeness. The default value is 1, and the value is greater than 0, which can be a decimal. This weight value will affect the weight of this specification and specification category in the model training process. For the shipment quantity and characteristics of these specifications, a time series model is established. For each specification, it should belong to multiple levels of specification categories. For each specification category, the shipment ratio of each specification category is calculated respectively. The calculation formula for the shipment ratio of the specification is: .

3. The algorithm for predicting the outbound quantity based on time series under the condition of frequent changes in product specifications according to claim 1 is characterized by: In the second stage, the ratio of the new product outbound quantity to the non-new product outbound quantity is calculated. The historical records are processed from old to new according to each time period, and the product specifications that appear in the outbound quantity for the first time in this time period are found. Its two setting values ​​are read: A: outbound time ratio, which indicates the outbound time ratio of the new product specification in this time period. For the first outbound time period, A can be less than 1, indicating that the new product specification is only outbound during part of the time period. For subsequent time periods, A defaults to 1, indicating that the new product specification is continuously outbound during the entire time period. B: New product feature multiplier, the adjustment coefficient of the outbound quantity is determined according to factors such as the market demand and promotional activities of the product specification. Let the outbound quantity of this product specification in this time period be x, the total outbound quantity of this product specification category in this time period be s, and the number of product specifications of this product specification category in this time period be q. Find: , this value is the ratio of the shipment volume of this product specification in the first time period to the average shipment volume of this product specification category in this time period. In the second to nth time period (the value of n is the same as the n defined for new products in the first phase), since A is 1, the above formula can be simplified to: , and set the ratio to 1 in the n+1th time period, use cubic spline interpolation to generate a smooth cubic polynomial curve for the ratios of all the above time periods, and fit the new product shipment curve Z for this product category.

4. The algorithm for predicting the outbound quantity based on time series under the condition of frequent changes in product specifications according to claim 1 is characterized by: The third stage comprehensively calculates the shipment volume of new and non-new products based on the results of the first two stages, and generates shipment volume forecasts for all specifications. For old specifications, the time series model generated in the first stage is directly used to predict their shipment volume. For new specifications with data less than three time periods, the linear extrapolation method is used to predict shipment volume for new specifications with historical data less than three time periods. The linear extrapolation method: when there is only data for one time period, the shipment volume ratio of the first time period is directly used. For reference, , when there are data from two time periods, read the ratio of the outbound volume in the first time period to the second time period and , and according to the changing trends of these two time periods, linear extrapolation is performed to predict the outbound volume ratio in the future time period t: in, and The first and second point values ​​of the new product specification outbound quantity curve, and The ratio of the outbound volume in the historical time period. The data is for new product specifications with three or more time periods. When the new product specification has historical data for three or more time periods, cubic spline interpolation is used to smooth the data to predict the future outbound volume. The cubic spline interpolation method is: read the outbound volume ratio of the previous n time periods from the historical data The corresponding point value of the new product specification outbound quantity curve , calculate the value of the interpolation curve in the future time period t , this value is the predicted ratio of new product specifications to inventory during this period: , calculate the predicted outbound volume of the new product specification. Whether the new product specification outbound volume ratio p is calculated by linear extrapolation or cubic spline interpolation, it can be combined with the product specification category outbound index i obtained in the first stage to finally calculate the predicted outbound volume of the new product specification in the future time period t: , where q is the number of new specifications in the specification category during the time period.

Images