KR20230081028A - Method for generating a product demand forecasting model and for predicting product demand using the same - Google Patents

Abstract

The present invention relates to a method for generating a model capable of accurately predicting future demand in selling a product at an online or offline store and predicting demand through the model.

Description

Method for generating a product demand forecasting model and for predicting product demand using the same}

The present invention relates to the field of economics and IT technology, and relates to a method for generating a model capable of accurately predicting future demand in selling a product at an online or offline store and predicting demand therethrough.

Product demand forecasting has a great impact on various factors such as inventory management, display, and ordering in stores that sell products. In stores such as marts where food materials with a short distribution period are sold, the accuracy of demand forecasting is closely related to profits. In the past, demand was predicted based on the experience of employees in stores, but recently, due to the development of statistical techniques and artificial intelligence, more sophisticated technology for predicting demand is being developed.

In the case of online stores, demand forecasting accuracy is reported to reach 90%, but offline stores are known to be much lower than this. This is due to the fact that the online store secures most of the personal information of product buyers. Since it is difficult to secure personal information in offline stores, various benefits are induced to sign up as members, but the ratio of product purchases by non-members who do not provide personal information is very high, making it difficult to accurately predict demand.

One of the known methods for predicting product demand is a method of predicting by time-series decomposition of hourly sales volume data of a product. Here, sales volume data by time of a product is data having time as the x-axis and sales volume as the y-axis, as shown in the left side of FIG. 1 .

As shown on the right side of FIG. 1 , time-series decomposition of actual observed values is classified into trend, seasonal, and residual. Since a specific decomposition method is a known technique, a detailed description thereof will be omitted. Trends represent the average flow of sales. Seasonality is a pattern that repeats in a certain time unit, and excluding these patterns, residuals representing randomness remain. Among them, trend and seasonality are mainly applied to future time to predict product demand.

In this method, the more data, the higher the prediction accuracy. However, since it can be assumed that the trend and/or seasonality are similar even though the individual sales volume is different in the same product group, demand is predicted by summing up data of all products in the same product group to increase accuracy. For example, although individual sales volumes of "mineral water A" and "mineral water B" within the product group of "mineral water" are different, the trend and/or seasonality will be similar. Therefore, when forecasting the demand for "bottled water A" product, rather than forecasting the demand by time series decomposition of only the hourly sales volume of "bottled water A", after summing up the hourly sales volume of all bottled water such as "bottled water A" and "bottled water B" If demand is predicted by time series decomposition, the accuracy is higher.

Relatively speaking, if products within the same product group are analyzed together, the accuracy will increase, but the degree is not sufficient. Accordingly, a method of increasing the accuracy of demand forecasting by adding more data is being studied.

In addition, the above-described method only statistically processes hourly sales data, but as the number of data increases, statistical processing takes excessive time, and accuracy may not increase sufficiently due to data errors and randomness.

(Patent Document 1) Chinese Patent Publication No. 108133391

(Patent Document 2) Korea Patent Registration No. 10-2264013

(Patent Document 3) Korea Patent Registration No. 10-22645526

The present invention has been made to solve the above problems.

By improving the method of predicting demand by time-series decomposition of sales volume data by time, to forecast the demand for a product, not only the product group data of that product is used, but also products of other product groups with similar trends and/or seasonality can be used as data. By enabling this, we want to improve the accuracy of demand forecasting by utilizing more data.

In addition, we want to predict demand by applying transfer learning to increase the accuracy of the results, not just statistical analysis.

One embodiment of the present invention for solving the above problems, (A) the weighted correlation calculation equation creation module 210, in a predetermined method, creating a weighted correlation equation; (B) After the clustering classification module 220 time-series decomposition of the hourly sales data of each of a plurality of products into trend, seasonal, and residual, the trend for every product pair combination of the plurality of products The correlation of trend, correlation of seasonality, and correlation of residual are calculated and obtained, and by substituting them into the weighted correlation equation, the weighted correlation of all product pair combinations is obtained. obtaining; (C) The cluster classification module 220 calculates the average of all weighted correlations included in any one product among the plurality of products, and performs this for all products, so that the weighted correlation average for each product is calculated step; (D) classifying, by the cluster classification module 220, the plurality of products into one or more clusters by classifying products having the weighted correlation average within a preset range into the same cluster; and (E) the demand forecasting model generation module 230 sequentially transfers the sales data by time of all products classified into the same cluster with respect to all clusters classified in step (D), It provides a method comprising generating a demand forecasting model for a product corresponding to.

In addition, the step (A) may include: (a1) obtaining, by the weighted correlation equation creation module 210, sales data for each hour of a plurality of modeling target products; (a2) time-series decomposition of the obtained sales volume data by time by the weighted correlation equation preparation module 210 into trend, seasonality, and train, respectively; and (a3) preparing, by the module 210, a weighted correlation equation using the time-series decomposed data of modeling target products.

Further, the step (a3) may include: (a31) generating a first demand forecasting model by sequentially performing transfer learning on the time-series decomposed data of modeling target products by the transfer learning module 290; and (a32) the transfer learning module 290 calculates permutation importance for each of the time-series decomposed trend, seasonality, and residual using the first-order demand forecasting model, and the weighted correlation equation It is preferable that the preparation module 210 includes a step of preparing a weighted correlation calculation formula such as the following formula using this. The formula here is:

Weighted Correlation = Correlation of Trend * Permutation Importance of Trend + Correlation of Seasonality * Permutation Importance of Seasonality + Correlation of Residual * Permutation Importance of Seasonality

In addition, it is preferable that the products to be modeled in the step (a1) and the plurality of products in the step (B) are the same.

In addition, in the step (E), the demand forecasting model generation module 230 transfer learning module 290 performs transfer learning on the hourly sales volume data of all products in the cluster classified in step (D). ) preferably includes generating a demand forecasting model.

In addition, the step (a31) includes the step of generating the first demand forecasting model based on the random forest by the transfer learning module 290, and the step (a32) includes the step of the transfer learning module 290 generating the It is preferable to include calculating permutation importance based on random forest in the first demand forecasting model.

In addition, after the step (E), (F) inputting a predetermined product and a time to be predicted into the demand forecasting unit 120; (G) determining which cluster the input product is included in by the demand predicting unit 120; and (H) predicting the demand of the time to be predicted by the demand forecasting unit 120 by calling the demand prediction model generated for the identified cluster and applying the time to be predicted. it is desirable

In addition, after the step (H), it is preferable that the (I) ordering unit 130 generates product ordering information for ordering the corresponding product by using the demand for the product predicted in the step (H).

In addition, the store management unit 100 is linked with the ordering unit 140 and the sales information acquisition unit 110, and the sales information acquisition unit 110 acquires sales information including hourly sales volume data of the plurality of products. And, the step (a1) includes the step of acquiring sales data by time of each of the plurality of products from the sales information acquisition unit 110, and the step (B) includes the step of the cluster classification module 220 It is preferable to include obtaining data on sales volume by time of each of the plurality of products.

By applying the method according to the present invention, a cluster can be formed by selecting all products of different product groups that are expected to have different trends and/or seasonality at first glance, but actually have similar trends, seasonality and/or residuals. has much more data than the corresponding product group, so the accuracy of the demand forecasting model increases.

In addition, by applying weighted correlation, it is possible to assign appropriate weights to trend, seasonality, and residuals, so that the product data to be used is determined with high accuracy, and cluster formation is elaborately performed, thereby increasing the accuracy of demand forecasting. .

As a result of generating a model by transfer learning and applying the method according to the present invention, but predicting demand by applying weighted correlation, it was confirmed that the accuracy greatly increased as shown in FIG. 6.

1 is a diagram for explaining a prior art for predicting demand by time-series analysis of hourly sales volume data.
Figure 2 shows a system for carrying out the method according to the invention.
3 is a flowchart of a method according to the present invention.
4 is one of the verification test results of the method according to the present invention, and shows the result of deriving the importance of permutation characteristics.
Figure 5 shows the weighted correlation of each product pair combination as one of the verification test results of the method according to the present invention.
Figure 6 is one of the verification test results of the method according to the present invention, Figure 6a is a verification result of the transfer learning use demand prediction model to which the cluster concept is applied according to the present invention, Figure 6b is a transfer learning use demand targeting a single product This is the verification result of the predictive model.

Hereinafter, "product" means all items sold in stores.

Hereinafter, a “group” is a collection of products, and as a concept different from a product group of the same type, products of the same or different types may form one cluster. In addition, any one product must be included in any one cluster.

1. Description of the system

Referring to Fig. 2, a system for performing the method according to the present invention is described.

The system for performing the method according to the present invention includes a store management unit 100, a sales information acquisition unit 110, a demand forecasting unit 120 and an ordering unit 130, and further includes a control unit 200. .

The store management unit 100 functions to manage stores where products are sold, and may be hardware that executes any kind of store management program conventionally used. You can manage sales, place orders, and manage inventory.

The sales information acquisition unit 110 acquires sales information from the store management unit 100, and the acquired sales information includes sales data for each product at each hour. Among such sales information, hourly sales volume data of products requested by the control unit 200 is transmitted to the control unit 200 .

The demand forecasting unit 120 performs a function of predicting demand for a specific product during a specific time using the generated demand forecasting model.

The ordering unit 130 automatically generates ordering information for a specific product based on the demand information predicted by the demand predicting unit 120 and sends it to the ordering party. The ordering information is transmitted to the store management unit 100 .

The control unit 200 functions to generate a demand forecasting model, and details will be described later in the description of the method below. The method according to the present invention is a program, which is operated by the control unit 200 and stored in a storage medium so that the corresponding method is executed.

2. Description of the method

Referring to Fig. 3, the method according to the present invention is described.

The method according to the present invention introduces the concept of clustering in order to secure more data to increase accuracy. As described above, all products that are determined to have similar trends, seasonality, and/or residuals are set as clusters without being limited to similar product groups of specific products. Since the demand forecasting model is created by using the data of all products included in one cluster, the accuracy is increased compared to the case of using only the data of a specific product group.

In order to determine which products can be formed as one cluster, the first demand forecasting model is created, and then a weighted correlation formula is prepared and utilized using the importance of permutation characteristics. For each product, a weighted correlation will be calculated for all other products. If N products are used as data, weighted correlations are calculated as many as N-1, and clusters are classified using the weighted correlation average, which is the average value.

The commodities included in the classified clusters include all commodities that are judged to have similar trends, seasonality, and/or residuals. It is not limited to similar product groups. A demand forecasting model for a cluster is created using all of these product data. The products included in the cluster have the same demand forecasting model. A demand forecasting model is created for each cluster.

If there is a product and time for which demand is to be predicted, the sales volume is calculated by checking which cluster the product belongs to, loading the demand prediction model of that cluster, and inputting the product and time. The calculated sales volume is the demand forecast for the corresponding product at the corresponding time.

Each step is explained in detail.

2.1 Writing a weighted correlation formula using the first demand forecasting model

Select the product to be modeled. It can be any product sold in the store.

The weighted correlation formula preparation module 210 obtains sales data for each hour of a plurality of modeling target products. This may be obtained from sales information obtained from the sales information acquisition unit 110 .

Next, the weighted correlation calculation formula preparation module 210 decomposes the acquired sales volume data by time into trend, seasonality, and time series, respectively. Since the time series decomposition method is a prior art, a detailed description thereof will be omitted.

Next, the weighted correlation formula preparation module 210 creates a weighted correlation formula using the time-series decomposed data of modeling target products.

Specifically, the transfer learning module 290 sequentially transfers the data of the time-series decomposed modeling target products to generate a first demand forecasting model. Through the first-order demand forecasting model, the permutation importance for each of the time series decomposed trend, seasonality, and residual is calculated. The permutation feature importance will be calculated by a conventionally known method. By this operation, the scores of Permutation Importance of Trend, Permutation Importance of Seasonality, and Permutation Importance of Seasonality of residuals are confirmed. Now, the weighted correlation formula preparation module 210 uses this to create a weighted correlation formula such as the following formula. The correlation of trend, correlation of seasonality, and correlation of residual in the formula will be determined for each pair of products subject to clustering, which will be described later.

Weighted Correlation = Correlation of Trend * Permutation Importance of Trend + Correlation of Seasonality * Permutation Importance of Seasonality + Correlation of Residual * Permutation Importance of Seasonality

Here, the first demand forecasting model is created using transfer learning. For example, after generating a demand prediction model A by learning using product A, which is one of the products to be modeled, a demand prediction model B is created by further using data on product B, another product to be modeled. By continuously learning in this way, demand forecasting model A through demand forecasting model N will be generated for N products to be modeled, and the demand forecasting model N in this case becomes the primary demand forecasting model.

In addition, such transfer learning can be performed based on random forest. In this case, the permutation importance is also calculated based on the random forest.

On the other hand, there are no restrictions on the products to be modeled for creating such a weighted correlation calculation formula, but it is preferable in terms of accuracy that all products sold in stores, that is, all products for demand forecasting as described below.

2.2 Classification of clusters

Now, the clusters of all products can be classified using the weighted correlation formula.

First, the clustering classification module 220 decomposes the time-series sales data of each of a plurality of products sold in the store into trend, seasonality, and total, and then the trend correlation, seasonality correlation, and total correlation for all product pair combinations of the plurality of products. calculate Any method may be used to calculate the correlation. The correlation score ranges from -1 to 1, where -1 indicates a negative correlation and 0 indicates a positive correlation. If correlations are calculated for each product pair combination for any one product out of N products in this way, each of N-1 trend correlations, seasonal correlations, and street correlations will be calculated.

N-1 weighted correlations are calculated by substituting the N-1 trend correlations, seasonal correlations, and street correlations obtained in this way into the above-described formula, and dividing them by N-1 to calculate the weighted correlation average. In this way, a weighted correlation average is calculated and secured for all N products.

One product has one weighted correlation average. This score is a score representing the change in sales volume of the product. Therefore, clusters can be classified by dividing scores into a specific range. For example, it can be classified into a first cluster ranging from -0.05 to 0 points, a second cluster ranging from 0 to 0.5 points, and a third cluster ranging from 0.5 to 10 points. Specific numerical values for determining the range of division are different for each store, so specific numerical values are not limited in this specification.

In this way, one or more clusters are formed, and all products are classified to be included in one cluster.

2.3 Demand forecasting model generation

The demand forecasting model generation module 230 generates a demand forecasting model for each cluster with respect to all classified clusters. Check again the sales volume data by hour for all products classified into the same cluster, and create a demand forecasting model by sequentially transfer learning one product at a time as described above. Similarly, it can be random forest based.

2.4 Demand Forecast

The user inputs a product for which demand is to be predicted and a time to be predicted in the demand prediction unit 120 . The demand forecasting unit 120 first checks in which cluster the input product is included, and calls the generated demand forecasting model for the identified cluster.

When product and time are applied to the imported demand forecasting model, sales volume is output. The output sales volume is the amount predicted by the demand for the product at that time.

On the other hand, if the demand forecasting unit 120 predicts the demand for a specific product at a specific time, the ordering unit 130 may automatically generate order information for the corresponding product by comparing it with the current stock amount. The generated ordering information may be automatically transmitted to the ordering party through the store management unit 100 .

3. Verification experiment

A verification experiment was conducted to confirm the accuracy of the demand forecasting model generated by the above method.

Among the products traded at B Mart located in A region, 50 products were selected and the sales volume data by hour was used. Of these, 40 were used as learning materials and 10 were used as verification materials. That is, 40 modeling products were selected. Then, the permutation feature importance was calculated using the random forest method, and the calculation results were 3.85 for the correlation of trend, 14.59 for the correlation of seasonality, and 6.8 for the correlation of residual, as shown in FIG. calculated Accordingly, the weighted correlation equation was determined as follows.

Weighted Correlation = Correlation of Trend * 3.85 + Correlation of Seasonality * 14.59 + Correlation of Residual * 6.8

Using this formula, the weighted correlation was calculated by making a pair of different products for every 40 products. Thirty-nine weighted correlations were calculated for each product. Figure 5 shows the result. In addition, the average of 39 weighted correlation scores was calculated. The average weighted correlation ranged from -0.05 to 10.

Now that 40 weighted correlations are secured, clusters can be formed by classifying them according to a predetermined criterion. In one example, the products formed in the same cluster as Gayasan Thousand Years 2L (6 pieces) were as follows.

- 2400 Seolreim Milk, Gaya Millennial Water 500ml_20, Southern) Bulgogi Flavor Frank Sausage, Southern) Oryukdo Flavor Bar, Wildflower Scent Wet Tissue Refill, Unpasteurized Bulgogi, B) Nougat Bar Gold, B) Pork Bar, B) Megaton, B) Babamba Gold, bar) Bibbic, bar) Seoul Melon Bar, bar) Watermelon Bar, f) Screw Bar, bar) Ssangssang Bar, b) Amatna, b) Okdongja, b) Joe’s Bar, bar) Quan Kva Choco, Bag King Oyster Mushroom (1) Mouth, 400g_Special_Bong_Domestic), New) Chamisul fresh, New) Cass fresh bottle, Jeju Samdasoo, Clean) Eco-friendly 60 eggs, Cas fresh can, Corn) Double Byanco, Corn) Burabo Vanilla, Corn) Breadpare Vanilla, Corn) Breadpare Chocolate, Corn) World Corn

In other words, results included in the same cluster even though they are not the same product group were derived.

For the above clusters, transfer learning was conducted again to finally complete the demand forecasting model.

Valid data prediction was performed using the remaining 10 products in the final completed demand forecasting model and compared with actual data. The result is shown in the upper figure of FIG. 6A, blue is actual data and red is predicted data. In addition, test data prediction for other random products was carried out. The result is shown in Fig. 6a lower figure.

In order to compare with this method, a demand forecasting model limited to a single product was formed without forming a cluster, and the same prediction was made. The valid data prediction result is shown in the upper figure of FIG. 6B, and the test data prediction result is shown in the lower figure of FIG. 6B. The statistical analysis results are shown in the table below.

division Transfer learning model applying cluster concept Single transfer learning model valid data test data valid data test data rmse 92.53 88.83 82.62 92.08 mae 73.26 66.92 63.88 65.18 r2 0.64 0.56 0.71 0.53 mape 0.51 0.14 0.49 0.15

Statistically, the accuracy of the transfer learning model itself was confirmed to be high in any case. However, when using a single transfer learning model, some exceptional data points with a very large difference between the predicted value and the actual value were identified. It was confirmed that the method application model according to the present invention to which the cluster concept was applied was more stable and had a statistically smaller difference from the actual value, resulting in higher accuracy.

100: store management department
110: sales information acquisition unit
120: demand forecasting unit
130: ordering department
200: control unit
210: weighted correlation formula creation module
220: cluster classification module
230: Demand forecasting model creation module
290: transfer learning module

Claims (10)

(A) preparing, by the weighted correlation equation creation module 210, a weighted correlation equation using a preset method;
(B) After the clustering classification module 220 time-series decomposition of the hourly sales data of each of the plurality of products into trend, seasonal, and residual, the trend for every product pair combination of the plurality of products The correlation of trend, correlation of seasonality, and correlation of residual are calculated and obtained, and by substituting them into the weighted correlation equation, the weighted correlation of all product pair combinations obtaining;
(C) The cluster classification module 220 calculates the average of all weighted correlations including any one product among the plurality of products, performs this for all products, and calculates the weighted correlation average for each product step;
(D) classifying, by the cluster classification module 220, the plurality of products into one or more clusters by classifying products having the weighted correlation average within a preset range into the same cluster; and
(E) For all clusters classified in step (D), the demand forecasting model generation module 230 sequentially transfers the sales data by time of all products classified in the same cluster to the corresponding cluster. Including the step of generating a demand forecasting model for the corresponding product,
method.
According to claim 1,
In step (A),
(a1) obtaining, by the weighted correlation formula preparation module 210, sales data for each of a plurality of modeling target products by time;
(a2) time-series decomposition of the obtained sales volume data by time by the weighted correlation equation preparation module 210 into trend, seasonality, and train, respectively; and
(a3) comprising the step of the weighted correlation equation creation module 210 creating a weighted correlation equation using the data of the time-series decomposed modeling target products,
method.
According to claim 2,
In the step (a3),
(a31) generating, by the transfer learning module 290, a first demand forecasting model by sequentially performing transfer learning on the time-series decomposed data of modeling target products; and
(a32) The transfer learning module 290 calculates the permutation importance of each of the trend, seasonality, and residual decomposed time series using the first demand forecasting model, and prepares the weighted correlation equation Module 210 uses this to create a weighted correlation equation such as the following equation,
Weighted Correlation = Correlation of Trend * Permutation Importance of Trend + Correlation of Seasonality * Permutation Importance of Seasonality + Correlation of Residual * Permutation Importance of Seasonality
method.
According to claim 2,
The products to be modeled in step (a1) and the plurality of products in step (B) are the same,
method.
According to claim 3,
In the step (E),
Generating, by the transfer learning module 290, the demand forecast model generation module 230, the demand forecast model, by transfer learning the hourly sales volume data of all products in the cluster classified in the step (D). including,
method.
According to claim 5,
In the step (a31),
The transfer learning module 290 includes generating a first demand forecasting model based on a random forest,
In the step (a32),
Comprising the transfer learning module 290 calculating permutation importance based on random forest in the first demand forecasting model,
method.
According to claim 4,
After the step (E),
(F) inputting a predetermined product and a time to be predicted into the demand forecasting unit 120;
(G) determining which cluster the input product is included in by the demand predicting unit 120; and
(H) further comprising the step of predicting the demand of the time to be predicted by the demand forecasting unit 120 calling the demand prediction model generated for the identified cluster and applying the time to be predicted,
method.
According to claim 6,
After the step (H),
(I) The ordering unit 130 generates product ordering information for ordering the product by using the demand for the product predicted in step (H),
method.
According to claim 8,
The store management unit 100 is linked with the ordering unit 140 and the sales information acquisition unit 110, and the sales information acquisition unit 110 acquires sales information including hourly sales volume data of the plurality of products,
The step (a1) includes acquiring sales data for each hour of each of the plurality of products from the sales information acquisition unit 110,
The step (B) includes the step of obtaining, by the cluster classification module 220, sales volume data by time of each of the plurality of products,
method.
A program stored in a storage medium and operated by the control unit 200 so that the method according to any one of claims 1 to 9 is performed.

Images