An intelligent ordering method for Nissin fresh food stores

Through the deep neural network model, multi-scale prediction of customer demand in Nissin fresh food stores has been solved, and the optimal ordering and economic benefits have been achieved.

CN113869938BActive Publication Date: 2024-11-08HANGZHOU MINGXIN INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111054206.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-09
Publication Date
2024-11-08
Estimated Expiration
2041-09-09

AI Technical Summary

Technical Problem

Due to the lack of inventory and the need to control the total order quantity and structure of Nissin fresh food stores, the existing technology single-item prediction model cannot guarantee the rationality of orders, resulting in unbalanced orders, which takes a long time and suffers huge losses.

Method used

The deep neural network model is used to predict the customer demand in three scales (total order amount, proportion of order amount of each product category and proportion of order amount of single product in major categories), and the single product order amount is determined through formulas. Combining store characteristics, date characteristics and weather characteristics, LSTM and MLP models are built for high-precision prediction.

Benefits of technology

The optimal ordering of Nissin fresh food stores is achieved, ensuring the rationality of the total order quantity and structure, improving economic benefits, and reducing labor time and losses.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113869938B_ABST
    Figure CN113869938B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of deep learning, and discloses a method for intelligent ordering of Nissin fresh food stores. A method for intelligent ordering of Nissin fresh food stores, which uses three deep learning models based on LSTM long short-term memory recurrent neural networks to predict customer demand at three scales: the total order amount of the store, the proportion of the order amount of each major commodity category and the number of SKUs of each major commodity category, and the proportion of the order amount of a single product in the major category, and determines the final order quantity of a single product. Different from the prediction of a single product on a single scale, the present invention can ensure the rationality of the total order quantity and the order structure; and introduces the key feature of the sold-out time into the deep learning model, realizes the optimal ordering method, and improves the economic benefits of the store.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and in particular to an intelligent ordering method for a Nissin fresh food store. Background Art

[0002] A new model has emerged in community fresh food stores, namely the daily clearance model. In this model, the goods received by the store on the same day must be sold out on the same day. In order to achieve daily clearance, the store starts to discount the unsold goods at 19:00 in the evening, with a 10% discount at 19:00, an 20% discount at 19:30, and a 30% discount at 20:00. The discounts continue to decrease until they are free. This daily clearance model gives consumers the intuitive feeling of "not selling overnight vegetables" and allows consumers to experience the true ultimate freshness.

[0003] After the community fresh food stores adopt the daily clearance model, they need to place orders every day. Generally, they place orders two days in advance, that is, if you place an order today, the goods will arrive the day after tomorrow. This model makes daily ordering a huge challenge. If you order too much, it will result in losses in the evening, and if you order too little, you will miss sales opportunities.

[0004] Community fresh food stores under the Nissin model (referred to as Nissin stores) have two major differences compared with other fresh food supermarkets and convenience stores.

[0005] First, Nissin stores do not carry inventory, so they cannot order from a safety stock perspective like in the traditional replenishment model.

[0006] Second, Nissin stores must control the total order quantity and the overall order structure, rather than just considering the order quantity of a single product (single product or single SKU). This is because the total daily demand and purchasing power of the customer base covered by each store are limited. If a Nissin store orders too much of a certain vegetable on a certain day, then other vegetables or even other products must be ordered less to keep the total quantity reasonable. Traditional fresh supermarkets and convenience stores are not strict in controlling the total quantity because they can stock up. They only need to ensure that the quantity of each single product is relatively sufficient.

[0007] The ordering systems in the prior art are all single-product forecasts on a single scale, which cannot guarantee the rationality of the total order quantity and order structure.

[0008] For example, in the prior art, the patent application number is CN201811599892.6, the patent application date is 2018-12-26, and the patent name is a smart ordering system for fresh food stores, including: a product sales forecasting module, which is used to determine the time point of discount promotions and predict the sales of each product before discount promotions; an ordering module, which is used to determine the actual order quantity of each product; and a promotion module, which is used to obtain the time point of discount promotions and the discount amount of the product after the time point of discount promotions from the product sales forecasting module. The ordering system for fresh food stores in this technology is actually based on single product predictions, which cannot be used in the daily clearance mode and will cause order imbalance.

[0009] In the past, Nissin stores’ ordering decisions mostly relied on the personal experience of the store manager, which had the following shortcomings: lack of quantitative analysis, inability to achieve optimal ordering; long ordering time per day, requiring a lot of manpower time to be wasted; there were cases of forgetting to order, resulting in huge losses; large individual differences, and the store manager’s experience could not be replicated. Therefore, an intelligent ordering model is urgently needed to solve these drawbacks and ensure the healthy and stable development of stores. Summary of the invention

[0010] In view of the problem that the existing technology cannot achieve optimal ordering for daily fresh food stores, the present invention provides an intelligent ordering method for daily fresh food stores. By using a deep neural network model to predict customer demand on three scales, the rationality of the total order amount and the ordering structure is solved, the optimal ordering method is achieved, and the economic benefits of the store are improved.

[0011] In order to achieve the above object, the present invention adopts the following technical solutions:

[0012] A method for intelligent ordering of fresh food stores in Nissin, the method comprising:

[0013] Demand forecasting: using three deep neural network models to forecast customer demand at three scales; the three-scale forecasts include the forecast of the total store order amount, the proportion of order amount in each major commodity category and the number of SKUs in each major commodity category, and the forecast of the proportion of single product order amount in the major category;

[0014] Determine the order quantity of a single product according to Formula 1;

[0015] N k =round(G*C k *Z k / E k ) Formula 1

[0016] Among them, N k is the order quantity of item k, G is the total order amount, C kis the proportion of the category to which single product k belongs in the total order amount, Z k is the proportion of the order amount of the single product in the category to which it belongs, E k is the unit specification order amount of single product k, and round() means rounding to an integer.

[0017] Three LSTM (Long Short-Term Memory Recurrent Neural Network) based deep neural network models are used to predict customer demand at three scales; these three deep neural network models are respectively recorded as deep neural network model 1, deep neural network model 2 and deep neural network model 3; deep neural network model 1 is used to predict the total order amount of the store, deep neural network model 2 is used to predict the proportion of order amount of each major commodity category and the number of SKUs of each major commodity category, and deep neural network model 3 is used to predict the proportion of single product order amount in the major category; thus determining the order quantity of a single product;

[0018] As a preferred method, the prediction of the total order amount of a store, the proportion of the order amount of each major commodity category and the number of SKUs of each major commodity category, and the prediction of the proportion of the order amount of a single product in the major category include store feature vectors, date feature vectors, and weather features;

[0019] Input the time series of the store's daily sold-out time, gross profit, transaction flow, and customer order composition from T-14 to T-1 into the LSTM model to determine the store feature vector;

[0020] Perform one-hot encoding on the seven days of the week to obtain the day of the week code corresponding to T+N days; perform one-hot encoding on whether it is a holiday to obtain the holiday code corresponding to T+N days; perform one-hot encoding on whether to make up for the class to obtain the make up for the class code corresponding to T+N days; thus determine the date feature vector;

[0021] Get the temperature of day T+N to determine the weather characteristics.

[0022] As a preferred method, the prediction of the total order amount of the store also uses the historical order feature vector and the T+N day store sell-out time.

[0023] The total daily order amount G from T-14 to T+N-1 T The constructed time series is input into the LSTM model to determine the historical order feature vector;

[0024] Where T represents the order date, N represents ordering N days in advance, and T+N-1 is the day before the arrival date;

[0025] Get the store sold-out time of T+N days.

[0026] Preferably, the method for determining the total order amount includes:

[0027] Input the historical order feature vector, the store's sold-out time feature, the store feature vector, the date feature vector, and the weather feature into the MLP multi-layer perceptron, and output the total order amount for T+N days;

[0028] The model is trained and tested based on historical order data, and the model with the highest test accuracy is selected to determine the total order amount.

[0029] As a preferred method, the major category coding feature vector, major category historical order feature vector and major category sold-out time are also used in the prediction of the order amount proportion of each major category and the number of SKUs of each major category.

[0030] Perform one-hot encoding on the major categories and determine the major category encoding feature vector;

[0031] The time series consisting of daily major category order amounts and major category order SKU numbers from days T-14 to T+N-1 is input into the LSTM model to determine the major category historical order feature vector;

[0032] Get the sold-out time of major categories in T+N days.

[0033] As a preferred method, the method of determining the proportion of the order amount of each category includes:

[0034] Input the major category coding feature vector, major category historical order feature vector, major category sold-out time feature, store feature vector, date feature vector and weather feature into MLP, and output the T+N day major category direct prediction order amount and major category SKU number;

[0035] Based on historical data, the model is trained and tested, and the model with the highest test accuracy is selected to predict the order amount and SKU number of major categories, thereby determining the proportion of order amount of major categories;

[0036] The proportion of major order amounts is determined according to Formula 2:

[0037] C i =B i / sum i (B i ) Formula 2

[0038] Among them, C i is the proportion of order amount of category i, B i is the direct forecast order amount of category i, sum i Indicates the sum of all categories.

[0039] As a preferred method, the prediction of the proportion of single product order amount in the major categories also uses single product coding feature vector, single product historical order feature vector and single product sold-out time;

[0040] Perform embedding encoding on all items in the order product pool to determine the item encoding feature vector;

[0041] The time series consisting of the order amount of each item from T-14 to T+N-1 is input into the LSTM model to determine the historical order feature vector of the item;

[0042] Get the T+N days single product sold out time.

[0043] Preferably, the method for determining the proportion of the order amount of a single product in a large category includes:

[0044] Input the single product code feature vector, single product historical order feature vector, single product sold-out time feature, store feature vector, date feature vector and weather feature into MLP, and output the direct predicted order amount of the single product in T+N days;

[0045] Based on historical data, the model is trained and tested, and the model with the highest test accuracy is selected to predict the order amount of a single product;

[0046] Determine whether the number of single products with a direct predicted amount greater than zero meets the constraints of the number of SKUs in the major category, control the number of single products within the constraint range, and thus determine the proportion of single products in the major category;

[0047] Determine the proportion of a single product in a major category based on Formula 3:

[0048] Z i =S i / sum i (S i ) Formula 3

[0049] Where Z i is the proportion of order amount of single product i in its category, S i is the direct predicted order amount of item i, sum i Indicates the sum of all items in the category to which the item belongs.

[0050] As a preference, the number of items is constrained according to the number of major SKUs, and the directly predicted order amounts of the items are sorted from large to small. If the number of items exceeds the number of major SKUs, the items that are ranked low and are not necessary to order are eliminated.

[0051] As a preferred method, the sold-out time is divided into the sold-out time during model training and model testing, and the sold-out time during model prediction. The sold-out time during model training and testing is the actual sold-out time, and the T+N-day sold-out time during model prediction is the expected optimal sold-out time; the optimal sold-out time is the time when the product is automatically discounted to 0 gross profit margin in the evening.

[0052] The present invention has significant technical effects due to the adoption of the above technical solution:

[0053] The present invention uses a deep neural network to predict the real daily demand of customers on three scales, including the total order amount, the proportion of each major category order amount, and the proportion of each major category SKU number and single product order quantity. The single product order quantity obtained thereby ensures the rationality of the total order amount and order structure;

[0054] The introduction of sold-out time can better reflect the objective and real user demand. The best order amount can be inferred based on the best sold-out time;

[0055] High-precision prediction is achieved through the reasonable selection and processing of deep learning feature vectors and the construction of deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a complete flow chart of order quantity calculation of the present invention.

[0057] Figure 2 It is a structural diagram of the store total demand prediction model of the present invention.

[0058] Figure 3 It is a structural diagram of the prediction model of major category order amount and major category SKU number of the present invention.

[0059] Figure 4 It is a structural diagram of the single product order amount prediction model of the present invention. DETAILED DESCRIPTION

[0060] The following is combined with Figure 1 To Attachment Figure 4 The present invention is further described in detail with examples.

[0061] Example 1

[0062] A smart ordering method for daily fresh food stores. The order quantity of a single product is determined according to formula 1.

[0063] N k =round(G*C k *Z k / E k ) Formula 1

[0064] Among them, N k is the order quantity of item k, G is the total order amount of the store, C k is the proportion of the category to which single product k belongs in the total order amount, Z k is the proportion of the order amount of the single product in the category to which it belongs, E k is the unit specification order amount of single product k, round() means rounding to an integer; G, C k ,Zk The predictions were obtained by deep neural network model one, deep neural network model two, and deep neural network model three respectively.

[0065] Determine the total order amount using a deep neural network model, including:

[0066] The total daily order amount G from T-14 to T+N-1 T The constructed time series is input into the LSTM model to determine the historical order feature vector; T represents the order date, N represents ordering N days in advance, and T+N-1 is the day before the arrival date;

[0067] Get the sold-out time of the entire store on T+N days;

[0068] Input the time series of daily sold-out time, gross profit, transaction flow, and customer order composition from T-14 to T-1 into the LSTM model to determine the store feature vector;

[0069] Perform one-hot encoding on the seven days of a week to obtain the day of the week code corresponding to T+N days; perform one-hot encoding on whether it is a holiday to obtain the holiday code corresponding to T+N days; perform one-hot encoding on whether to make up for the class to obtain the make up for the class code corresponding to T+N days; thus obtain the date feature vector;

[0070] Get the temperature of day T+N to determine the weather characteristic vector;

[0071] Input the historical total order feature vector, the store sell-out time feature, the store feature vector, the date feature vector and the weather feature into the MLP, and output the total order amount for T+N days.

[0072] Substitute historical data into the above model, train and test the model, and select the model with the highest test accuracy to predict the total order amount.

[0073] Use the deep neural network model 2 to determine the proportion of order amounts for each major category of goods, including:

[0074] Perform one-hot encoding on the major categories and determine the major category encoding feature vector;

[0075] The time series consisting of daily major category order amounts and major category order SKU numbers from days T-14 to T+N-1 is input into the LSTM model to determine the historical major category order feature vector;

[0076] Get the sold-out time of major categories in T+N days;

[0077] Input the major category code feature vector, the historical major category order feature vector, the major category sold-out time feature vector, the store feature vector, the date feature vector, and the weather feature vector into the MLP, and output the major category order amount and the major category SKU number for day T+N;

[0078] Substitute historical data into the above model, train and test the model, and select the model with the highest test accuracy to predict the order amount and SKU number of major categories; thus determine the proportion of order amount of major categories;

[0079] Determine the proportion of major order amounts based on Formula 2.

[0080] C i =B i / sum i (B i ) Formula 2

[0081] Among them, C i is the proportion of order amount of category i, B i is the direct forecast order amount of category i, sum i Indicates the sum of all categories.

[0082] Use deep neural network model 3 to determine the proportion of order amounts of single products in major categories, including:

[0083] Embed the single product and determine the single product coding feature vector;

[0084] The time series of daily single product order amounts from days T-14 to T+N-1 is input into the LSTM model to determine the historical single product order feature vector;

[0085] Get the T+N days single product sold out time;

[0086] Input the single product code feature vector, historical single product order feature vector, single product sold-out time feature vector, store feature vector, date feature vector and weather feature vector into MLP, and output the single product order amount for T+N days.

[0087] Substitute historical data into the above model, train and test the model, select the model with the highest test accuracy, and predict the order amount of a single product;

[0088] Determine whether the single product prediction meets the constraints of the number of major SKUs, and sort the single product order amounts predicted by the model from large to small. If the number of non-zero single products exceeds the number of major SKUs, the low-ranking and non-must-order products will be eliminated.

[0089] Determine the proportion of a single product in its category using Formula 3:

[0090] Z i =S i / sum i (S i ) Formula 3

[0091] Where Zi is the proportion of order amount of single product i in its category, S i is the direct predicted order amount of item i, sum i Indicates the sum of all items in the category to which the item belongs.

[0092] The above-mentioned sold-out time includes the sold-out time during model training and testing, as well as the sold-out time during model prediction. The sold-out time during model training and testing is the actual sold-out time, which can be obtained from historical receipt data. However, in the prediction process, we need to give the best sold-out time for T+N days. The best sold-out time is the time when the gross profit margin is 0 after the discount in the evening. The gross profit margin of the whole store and the gross profit margin of each category can be calculated by combining the prediction results of the proportion of each category and the proportion of each single product with the original gross profit margin of each single product.

[0093] Except for the encoded features, all other features need to be cleaned of outliers, filled with missing values, and standardized before entering the model.

[0094] Example 2

[0095] Based on Example 1, the product coding of this embodiment adopts 64-dimensional embedding coding. Embedding coding can handle thousands of products well. Compared with one-hot coding, embedding coding effectively reduces the characteristic dimension and can well characterize the similarity between different products.

[0096] Example 3

[0097] Based on the above embodiment, this embodiment adopts the T+2 ordering mode, that is, if you order today, the goods will arrive the day after tomorrow. T-14 means the 14th day before the ordering day, T-1 means the 1st day before the ordering day, T+1 means the day after the ordering day, and T+2 means two days after the ordering day, that is, the arrival day.

[0098] Example 4

[0099] On the basis of the above embodiment, the LSTM module in this embodiment adopts a two-layer structure, and the dropout parameter between the two layers is 0.5 to prevent overfitting, thereby improving the accuracy of prediction.

[0100] Example 5

[0101] On the basis of the above embodiment, the MLP module in this embodiment adopts a two-layer structure, and the dropout parameter between the two layers is 0.4 to prevent overfitting, thereby improving the accuracy of prediction.

[0102] Example 6

[0103] Based on the above embodiment, the deep neural network model in this embodiment is implemented using pytorch, and NVIDIA A100 GPU is used for model training, model testing and model prediction.

[0104] Example 7

[0105] Based on the above embodiment, in actual application, the operating data, date data and weather data of 150 stores in the past year are obtained. All input and output features of a store in one day constitute a data point. These data points are divided into training set, validation set and test set. The training set data is input into the model in a random batch manner for training, and the training results are cross-validated with the validation set data. Finally, the test is performed on the test set, and the model with the highest test accuracy is selected as the prediction model.

[0106] Example 8

[0107] Based on the above embodiment, this embodiment uses a variant of LSTM, the GRU model, to replace the LSTM model, which has a faster training speed.

[0108] Example 9

[0109] Based on the above embodiment, the major categories of this embodiment include dried miscellaneous items, aquatic products, fruits, beef, mutton, pork, poultry, vegetables, and eggs.

Claims

1. A smart ordering method for Nissin fresh food stores. Features: Methods include, Prediction of customer demand: using three deep neural network models to predict customer demand at three scales; The three-scale forecasts include the forecast of the total store order amount, the forecast of the proportion of the order amount of each major product category and the number of SKUs in each major product category, and the forecast of the proportion of the order amount of a single product in the major category; Determine the order quantity of a single product according to Formula 1; N k =round(G*C k *Z k / E k )Formula 1 Among them, N k is the order quantity of item k, G is the total order amount, C k is the proportion of the category to which single product k belongs in the total order amount, Z k is the proportion of the order amount of the single product in the category to which it belongs, E k is the order amount per unit of product k, round() means rounding to an integer; Methods for determining the total order amount include: Input the historical order feature vector, the store's sold-out time feature, the store feature vector, the date feature vector, and the weather feature into the MLP multi-layer perceptron, and output the total order amount for T+N days; Based on historical data, the model is trained and tested, and the model with the highest test accuracy is selected to predict the total order amount; Methods for determining the proportion of major order amounts include: Input the major category coding feature vector, major category historical order feature vector, major category sold-out time feature, store feature vector, date feature vector and weather feature into MLP, and output the major category order amount and major category SKU number for day T+N; Based on historical data, the model is trained and tested, and the model with the highest test accuracy is selected to predict the order amount and SKU number of major categories, thereby determining the proportion of order amount of major categories; The proportion of major order amounts is determined according to Formula 2: C i =B i / sum i (B i )Formula 2 Among them, C i is the proportion of order amount of category i, B i is the predicted order amount for category i, sum i It means summing up all major categories; Methods for determining the proportion of single product order amount in a major category include: Input the single product code feature vector, single product historical order feature vector, single product sold-out time feature, store feature vector, date feature vector and weather feature into MLP, and output the single product order amount on T+N days, where T represents the order date, N represents ordering N days in advance, and T+N is the arrival date; Based on historical data, the model is trained and tested, and the model with the highest test accuracy is selected to predict the order amount of a single product; Determine whether the number of single products with a direct predicted amount greater than zero meets the constraints of the number of SKUs in the major category, control the number of single products within the constraint range, and thus determine the proportion of single products in the major category; Determine the proportion of a single product in a major category based on Formula 3: Z i =S i / sum i (S i )Formula 3 Where Z i is the proportion of order amount of single product i in its category, S i is the direct forecast order amount of item i, sum i Indicates the sum of all items in the category to which the item belongs.

2. According to claim 1, a method for intelligent ordering of fresh food stores in Nissin, Features: The store feature vector, date feature vector and weather feature are used in the prediction of the total store order amount, the proportion of order amount in each major product category, the number of SKUs in each major product category, and the proportion of single product order amount in the major category; Input the time series of the store's daily sold-out time, gross profit, transaction flow, and customer order composition from T-14 to T-1 into the LSTM model to determine the store feature vector; Perform one-hot encoding on the seven days of the week to obtain the day of the week code corresponding to T+N days; Perform one-hot encoding on whether it is a holiday and obtain the holiday code corresponding to T+N days; Perform one-hot encoding on whether to make up the class, and obtain the make-up class code corresponding to T+N days; thus determining the date feature vector; Get the temperature of day T+N to determine the weather characteristics.

3. According to claim 2, a method for intelligent ordering in a Nissin fresh food store, Features: The prediction of the total order amount of the store also uses the historical order feature vector and the store's sell-out time. The total daily order amount G from T-14 to T+N-1 T The constructed time series is input into the LSTM model to determine the historical order feature vector; Get the store sold-out time of T+N days.

4. According to claim 2, a method for intelligent ordering in a Nissin fresh food store, Features: The prediction of the order amount proportion of each major category and the number of SKUs of each major category also uses the major category coding feature vector, the major category historical order feature vector and the major category sold-out time. Perform one-hot encoding on the major categories and determine the major category encoding feature vector; The time series consisting of daily major category order amounts and major category order SKU numbers from days T-14 to T+N-1 is input into the LSTM model to determine the major category historical order feature vector; Get the sold-out time of major categories in T+N days.

5. According to claim 2, a method for intelligent ordering in a Nissin fresh food store, Features: The prediction of the proportion of single product order amount in the major categories also uses the single product code feature vector, single product historical order feature vector and single product sold-out time; Embed all items in the order product pool to determine the item encoding feature vector; The time series consisting of the order amount of each item from T-14 to T+N-1 is input into the LSTM model to determine the historical order feature vector of the item; Get the T+N days single product sold out time.

6. According to claim 1, a method for intelligent ordering in a Nissin fresh food store, Features: Constrain the number of items according to the number of major SKUs, and sort the directly predicted order amounts of the items from large to small. If the number of items exceeds the number of major SKUs, the items that are ranked low and are not necessary to order will be eliminated.

7. A method for intelligent ordering of fresh food stores in Nissin according to claim 2, Features: The sold-out time is divided into the sold-out time during model training and model testing, and the sold-out time during model prediction. The sold-out time during model training and testing is the actual sold-out time; the T+N day sold-out time during model prediction is the expected optimal sold-out time.