CN113869938A - Intelligent ordering method for Nisshink fresh stores - Google Patents

Intelligent ordering method for Nisshink fresh stores Download PDF

Info

Publication number
CN113869938A
CN113869938A CN202111054206.9A CN202111054206A CN113869938A CN 113869938 A CN113869938 A CN 113869938A CN 202111054206 A CN202111054206 A CN 202111054206A CN 113869938 A CN113869938 A CN 113869938A
Authority
CN
China
Prior art keywords
class
amount
ordering
order
days
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111054206.9A
Other languages
Chinese (zh)
Inventor
熊礼平
顾宏光
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Mingxin Information Technology Co ltd
Original Assignee
Hangzhou Mingxin Information Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Mingxin Information Technology Co ltd filed Critical Hangzhou Mingxin Information Technology Co ltd
Priority to CN202111054206.9A priority Critical patent/CN113869938A/en
Publication of CN113869938A publication Critical patent/CN113869938A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G06Q30/0203—Market surveys; Market polls
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0207—Discounts or incentives, e.g. coupons or rebates
    • G06Q30/0224—Discounts or incentives, e.g. coupons or rebates based on user history
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0207—Discounts or incentives, e.g. coupons or rebates
    • G06Q30/0235—Discounts or incentives, e.g. coupons or rebates constrained by time limit or expiration date
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S50/00—Market activities related to the operation of systems integrating technologies related to power network operation or related to communication or information technologies
    • Y04S50/14—Marketing, i.e. market research and analysis, surveying, promotions, advertising, buyer profiling, customer management or rewards

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Development Economics (AREA)
  • Accounting & Taxation (AREA)
  • Physics & Mathematics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to the field of deep learning and discloses an intelligent ordering method for a Nisshinbo fresh shop. An intelligent ordering method for a daily fresh-keeping store is characterized in that three deep learning models based on an LSTM long-short term memory recurrent neural network are used for respectively predicting three scales of customer demand, total ordering amount of the store, the ratio of the ordering amount of each large class of commodities, the number of SKUs of each large class of commodities and the ratio of the ordering amount of a single commodity in the large class, and determining the final ordering amount of the single commodity. The invention is different from single-product prediction on a single scale, and can ensure the rationality of the total order quantity and the order structure; and the key characteristic of sold-out time is introduced into the deep learning model, so that an optimal ordering mode is realized, and the economic benefit of stores is improved.

Description

Intelligent ordering method for Nisshink fresh stores
Technical Field
The invention relates to the field of deep learning, in particular to an intelligent ordering method for a fresh store in Nisshink.
Background
A new mode, namely a daybreak mode, appears in the current community fresh store. In this mode, the store must empty the day's incoming items. To achieve daytime cleanliness, the store begins to discount unsold-out goods at 19 pm, 19: 00, 9 folds, 19: 30, 8-fold, 20: and 00, 7 folds are made, and the folds are continuously made until the folds are free. The day-to-night dish selling mode brings intuitive feeling of not selling overnight dishes to consumers, and the consumers experience real extreme freshness.
After the fresh shop of the community adopts the daily cleaning mode, the shop needs to order every day. The order is generally made two days in advance, that is, the order is made the day after today. The mode makes daily ordering face a great challenge, more orders can cause damage at night, and less orders can miss sales opportunities.
The community fresh shop (abbreviated as the Nissin shop) in the Nissin mode has two differences compared with other fresh supermarkets and convenience stores.
First, the day store has no inventory and therefore cannot order from a safe inventory perspective as in the traditional replenishment model.
Second, the daily store clearance must control the overall order size and overall order structure, rather than just from the point of view of the order size for a single item (referred to as a single item or single SKU). This is because the total daily demand and purchasing power of the customer base covered by each store is limited. If a day-to-day store orders more for a certain vegetable on a certain day, then other vegetables and even other products must be ordered less to keep the total amount reasonable. The conventional fresh supermarket and convenience store can be stocked, so that the total amount is not strictly controlled, and the quantity of each single product is ensured to be relatively sufficient.
The ordering system in the prior art is used for single-product prediction on a single scale, and the rationality of the total ordering quantity and the ordering structure cannot be ensured.
For example, in the prior art, patent application No. CN201811599892.6, patent application No. 2018-12-26, and patent name "a fresh shop intelligent ordering system", includes: the commodity sales predicting module is used for determining the time point of discount sales promotion and predicting the sales of each commodity before discount sales promotion; the ordering module is used for determining the actual ordering quantity of each commodity; and the sales promotion module is used for acquiring the time point of discount sales promotion and the discount number of the commodity after the time point of discount sales promotion from the commodity sales volume prediction module. The ordering system for the fresh shop in the technology is predicted on the basis of single products, so that the ordering system cannot be used in a day-to-day mode, and the ordering system can cause unbalanced ordering.
The conventional order decision of the Nisshink store mostly depends on personal experience of a store keeper, and has the following defects: lack of quantitative analysis and failure to make optimal orders; the daily ordering is long, and a large amount of labor time needs to be wasted; the situation that the order is forgotten causes huge loss; individual differences are large and store keeper experience is not reproducible. Therefore, an intelligent ordering model is urgently needed to solve the defects and ensure the healthy and stable development of stores.
Disclosure of Invention
Aiming at the problem that the prior art can not optimally order the daily fresh store, the invention provides the daily fresh store intelligent ordering method, which solves the problems of the total ordering amount and the rationality of the ordering structure by predicting the demand of a customer on three scales by using a deep neural network model, realizes an optimal ordering mode and improves the economic benefit of the store.
In order to achieve the purpose, the invention adopts the following technical scheme:
an intelligent ordering method for a fresh-keeping daily store, which comprises the steps of,
forecasting the demand, namely forecasting the demand of the customer on three scales by using three deep neural network models; the three-scale prediction comprises the prediction of the total ordering amount of the store, the prediction of the order amount proportion of each commodity and the number of SKUs of each commodity in the large class, and the prediction of the proportion of the single-product order amount in the large class;
determining the order quantity of the single products, namely determining the order quantity of the single products according to a formula 1;
Nk＝round(G*Ck*Zk/Ek) Equation 1
Wherein N iskThe number of orders to be made for the single item k, the total order amount G, CkIs the ratio of the main class to which the single item k belongs to the total order amount, ZkFor the order-amount ratio of the individual articles in the class to which they belong, EkThe round () representation is rounded to an integer for the unit specification order amount for singleton k.
Predicting the demand of customers on three scales by using three LSTM (long-short term memory recurrent neural network) -based deep neural network models; the three deep neural network models are respectively marked as a deep neural network model I, a deep neural network model II and a deep neural network model III; the deep neural network model I is used for predicting the total ordering amount of the store, the deep neural network model II is used for predicting the order amount proportion of each commodity and the SKU number of each commodity class, and the deep neural network model III is used for predicting the proportion of the single-item order amount in the classes; thereby determining the order quantity of the single goods;
preferably, the prediction of the total ordering amount of the store, the ratio of the ordering amount of each large class of commodities, the number of the SKUs of each large class of commodities and the prediction of the ratio of the ordering amount of each single commodity in the large class comprise a store characteristic vector, a date characteristic vector and a weather characteristic;
inputting time sequences consisting of time of sold-out, gross profit, passenger flow and passenger lists of whole shops from T-14 to T-1 days into an LSTM model, and determining shop characteristic vectors;
performing one-hot unique coding on seven days in a week to obtain week codes corresponding to T + N days; performing one-hot coding on the holidays to obtain holiday codes corresponding to the T + N days; performing one-hot coding on whether to carry out the shift compensation or not, and acquiring a shift compensation code corresponding to the T + N days; thereby determining a date feature vector;
and acquiring the air temperature of T + N days so as to determine the weather characteristics.
Preferably, the prediction of the total order amount at the store also uses a historical order feature vector and a time of sale at out of the store of T + N days,
total daily order amount G for T-14 to T + N-1 daysTInputting the formed time sequence into an LSTM model, and determining a historical ordering feature vector;
wherein T represents the order date, N represents the order N days in advance, and T + N-1 is the day before the order date;
and obtaining the time of sold out in the whole store of T + N days.
Preferably, the method of determining the total order amount includes,
inputting historical ordering characteristic vectors, the sold-out time characteristics of the whole store, the store characteristic vectors, the date characteristic vectors and the weather characteristics into an MLP multilayer sensing machine, and outputting total ordering amount of T + N days;
and training and testing the model according to the historical ordering data, and selecting the model with the highest testing accuracy to determine the total ordering amount.
Preferably, the large-class code characteristic vector, the large-class historical ordering characteristic vector and the large-class sold-out time are also used for predicting the large-class order ratio and the large-class SKU number of each commodity,
carrying out one-hot coding on the large class to determine a large class coding feature vector;
inputting a time sequence formed by the large-class order money amount and the large-class order SKU number of each day from T-14 to T + N-1 days into an LSTM model so as to determine a large-class historical order characteristic vector;
and obtaining the time of T + N days when the samples are sold out.
Preferably, the method of determining the amount of the general order includes,
inputting the large-class coded feature vector, the large-class historical ordering feature vector, the large-class sold-out time feature, the store feature vector, the date feature vector and the weather feature into an MLP (MLP), and outputting the T + N days of large-class direct prediction ordering amount and the large-class SKU number;
training and testing the models according to historical data, and selecting the model with the highest testing accuracy for predicting the large-class order amount and the large-class SKU number; thereby determining the ratio of the large order amount;
the ratio of the large order amount is determined according to formula 2:
Ci＝Bi/sumi(Bi) Equation 2
Wherein, CiOrder amount ratio of the large class i, BiDirect forecast order amount, sum, for large class iiIndicating summing all the large classes.
Preferably, the single product code characteristic vector, the single product historical ordering characteristic vector and the single product sold-out time are also used for predicting the single product ordering amount in the large class;
carrying out embedding coding on all single products in the ordered commodity pool, and determining single product coding characteristic vectors;
inputting a time sequence formed by the order amount of each single product every day from T-14 days to T + N-1 days into an LSTM model so as to determine a historical order feature vector of the single product;
and obtaining the time of sold out of the single product in T + N days.
Preferably, the method of determining the amount of the order for the single item over the broad category includes,
inputting the single product coding characteristic vector, the single product historical ordering characteristic vector, the single product sold-out time characteristic, the store characteristic vector, the date characteristic vector and the weather characteristic into an MLP (maximum likelihood ratio), and outputting the T + N days of single products to directly predict ordering amount;
training and testing the models according to historical data, selecting the model with the highest testing accuracy, and predicting the single-product order amount;
judging whether the number of the single products with the directly predicted money amount larger than zero accords with the constraint of the large SKU number, and controlling the number of the single products in the constraint range so as to determine the occupation ratio of the single products in the large SKU number;
determining the proportion of the single product in the large class according to formula 3:
Zi＝Si/sumi(Si) Equation 3
Wherein ZiFor the order-amount ratio of the singles i in the class to which they belong, SiDirect prediction of order amount, sum, for an individual item iiMeaning that all singlets in the class to which they belong are summed.
Preferably, the number of the single products is restricted according to the number of the large-class SKUs, the direct prediction ordering amount of the single products is sorted from large to small, and if the number of the single products exceeds the number of the large-class SKUs, the sorted and unnecessary ordering products are removed.
Preferably, the sold-out time is divided into sold-out time during model training and model testing and sold-out time during model prediction. The sold-out time during model training and testing is actually generated sold-out time, and the sold-out time in T + N days during model prediction is the expected optimal sold-out time; the optimal sold-out time is the time for automatically discounting to 0 gross interest rate at night.
Due to the adoption of the technical scheme, the invention has the remarkable technical effects that:
the invention utilizes the deep neural network to predict the real daily demand of the customer on three scales, including the total order amount, the ratio of each large order amount and the ratio of each large SKU number and the single order amount, thereby obtaining the single order amount and ensuring the total order amount and the rationality of the order structure;
the introduction of the sold-out time can better reflect the objective and real user demand. The optimal ordering amount can be reversely deduced according to the optimal sold-out time;
high-precision prediction is realized through reasonable selection and processing of deep learning feature vectors and construction of a deep learning model.
Drawings
FIG. 1 is a flow chart of the order quantity calculation of the present invention.
Fig. 2 is a diagram showing a total demand prediction model according to the present invention.
FIG. 3 is a block diagram of a large order amount and large SKU number prediction model of the present invention.
FIG. 4 is a block diagram of an order amount prediction model according to the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings 1 to 4 and the embodiments.
Example 1
An intelligent ordering method for a daily fresh store, the ordering quantity of single products is determined according to a formula 1,
Nk＝round(G*Ck*Zk/Ek) Equation 1
Wherein N iskThe number of orders to be made for a single item k, the total order amount of the whole store G and CkIs the ratio of the main class to which the single item k belongs to the total order amount, ZkFor the order-amount ratio of the individual articles in the class to which they belong, EkOrdering an amount for a unit specification of singleton k, round () means rounding to an integer; g, Ck,ZkThe prediction is respectively obtained through a first deep neural network model, a second deep neural network model and a third deep neural network model.
Determining a total order amount using a first deep neural network model, comprising:
total daily order amount G for T-14 to T + N-1 daysTInputting the formed time sequence into an LSTM model so as to determine a historical ordering feature vector; wherein T represents the order date, N represents the order N days in advance, and T + N-1 is the day before the order date;
obtaining the time of selling out of the whole store in T + N days;
inputting a time sequence consisting of sold-out time, gross profit, traffic flow and passenger list of T-14 to T-1 days into an LSTM model, and determining a store characteristic vector;
carrying out single-hot coding on seven days of a week to obtain week codes corresponding to T + N days; performing one-hot coding on the holidays to obtain holiday codes corresponding to the T + N days; performing one-hot coding on whether to carry out the shift compensation or not, and acquiring a shift compensation code corresponding to the T + N days; thereby obtaining a date feature vector;
acquiring the air temperature of T + N days so as to determine a weather feature vector;
inputting the historical total ordering characteristic vector, the sold-out time characteristic of the whole store, the store characteristic vector, the date characteristic vector and the weather characteristic into an MLP (MLP), and outputting the total ordering amount of T + N days.
And substituting the historical data into the model, training and testing the model, and selecting the model with the highest test accuracy to predict the total order amount.
Determining the ratio of the large-class order of each commodity by using a deep neural network model II, wherein the method comprises the following steps:
carrying out one-hot coding on the large class to determine a large class coding feature vector;
inputting a time sequence formed by the large-class order money amount and the large-class order SKU number of each day from T-14 to T + N-1 days into an LSTM model so as to determine a historical large-class order characteristic vector;
obtaining the time of T + N days when the samples are sold out;
inputting the large-class coded feature vector, the historical large-class order feature vector, the large-class sold-out time feature vector, the store feature vector, the date feature vector and the weather feature vector into an MLP (Multi-level queuing) and outputting the T + N-day large-class order amount and the large-class SKU number;
substituting the historical data into the model, training and testing the model, and selecting the model with the highest testing accuracy for predicting the large-class order amount and the large-class SKU number; thereby determining the ratio of the large order amount;
the large class order share ratio is determined according to formula 2,
Ci＝Bi/sumi(Bi) Equation 2
Wherein, CiOrder amount ratio of the large class i, BiDirect forecast order amount, sum, for large class iiIndicating summing all the large classes.
Determining the order amount ratio of the single product in the large class by using a deep neural network model III, which comprises the following steps:
performing embedded coding on the singles, and determining singles coding feature vectors;
inputting the time sequence of the daily single-product ordering amount from T-14 days to T + N-1 days into an LSTM model so as to determine a historical single-product ordering feature vector;
obtaining the time of sold out of the single product in T + N days;
and inputting the single product coding characteristic vector, the historical single product ordering characteristic vector, the single product selling-out time characteristic vector, the store characteristic vector, the date characteristic vector and the weather characteristic vector into an MLP (MLP), and outputting the T + N days of single product ordering amount.
Substituting the historical data into the model, training and testing the model, selecting the model with the highest testing accuracy, and predicting the order amount of the single product;
and judging whether the single product prediction conforms to the constraint of the large SKU number, sorting the ordered money of the single products obtained by the model prediction from large to small, and if the non-zero single product number exceeds the large SKU number, rejecting the non-ordered products which are ranked later.
Determining the proportion of the single product in the class to which the single product belongs according to formula 3:
Zi＝Si/sumi(Si) Equation 3
Wherein ZiFor the order-amount ratio of the singles i in the class to which they belong, SiOf singles iDirect prediction of order amount, sumiMeaning that all singlets in the class to which they belong are summed.
The sold-out time comprises sold-out time during model training and testing and sold-out time during model prediction, and the sold-out time during the model training and testing is real sold-out time and can be acquired from historical receipt data. But 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 rate is 0 at night. The gross profit rate of the whole store and the gross profit rate of the large store can be obtained by combining the prediction results of the gross occupation ratio and the single-product occupation ratio with the single-product raw profit rate.
Except for the coding features, all other features need to be cleaned of outliers, filled with missing values and normalized before being input into the model.
Example 2
In addition to embodiment 1, the singleton coding of the present embodiment adopts 64-dimensional embedding (embedding) coding. Embedded coding works well with thousands of singles. Compared with single hot coding, the embedded coding effectively reduces characteristic dimensionality and can well depict the similarity between different single products.
Example 3
On the basis of the above embodiments, the present embodiment adopts the order of the T +2 mode, that is, the order of today comes the next day. T-14 denotes day 14 before the current day of ordering, T-1 denotes day 1 before the current day of ordering, T +1 denotes the day after the current day of ordering, and T +2 denotes the day two days after the current day of ordering, that is, the arrival day.
Example 4
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, which is used to prevent overfitting, thereby improving the accuracy of prediction.
Example 5
On the basis of the above embodiment, the MLP module in this embodiment adopts a two-layer structure, and the dropout parameter between two layers is 0.4, which is used to prevent overfitting, thereby improving the accuracy of prediction.
Example 6
On the basis of the above embodiment, the deep neural network model in this embodiment is implemented by a pytorch, and model training, model testing, and model prediction are performed by using the NVIDIA a100 GPU.
Example 7
On the basis of the above-described embodiment, in practical application, the operation data, date data and weather data of the past year of 150 stores are acquired. All the input and output characteristics of a store during a day are combined into a data point. These data points are divided into a training set, a validation set, and a test set. And inputting the training set data into the model in a random batch mode for training, and performing cross validation on the training result by using the data of the validation set. And finally, testing on the test set, and selecting the model with the highest test accuracy as the prediction model.
Example 8
Based on the above embodiments, the present embodiment replaces the LSTM model with the variant GRU model of LSTM, which is faster in training speed.
Example 9
On the basis of the above embodiments, the general categories of the present embodiment include dry trash, aquatic products, fruits, beef and mutton, pork, poultry, vegetables, and eggs.

Claims (10)

1. An intelligent ordering method for a fresh-keeping daily store is characterized in that: the method comprises the steps of (1) carrying out,
forecasting the customer demand, namely forecasting the customer demand on three scales by using three deep neural network models; the three-scale prediction comprises the prediction of the total ordering amount of the store, the prediction of the order amount proportion of each commodity and the number of SKUs of each commodity in the large class, and the prediction of the proportion of the single-product order amount in the large class;
determining the order quantity of the single products, namely determining the order quantity of the single products according to a formula 1;
Nk＝round(G*Ck*Zk/Ek) Equation 1
Wherein N iskThe order amount of the single item k, the total order amount G and CkThe main category to which the single item k belongs is in general orderProportion in the amount of goods, ZkFor the order-amount ratio of the individual articles in the class to which they belong, EkFor an order amount of unit size for singleton k, round () represents rounding to an integer.
2. The intelligent ordering method for the Nisshink fresh store according to claim 1, wherein the method comprises the following steps: the prediction of the total ordering amount of the store, the ratio of the ordering amount of each large class of commodities, the number of the SKUs of each large class of commodities and the ratio of the ordering amount of each single commodity in the large class are realized by using a store characteristic vector, a date characteristic vector and weather characteristics;
inputting time sequences consisting of time of sold-out, gross profit, passenger flow and passenger lists of whole shops from T-14 to T-1 days into an LSTM model, and determining shop characteristic vectors;
performing one-hot unique coding on seven days in a week to obtain week codes corresponding to T + N days; performing one-hot coding on the holidays to obtain holiday codes corresponding to the T + N days; performing one-hot coding on whether to carry out the shift compensation or not, and acquiring a shift compensation code corresponding to the T + N days; thereby determining a date feature vector;
acquiring the air temperature of T + N days so as to determine weather characteristics; wherein T represents the order date, N represents an order N days in advance, and T + N is the arrival date.
3. The intelligent ordering method for the Nisshink fresh store according to claim 2, wherein the method comprises the following steps: the use of historical order characteristics vectors and the time sold out of the store are also used in the prediction of the total order amount at the store,
total daily order amount G for T-14 to T + N-1 daysTInputting the formed time sequence into an LSTM model, and determining a historical ordering feature vector;
and obtaining the time of sold out in the whole store of T + N days.
4. The intelligent ordering method for the Nisshink fresh store according to claim 3, wherein the method comprises the following steps: a method of total order amount determination includes,
inputting historical ordering characteristic vectors, the sold-out time characteristics of the whole store, the store characteristic vectors, the date characteristic vectors and the weather characteristics into an MLP multilayer sensing machine, and outputting total ordering amount of T + N days;
and training and testing the models according to the historical data, and selecting the model with the highest testing accuracy to predict the total order amount.
5. The intelligent ordering method for the Nisshink fresh store according to claim 2, wherein the method comprises the following steps: the prediction of the large-class order quantity ratio and the large-class SKU number of each commodity also uses a large-class coding characteristic vector, a large-class historical order characteristic vector and a large-class sold-out time,
carrying out one-hot coding on the large class to determine a large class coding feature vector;
inputting a time sequence formed by the large-class order money amount and the large-class order SKU number of each day from T-14 to T + N-1 days into an LSTM model so as to determine a large-class historical order characteristic vector;
and obtaining the time of T + N days when the samples are sold out.
6. The intelligent ordering method for the Nisshink fresh store according to claim 5, wherein the method comprises the following steps: a method of broad order amount proportion determination includes,
inputting the large-class coded feature vector, the large-class historical ordering feature vector, the large-class sold-out time feature, the store feature vector, the date feature vector and the weather feature into an MLP (MLP), and outputting the large-class ordering amount of T + N days and the large-class SKU number;
training and testing the models according to historical data, and selecting the model with the highest testing accuracy for predicting the large-class order amount and the large-class SKU number; thereby determining the ratio of the large order amount;
the ratio of the large order amount is determined according to formula 2:
Ci＝Bi/sumi(Bi) Equation 2
Wherein, CiOrder amount ratio of the large class i, BiPredicted order amount, sum, for class iiIndicating summing all the large classes.
7. The intelligent ordering method for the Nisshink fresh store according to claim 2, wherein the method comprises the following steps: the single product ordering amount is used in the prediction of the large class proportion, and the single product coding characteristic vector, the single product historical ordering characteristic vector and the single product sold-out time are also used;
carrying out embedding coding on all single products in the ordered commodity pool, and determining single product coding characteristic vectors;
inputting a time sequence formed by the order amount of each single product every day from T-14 days to T + N-1 days into an LSTM model so as to determine a historical order feature vector of the single product;
and obtaining the time of sold out of the single product in T + N days.
8. The intelligent ordering method for the Nisshink fresh store according to claim 7, wherein the method comprises the following steps: a method of determining the proportion of an order for a single good in a broad category includes,
inputting the single product coding characteristic vector, the single product historical ordering characteristic vector, the single product selling-out time characteristic, the store characteristic vector, the date characteristic vector and the weather characteristic into an MLP (maximum likelihood ratio), and outputting the ordering amount of the single product for T + N days;
training and testing the models according to historical data, selecting the model with the highest testing accuracy, and predicting the single-product order amount;
judging whether the number of the single products with the directly predicted money amount larger than zero accords with the constraint of the large SKU number, and controlling the number of the single products in the constraint range so as to determine the occupation ratio of the single products in the large SKU number;
determining the proportion of the single product in the large class according to formula 3:
Zi＝Si/sumi(Si) Equation 3
Wherein ZiFor the order-amount ratio of the singles i in the class to which they belong, SiDirect prediction of order amount, sum, for an individual item iiMeaning that all singlets in the class to which they belong are summed.
9. The intelligent ordering method for the Nisshink fresh store according to claim 8, wherein the method comprises the following steps: and (4) restricting the number of the single products according to the number of the large-class SKUs, sequencing the directly predicted ordering amount of the single products from large to small, and removing the non-ordered products after ranking if the number of the single products exceeds the number of the large-class SKUs.
10. The intelligent ordering method for the Nisshink fresh store according to claim 2, wherein the method comprises the following steps: the sold-out time is divided into sold-out time during model training and model testing and sold-out time during model prediction; the sold-out time during model training and testing is the actually occurred sold-out time; the T + N days of predicted time are predicted to be the expected best time.
CN202111054206.9A 2021-09-09 2021-09-09 Intelligent ordering method for Nisshink fresh stores Pending CN113869938A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111054206.9A CN113869938A (en) 2021-09-09 2021-09-09 Intelligent ordering method for Nisshink fresh stores

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111054206.9A CN113869938A (en) 2021-09-09 2021-09-09 Intelligent ordering method for Nisshink fresh stores

Publications (1)

Publication Number Publication Date
CN113869938A true CN113869938A (en) 2021-12-31

Family

ID=78995038

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111054206.9A Pending CN113869938A (en) 2021-09-09 2021-09-09 Intelligent ordering method for Nisshink fresh stores

Country Status (1)

Country Link
CN (1) CN113869938A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115271256A (en) * 2022-09-20 2022-11-01 华东交通大学 Intelligent ordering method under multi-dimensional classification

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115271256A (en) * 2022-09-20 2022-11-01 华东交通大学 Intelligent ordering method under multi-dimensional classification
CN115271256B (en) * 2022-09-20 2022-12-16 华东交通大学 Intelligent ordering method under multi-dimensional classification

Similar Documents

Publication Publication Date Title
Goyal et al. Recent trends in modeling of deteriorating inventory
Alfares et al. EOQ and EPQ production-inventory models with variable holding cost: state-of-the-art review
Janssen et al. A stochastic micro-periodic age-based inventory replenishment policy for perishable goods
CN109741083B (en) Material demand weighted prediction method based on enterprise MRP
CN112785354B (en) Commodity recommendation system for retail management platform
Paul et al. An artificial neural network model for optimization of finished goods inventory
Sazvar et al. A novel mathematical model for a multi-period, multi-product optimal ordering problem considering expiry dates in a FEFO system
CN108470261B (en) Ordering method and device
Pal et al. A stochastic production inventory model for deteriorating items with products’ finite life-cycle
Indrajitsingha et al. A fuzzy two-warehouse inventory model for single deteriorating item with selling-price-dependent demand and shortage under partial-backlogged condition
CN111652653A (en) Price determination and prediction model construction method, device, equipment and storage medium
CN113487359A (en) Multi-modal feature-based commodity sales prediction method and device and related equipment
CN113869938A (en) Intelligent ordering method for Nisshink fresh stores
Khan et al. A prepayment installment decision support framework in an inventory system with all-units discount against link-to-order prepayment under power demand pattern
Khanra et al. An EOQ model for perishable item with stock and price dependent demand rate
Jaggi et al. Impact of trade credit on inventory models for Weibull distribution deteriorating items with partial backlogging in two-warehouse environment
Kumar et al. Optimized warehouse management of perishable goods.
Ebrahimi et al. Bi-objective build-to-order supply chain problem with customer utility
Gunjal et al. Fusing clustering and machine learning techniques for Big-Mart sales predication
Lawal et al. A product backorder predictive model using recurrent neural network
Bardeji et al. Perishable inventory management using GA-ANN and ICA-ANN
Kaipov et al. Sales forecasting of goods in shoe retail
Mauricio et al. Predicting customer lifetime value through data mining technique in a direct selling company
Nodoust et al. A Genetic Algorithm for an inventory system under belief structure inflationary conditions
Harale et al. Dynamic small-series fashion order allocation and supplier selection: a ga-topsis-based model

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination