CN111192083A - Method for predicting intermittent demand - Google Patents

Method for predicting intermittent demand Download PDF

Info

Publication number
CN111192083A
CN111192083A CN201911369318.6A CN201911369318A CN111192083A CN 111192083 A CN111192083 A CN 111192083A CN 201911369318 A CN201911369318 A CN 201911369318A CN 111192083 A CN111192083 A CN 111192083A
Authority
CN
China
Prior art keywords
demand
prediction
clustering
products
training
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
CN201911369318.6A
Other languages
Chinese (zh)
Inventor
周伟华
周云
钱仲文
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.)
Zhejiang University ZJU
State Grid Zhejiang Electric Power Co Ltd
Original Assignee
Zhejiang University ZJU
State Grid Zhejiang Electric Power 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 Zhejiang University ZJU, State Grid Zhejiang Electric Power Co Ltd filed Critical Zhejiang University ZJU
Priority to CN201911369318.6A priority Critical patent/CN111192083A/en
Publication of CN111192083A publication Critical patent/CN111192083A/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
    • G06Q30/0202—Market predictions or forecasting for commercial activities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques

Abstract

The invention relates to a method for predicting intermittent demands, which is characterized by comprising the following steps: training features by extracting classification models
Figure DDA0002339249310000011
Predictive features
Figure DDA0002339249310000012
Classification model training features
Figure DDA0002339249310000013
Predictive features
Figure DDA0002339249310000014
And clustering feature F3(ii) a Firstly clustering products, and respectively training a classification model M whether each class has requirements1(ii) a Training regression prediction model M of demand quantity for each category respectively2(ii) a By M1Predicting whether each product has a demand in the future; if there is a demand, then use M2Predicting specific demand; and repeating the clustering and model training processes at intervals. The method is decomposed through multiple tasks, and the representation capability of the model is improved.

Description

Method for predicting intermittent demand
Technical Field
The present invention relates to the cross-domain of machine learning and supply chain management, and more particularly to a method of pattern mining and forecasting intermittent demand by consumers.
Background
Intermittent demand means that in the time series of product sales, there are some intermittent periods when the product is in stock and the demand or sales amount is 0. In supply chain management, intermittent demand often means that products are easily out of date, outdated, or lost, which can pose significant challenges for inventory efficiency optimization and operational cost reduction. In fact, in any supply chain, there may be intermittent demand for any product. Such products tend to occupy a relatively high proportion of the total inventory value, such as 60% (Johnston et al, 2003), particularly in aerospace parts, automotive retail, luxury goods, large machinery, and the like; in the case of the Molenaers (2010) analysis, 54% of the inventory products of petrochemical enterprises have not been in stock for 5 years. Therefore, the prediction method for improving the intermittent demand has great practical significance and application value for assisting the management decision of the supply chain, improving the operation efficiency of the supply chain of an organization and reducing the inventory cost.
Current prediction methods focus more on traditional time series frameworks, including ARIMA, exponential smoothing, moving average, and crosston methods, among others (1972). Among them, the crossbar method has many applications in practice, and ERP systems such as SAP, Forecast pro. The method considers the decomposition of the demand occurrence probability and the demand, and respectively updates the positive demand and the demand interval of each phase by utilizing exponential smoothing, wherein the demand/demand interval is the prediction result of each phase. The method has 3 disadvantages: 1) systematic positive bias exists in the predicted results, 2) after several periods of 0 demand, the predicted results become unavailable; 3) a large amount of service-related data is not fully utilized.
Under the era of 'internet +', the data volume of each industry is increased explosively, and a lot of data which cannot be collected in the past are now possible. It is necessary to develop a method for predicting intermittent demand by fully utilizing a large amount of newly added data. According to the concept of divide-and-conquer, we divide the prediction of the intermittent demand into two stages: 1) classification prediction of whether there is a demand, 2) regression prediction of demand. And the method is combined with a machine learning method, and full channel and full link information under a big data background is fully utilized.
Disclosure of Invention
In order to solve the technical problems, the invention adopts the following scheme:
a method for predicting intermittent demand, comprising: the method comprises the following steps:
1) determining a task objective to predict the demand for N products for a future T-phase;
2) collecting historical demand information, product related information and future marketing plan data;
3) respectively extracting N products and history T0Historical information construction classification training feature of period
Figure BDA0002339249290000021
N products and classification prediction feature constructed by historical information of future T-period
Figure BDA0002339249290000022
Marking historical requirements, if the requirements are greater than 0, marking the historical requirements as 1, otherwise marking the historical requirements as 0;
4) extracting N products and history T0Historical information of period building regression model training features
Figure BDA0002339249290000023
N products and prediction characteristics of regression model built by historical information of future T period
Figure BDA0002339249290000024
5) Extracting clustering characteristics F for N products3Clustering to obtain K categories;
6) respectively training a classification prediction model M for K classes according to clustering results1And regression prediction model M2；
7) By M1Carrying out classification prediction on whether N products have demands in the future T period;
8) if the classification result of the product n and the future t-th stage is in demand, namely the predicted value is 1, M is used2Predicting demand values, where N is [1, N ]],t∈[1,T]；
9) Every other TitAnd repeating the clustering and model training process.
The method for predicting the intermittent demand is characterized in that: the historical requirements in the step 2) comprise: the sum of the actual sales, truncated demand, and actual sales.
The method for predicting the intermittent demand is characterized in that: the clustering characteristic F in the step 5)3The method comprises the following steps: product attributes, correlations between products, product sales, and statistics of features.
The method for predicting the intermittent demand is characterized in that: and 5) clustering comprises a division method, a hierarchy method, a density algorithm, a graph theory clustering method, a grid algorithm, a model algorithm, association clustering, subspace clustering and mode clustering.
The method for predicting the intermittent demand is characterized in that: step 6) the classification prediction model M1The method comprises decision trees, random forests, neural networks, logistic regression, support vector machines, xgboost, GBDT and Bayesian classifiers.
The method for predicting the intermittent demand is characterized in that: step 6) the regression prediction model M2The method comprises decision trees, random forests, neural networks, support vector machines, xgboost, GBDT and linear regression.
The method for predicting the intermittent demand is characterized in that: the classification prediction features in the step 3) are the same as or different from the regression prediction features in the step 4).
The method for predicting the intermittent demand is characterized in that: step 9) the model training comprises a rolling training method, wherein each training adopts a fixed T0Historical information of the period.
The method for predicting the intermittent demand has the following beneficial effects:
the discontinuous demand prediction method can fully utilize a large amount of newly added data to predict the discontinuous demand. According to the concept of divide-and-conquer, the prediction of the discontinuous demand is divided into two stages: 1) whether classification prediction of demands exists or not, 2) regression prediction of demand, and a machine learning method are combined, so that full channel and full link information under a big data background are fully utilized. Decomposition is carried out through multiple tasks, and the representation capability of the model is improved. The prediction result has no systematic positive bias, and the prediction result can be well utilized.
Drawings
FIG. 1: the invention discloses a flow chart for predicting intermittent demand.
Detailed Description
The present invention will be described in further detail with reference to the following drawings and specific examples, which are only used for explaining and explaining the present invention and are not to be construed as limiting the scope of the present invention. The scope of the present invention is defined by the appended claims, and all changes, substitutions and the like that occur to some elements are intended to be embraced therein.
The technical scheme of the invention is that the training characteristics of the classification model are extracted
Figure BDA0002339249290000031
Predictive features
Figure BDA0002339249290000032
Classification model training features
Figure BDA0002339249290000033
Predictive features
Figure BDA0002339249290000034
And clustering feature F3(ii) a Firstly clustering products, and respectively training a classification model M whether each class has requirements1(ii) a Training regression prediction model M of demand quantity for each category respectively2(ii) a By M1Predicting whether each product has a demand in the future; if there is a demand, then use M2Predicting specific demand; repeating clustering and model training at intervalsAnd (6) carrying out the process.
Suppose the demand for N products for the future T phase is to be predicted.
1) Collecting historical demand information, product related information and future marketing plan data;
2) respectively extracting N products and history T0Historical information construction classification training feature of period
Figure BDA0002339249290000035
N products and classification prediction feature constructed by historical information of future T-period
Figure BDA0002339249290000036
Marking historical requirements, if the requirements are greater than 0, marking the historical requirements as 1, otherwise marking the historical requirements as 0;
3) extracting N products and history T0Historical information of period building regression model training features
Figure BDA0002339249290000041
N products and prediction characteristics of regression model built by historical information of future T period
Figure BDA0002339249290000042
4) Extracting clustering characteristics F for N products3Clustering to obtain K categories;
5) respectively training a classification prediction model M for K classes according to clustering results1And regression prediction model M2；
6) By M1Carrying out classification prediction on whether N products have demands in the future T period;
7) if the classification result of the product n and the future t-th stage is in demand, namely the predicted value is 1, M is used2Predicting demand values, where N is [1, N ]],t∈[1,T]；
8) Every other TitAnd repeating the clustering and model training process.
In the above aspect, 1) said T0A range of values includes, but is not limited to, any time period in the range of 1 hour to 10 years into the future.
In the foregoing aspect, 1) the demand includes: the sum of the actual sales, truncated demand, and actual sales.
In the above schemes, the classification model features and the regression models in 2) and 3) may be different, and both include but are not limited to: historical sales, historical prices, reviews, marketing campaigns, advertising information, customer purchasing behavior records, trend information, period information, season information.
In the above scheme, 4) the clustering characteristic F3Including but not limited to brand, category, functional parameters, price, channel characteristics, social attributes, longevity, consumption speed, etc.
In the foregoing solution, 4) the clustering method includes: the method comprises the following steps of division, hierarchy, density algorithm, graph theory clustering method, grid algorithm, model algorithm, association clustering, subspace clustering and pattern clustering.
In the above scheme, 5) the classification model M1The method comprises the following steps: decision trees, random forests, neural networks, logistic regression, support vector machines, xgboost, GBDT, Bayesian classifiers.
In the above embodiment, 5) the regression model M2The method comprises the following steps: decision trees, random forests, neural networks, support vector machines, xgboost, GBDT, linear regression.
In the scheme, 8) the model training adopts a rolling training method, and every T istrPerforming a rolling process, repeating clustering and model training, and using a fixed T0Historical information of the period.
Example one
Statistically, a chain of luxury companies has 100 SKUs for sale, and the headquarters needs to allocate product inventory to multiple stores and predict the daily demand of 2019 from 5 months, 19 days to 26 days. Collecting information of historical sales, price change, marketing activities, customer repurchase, future marketing plans and the like of the products in the last 3 years, constructing a classification model training characteristic and a demand marking matrix of the product i, carrying out category marking on the demand, and recording the category marking as the demand marking
Figure BDA0002339249290000051
The demand greater than 0 is marked as 1, otherwise the demand is marked as 0:
Figure BDA0002339249290000052
wherein i ∈ [1,100 ]]。
Figure BDA0002339249290000053
The first column of the matrix is a demand mark, and the column vectors starting from the second column respectively represent the sales volume of the same day in the last week, the selling price of the same day, the selling price of the last week, whether sales promotion exists or not, whether money is exploded or not, the week day, the month and the shop advertisement delivery. The number of rows is the number of days in the history, 3 x 365 ═ 1095.
Constructing a prediction characteristic and a demand marking matrix of a future T-period classification model of a product i, wherein the statistics of demand marking is initialized to be 0:
Figure BDA0002339249290000054
wherein i ∈ [1,100 ]]。
Figure BDA0002339249290000055
The first column of the matrix is demand, where demand is represented by actual sales; the second column is a demand mark; the column vectors starting from the third column indicate sales volume on the same day of the last week, sales price on the same day, sales price on the last week, sales promotion, whether money is exploded, day of week, month, and shop advertisement delivery, respectively. The number of rows is the number of days in the history, 3 x 365 ═ 1095.
Constructing a regression model training characteristic and demand matrix of the product i,
Figure BDA0002339249290000056
representing the demand of product i at time t:
Figure BDA0002339249290000057
wherein i ∈ [1,100 ]]。
Figure BDA0002339249290000058
The first column of the matrix is the demand, expressed in actual sales; the column vectors starting at the second column respectively represent the sales volume of the same day in the last week, the selling price of the same day, the selling price of the last week, whether sales promotion exists or not, whether money is exploded or not, the day of the week, the month and the advertisement delivery of stores. The number of rows is the number of days in the history, 3 x 365 ═ 1095.
Constructing a regression model prediction characteristic and demand matrix of a product i in the future T period, and uniformly initializing the demand to be 0:
Figure BDA0002339249290000059
wherein i ∈ [1,100 ]]。
Figure BDA00023392492900000510
The first column of the matrix is the demand, expressed in actual sales; the column vectors starting at the second column respectively represent the sales volume of the same day in the last week, the selling price of the same day, the selling price of the last week, whether sales promotion exists or not, whether money is exploded or not, the day of the week, the month and the advertisement delivery of stores. The number of rows is the number of days in the history, 3 x 365 ═ 1095.
Correspondingly extracting attribute characteristics F of the product i3,i
F3,i＝[PLi,Cai,Dai,Pui,PTi,PPi],i∈[1,100]
Wherein PLi,Cai,Dai,Pui,PTi,PPiIndicating the price level, category, presence or absence of diamonds, purity, type of packaging, origin of product i. Based on F3,iAnd (5) performing hierarchical clustering on 100 products, and dividing the products into 8 types.
Model training is respectively carried out on 8 classes by adopting a random forest to obtain a classification model set M1(ii) a Training regression model set M by adopting double-hidden-layer neural network and characteristic standardization2。
Using a set of models M1Respectively carrying out classification prediction on whether the future 7 days of 100 SKUs have demands or not, and if the prediction result is 1, then carrying out classification prediction on whether the future 7 days of the SKUs have demands or not
Figure BDA0002339249290000061
The corresponding date in (1); otherwise it is not necessary to
Figure BDA0002339249290000062
And (5) re-marking.
Using a set of models M2Forecasting the demand of 100 SKUs in 7 days in the future and every day respectively
Figure BDA0002339249290000063
Is listed in
Figure BDA0002339249290000064
Filling the predicted demand value with the corresponding date marked as 1; if the actual predicted required value is 0 and marked as 1, the required value of the previous date is filled in (the selected predicted value in the prediction set and the selected actual value in the training set), and if the required value of the previous date is also 0, the required values which are not 0 are searched in turn in the previous day.
If long-term prediction is used, T is measured every 30 daysitAnd (6) repeating the processes of feature extraction, clustering, classification model and regression model training and prediction.
Example two
Statistically, a chain of luxury companies has 200 SKUs for sale, and the headquarters needs to allocate product inventory to multiple stores and predict the daily demand of 2019 from 6 months, 15 days to 21 days. Collecting information of historical sales, price change, marketing activities, customer repurchase, future marketing plans and the like of the products in the last 3 years, constructing a classification model training characteristic and a demand marking matrix of the product i, carrying out category marking on the demand, and recording the category marking as the demand marking
Figure BDA0002339249290000065
The demand greater than 0 is marked as 1, otherwise the demand is marked as 0:
Figure BDA0002339249290000066
wherein i ∈ [1,200 ]]。
Figure BDA0002339249290000067
The first column of the matrix is a demand mark, and the column vectors starting from the second column respectively represent the sales volume of the same day in the last week, the selling price of the same day, the selling price of the last week, whether sales promotion exists or not, whether money is exploded or not, the week day, the month and the shop advertisement delivery. The number of rows is the number of days in the history, 3 x 365 ═ 1095.
Constructing a prediction characteristic and a demand marking matrix of a future T-period classification model of a product i, wherein the statistics of demand marking is initialized to be 0:
Figure BDA0002339249290000068
wherein i ∈ [1,200 ]]。
Figure BDA0002339249290000071
The first column of the matrix is demand, where demand is represented by actual sales; the second column is a demand mark; the column vectors starting from the third column indicate sales volume on the same day of the last week, sales price on the same day, sales price on the last week, sales promotion, whether money is exploded, day of week, month, and shop advertisement delivery, respectively. The number of rows is the number of days in the history, 3 x 365 ═ 1095.
Constructing a regression model training characteristic and demand matrix of the product i,
Figure BDA0002339249290000072
representing the demand of product i at time t:
Figure BDA0002339249290000073
wherein i ∈ [1,200 ]]。
Figure BDA0002339249290000074
The first column of the matrix is the demand, expressed in actual sales; the column vectors starting from the second column respectively represent the sales volume of the same day in the last week, the sales price of the same day, the sales price of the last week, whether sales promotion exists or not, whether money is exploded or not, the day of the week, the month, the advertisement delivery of a store and the average repurchase rate of SKU in the last 3 years.The number of rows is the number of days in the history, 3 x 365 ═ 1095.
Constructing a regression model prediction characteristic and demand matrix of a product i in the future T period, and uniformly initializing the demand to be 0:
Figure BDA0002339249290000075
wherein i ∈ [1,200 ]]。
Figure BDA0002339249290000076
The first column of the matrix is the demand, expressed in actual sales; the column vectors starting from the second column respectively represent the sales volume of the same day in the last week, the sales price of the same day, the sales price of the last week, whether sales promotion exists or not, whether money is exploded or not, the day of the week, the month, the advertisement delivery of a store and the average repurchase rate of SKU in the last 3 years. The number of rows is the number of days in the history, 3 x 365 ═ 1095.
Correspondingly extracting attribute characteristics F of the product i3,i
F3,i＝[PLi,Cai,Dai,Pui,PTi,PPi],i∈[1,200]
Wherein PLi,Cai,Dai,Pui,PTi,PPiIndicating the price level, category, presence or absence of diamonds, purity, type of packaging, origin of product i. Based on F3,iAnd (4) performing hierarchical clustering on 200 products, wherein the products are classified into K-9 types.
Using logistic regression to
Figure BDA0002339249290000077
Respectively carrying out model training on the 9 classes as a target value to obtain a classification model set M1(ii) a Using random forests to
Figure BDA0002339249290000078
Training a set of regression models M for the target value2。
Using a set of models M1Respectively carrying out classification prediction on whether the future 7 days of 200 SKUs have demands or not, and if the prediction result is 1, then carrying out classification prediction on whether the future 7 days of the SKUs have demands or not
Figure BDA0002339249290000079
The corresponding date in (1); otherwise it is not necessary to
Figure BDA00023392492900000710
And (5) re-marking.
Using a set of models M2Respectively carrying out demand quantity prediction on 200 SKUs in 7 days in the future
Figure BDA0002339249290000081
Is listed in
Figure BDA0002339249290000082
Filling the predicted demand value with the corresponding date marked as 1; if the actual predicted required value is 0 and marked as 1, the required value of the previous date is filled in (the selected predicted value in the prediction set and the selected actual value in the training set), and if the required value of the previous date is also 0, the required values which are not 0 are searched in turn in the previous day.
If long-term prediction, every two weeks, i.e. TitThe feature extraction, clustering, classification model and regression model training, prediction process was repeated 14.
EXAMPLE III
Statistically, a chain of luxury companies has 500 SKUs on sale, and the headquarters needs to allocate product inventory to multiple stores, and needs to predict the daily demand of 2019 from 7 months and 10 days to 16 days. Collecting information of historical sales, price change, marketing activities, customer repurchase, future marketing plans and the like of products in the past 2 years, constructing a classification model training characteristic and a demand marking matrix of a product i, carrying out category marking on demands, and recording the category marking as
Figure BDA0002339249290000083
The demand greater than 0 is marked as 1, otherwise the demand is marked as 0:
Figure BDA0002339249290000084
wherein i ∈[1,500]。
Figure BDA0002339249290000085
The first column of the matrix is a demand mark, and the column vectors starting from the second column respectively represent the sales volume of the same day in the last week, the selling price of the same day, the selling price of the last week, whether sales promotion exists or not, whether money is exploded or not, the week day, the month, the advertisement delivery of a store and the average repurchase rate of the product in the last 2 years. The number of rows is the number of days in the history, 2 x 365 — 730.
Constructing a prediction characteristic and a demand marking matrix of a future T-period classification model of a product i, wherein the statistics of demand marking is initialized to be 0:
Figure BDA0002339249290000086
wherein i ∈ [1,500 ]]。
Figure BDA0002339249290000087
The first column of the matrix is demand, where demand is represented by actual sales; the second column is a demand mark; the column vectors beginning in the third column respectively represent sales volume on the same day of the last week, sales price on the same day, sales price on the last week, sales promotion, whether money is exploded, day of week, month, shop advertising, and average repurchase rate of the product in the last 2 years. The number of rows is the number of days in the history, 2 x 365 — 730.
Constructing a regression model training characteristic and demand matrix of the product i,
Figure BDA0002339249290000088
representing the demand of product i at time t:
Figure BDA0002339249290000089
wherein i ∈ [1,500 ]]。
Figure BDA00023392492900000810
The first column of the matrix is the demand, expressed in actual sales; the column vectors starting at the second column represent the sales volume on the same day of the previous week, the selling price on the same day, the selling price on the previous week, and the presence or absence of promotionSale, whether the money is exploded, week day, month and store advertising. The number of rows is the number of days in the history, 2 x 365 — 730.
Constructing a regression model prediction characteristic and demand matrix of a product i in the future T period, and uniformly initializing the demand to be 0:
Figure BDA0002339249290000091
wherein i ∈ [1,500 ]]。
Figure BDA0002339249290000092
The first column of the matrix is the demand, expressed in actual sales; the column vectors starting at the second column respectively represent the sales volume of the same day in the last week, the selling price of the same day, the selling price of the last week, whether sales promotion exists or not, whether money is exploded or not, the day of the week, the month and the advertisement delivery of stores. The number of rows is the number of days in the history, 2 x 365 — 730.
Correspondingly extracting attribute characteristics F of the product i3,i
F3,i＝[PLi,Cai,Dai,Pui,PTi,PPi,Mi],i∈[1,500]
Wherein PLi,Cai,Dai,Pui,PTi,PPiIndicating the price level, category, presence or absence of diamonds, purity, type of packaging, place of origin and material of product i. Based on F3,iAnd performing hierarchical clustering on 500 products, wherein the products are classified into K-9 types.
Using a support vector machine to
Figure BDA0002339249290000093
Respectively carrying out model training on the 9 classes as a target value to obtain a classification model set M1(ii) a Using xgboost, to
Figure BDA0002339249290000094
Training a set of regression models M for the target value2。
Using a set of models M1Respectively carrying out the judgment on whether the future 7 days of 500 SKUs are in need every dayClassifying and predicting, if the prediction result is 1, then
Figure BDA0002339249290000095
The corresponding date in (1); otherwise it is not necessary to
Figure BDA0002339249290000096
And (5) re-marking.
Using a set of models M2Respectively carrying out demand quantity prediction on 7 days in the future for 500 SKUs every day
Figure BDA0002339249290000097
Is listed in
Figure BDA0002339249290000098
Filling the predicted demand value with the corresponding date marked as 1; if the actual predicted required value is 0 and marked as 1, the required value of the previous date is filled in (the selected predicted value in the prediction set and the selected actual value in the training set), and if the required value of the previous date is also 0, the required values which are not 0 are searched in turn in the previous day.
If long-term prediction, every other week, i.e. TitAnd 7, repeating the processes of feature extraction, clustering, training of classification models and regression models and prediction.

Claims (8)

1. A method for predicting intermittent demand, comprising: the method comprises the following steps:
1) determining a task objective to predict the demand for N products for a future T-phase;
2) collecting historical demand information, product related information and future marketing plan data;
3) respectively extracting N products and history T0Historical information construction classification training feature of period
Figure FDA0002339249280000011
N products and classification prediction feature constructed by historical information of future T-period
Figure FDA0002339249280000012
Marking historical requirements, if the requirements are greater than 0, marking the historical requirements as 1, otherwise marking the historical requirements as 0;
4) extracting N products and history T0Historical information of period building regression model training features
Figure FDA0002339249280000013
N products and prediction characteristics of regression model built by historical information of future T period
Figure FDA0002339249280000014
5) Extracting clustering characteristics F for N products3Clustering to obtain K categories;
6) respectively training a classification prediction model M for K classes according to clustering results1And regression prediction model M2；
7) By M1Carrying out classification prediction on whether N products have demands in the future T period;
8) if the classification result of the product n and the future t-th stage is in demand, namely the predicted value is 1, M is used2Predicting demand values, where N is [1, N ]],t∈[1,T]；
9) Every other TitAnd repeating the clustering and model training process.
2. A method of demand break prediction as claimed in claim 1, characterised in that: the historical requirements in the step 2) comprise: the sum of the actual sales, truncated demand, and actual sales.
3. A method of demand break prediction as claimed in claim 1, characterised in that: the clustering characteristic F in the step 5)3The method comprises the following steps: product attributes, correlations between products, product sales, and statistics of features.
4. A method of demand break prediction as claimed in claim 1, characterised in that: and 5) clustering comprises a division method, a hierarchy method, a density algorithm, a graph theory clustering method, a grid algorithm, a model algorithm, association clustering, subspace clustering and mode clustering.
5. A method of demand break prediction as claimed in claim 1, characterised in that: step 6) the classification prediction model M1The method comprises decision trees, random forests, neural networks, logistic regression, support vector machines, xgboost, GBDT and Bayesian classifiers.
6. A method of demand break prediction as claimed in claim 1, characterised in that: step 6) the regression prediction model M2The method comprises decision trees, random forests, neural networks, support vector machines, xgboost, GBDT and linear regression.
7. A method of demand break prediction as claimed in claim 1, characterised in that: the classification prediction features in the step 3) are the same as or different from the regression prediction features in the step 4).
8. A method of demand break prediction as claimed in claim 1, characterised in that: step 9) the model training comprises a rolling training method, wherein each training adopts a fixed T0Historical information of the period.
CN201911369318.6A 2019-12-26 2019-12-26 Method for predicting intermittent demand Pending CN111192083A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911369318.6A CN111192083A (en) 2019-12-26 2019-12-26 Method for predicting intermittent demand

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911369318.6A CN111192083A (en) 2019-12-26 2019-12-26 Method for predicting intermittent demand

Publications (1)

Publication Number Publication Date
CN111192083A true CN111192083A (en) 2020-05-22

Family

ID=70709603

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911369318.6A Pending CN111192083A (en) 2019-12-26 2019-12-26 Method for predicting intermittent demand

Country Status (1)

Country Link
CN (1) CN111192083A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114662997A (en) * 2022-05-20 2022-06-24 成都运荔枝科技有限公司 Cold chain transport capacity demand prediction method and cold chain transport capacity distribution method
CN116402321A (en) * 2023-06-08 2023-07-07 北京京东乾石科技有限公司 Method and device for determining demand of article, electronic equipment and storage medium

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109754118A (en) * 2018-12-26 2019-05-14 复旦大学 A kind of prediction technique of system self-adaption
CN109784979A (en) * 2018-12-19 2019-05-21 重庆邮电大学 A kind of supply chain needing forecasting method of big data driving
CN109886445A (en) * 2018-12-13 2019-06-14 国网浙江省电力有限公司衢州供电公司 A kind of tomorrow requirement prediction technique based on material requirements property quantification
CN110135630A (en) * 2019-04-25 2019-08-16 武汉数澎科技有限公司 The short term needing forecasting method with multi-step optimization is returned based on random forest

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109886445A (en) * 2018-12-13 2019-06-14 国网浙江省电力有限公司衢州供电公司 A kind of tomorrow requirement prediction technique based on material requirements property quantification
CN109784979A (en) * 2018-12-19 2019-05-21 重庆邮电大学 A kind of supply chain needing forecasting method of big data driving
CN109754118A (en) * 2018-12-26 2019-05-14 复旦大学 A kind of prediction technique of system self-adaption
CN110135630A (en) * 2019-04-25 2019-08-16 武汉数澎科技有限公司 The short term needing forecasting method with multi-step optimization is returned based on random forest

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
张瑞: "不常用备件需求预测模型与方法研究" *
张瑞: "不常用备件需求预测模型与方法研究", 中国博士学位论文全文数据库, no. 9, pages 9 - 19 *
林琳等: "一种新的间断型备件需求预测方法", vol. 48, no. 1, pages 40 - 45 *
许绍杰等: "基于SVM的雷达装备器材间断性需求预测", vol. 33, no. 6, pages 87 - 90 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114662997A (en) * 2022-05-20 2022-06-24 成都运荔枝科技有限公司 Cold chain transport capacity demand prediction method and cold chain transport capacity distribution method
CN114662997B (en) * 2022-05-20 2022-08-09 成都运荔枝科技有限公司 Cold chain transport capacity demand prediction method and cold chain transport capacity distribution method
CN116402321A (en) * 2023-06-08 2023-07-07 北京京东乾石科技有限公司 Method and device for determining demand of article, electronic equipment and storage medium
CN116402321B (en) * 2023-06-08 2023-09-22 北京京东乾石科技有限公司 Method and device for determining demand of article, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
Zunic et al. Application of facebook's prophet algorithm for successful sales forecasting based on real-world data
Li et al. The inventory management system for automobile spare parts in a central warehouse
CN110392899A (en) The dynamic feature selection generated for model
CN111295681A (en) Demand prediction using a weighted hybrid machine learning model
Abdolazimi et al. Designing a new mathematical model based on ABC analysis for inventory control problem: A real case study
CN103430196A (en) Sales prediction and recommendation system
CA3057521A1 (en) Method and system for generation of at least one output analytic for a promotion
Jain et al. Sales forecasting for retail chains
Birim et al. The derived demand for advertising expenses and implications on sustainability: a comparative study using deep learning and traditional machine learning methods
Punia et al. Predictive analytics for demand forecasting: A deep learning-based decision support system
CN114548871B (en) Big data driven supply chain AI intelligent replenishment system
US20240086726A1 (en) Systems and methods for big data analytics
Januschowski et al. 3.2 A CLASSIFICATION OF BUSINESS FORECASTING PROBLEMS
CN111192083A (en) Method for predicting intermittent demand
Theodorou et al. Optimizing inventory control through a data-driven and model-independent framework
Nong An application of delphi and dea to performance efficiency assessment of retail stores in fashion industry
Lahrichi et al. A first MILP model for the parameterization of Demand-Driven MRP
CN114372848A (en) Tobacco industry intelligent marketing system based on machine learning
Pandey et al. Machine Learning based Food Demand Estimation for Restaurants
WO2022055615A1 (en) Short life cycle sales curve estimation
Kar et al. A study on using business intelligence for improving marketing efforts
CN111127072A (en) Multi-stage real-time prediction method for new product requirements
Ramya et al. An advanced sales forecasting using machine learning algorithm
Prabhakar et al. Demand Forecasting for a large grocery chain in Ecuador
Samarghandi Studying the impact of merged and divided storage policies on the profitability of a remanufacturing system with deteriorating revenues

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