CN115204535B - Purchasing business volume prediction method based on dynamic multivariate time sequence and electronic equipment - Google Patents

Purchasing business volume prediction method based on dynamic multivariate time sequence and electronic equipment Download PDF

Info

Publication number
CN115204535B
CN115204535B CN202211126215.9A CN202211126215A CN115204535B CN 115204535 B CN115204535 B CN 115204535B CN 202211126215 A CN202211126215 A CN 202211126215A CN 115204535 B CN115204535 B CN 115204535B
Authority
CN
China
Prior art keywords
time
sequence
input
attention
stage
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.)
Active
Application number
CN202211126215.9A
Other languages
Chinese (zh)
Other versions
CN115204535A (en
Inventor
卢华
陈晶
陈挚
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.)
Hubei Xintong Communication Co ltd
Original Assignee
Hubei Xintong Communication 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 Hubei Xintong Communication Co ltd filed Critical Hubei Xintong Communication Co ltd
Priority to CN202211126215.9A priority Critical patent/CN115204535B/en
Publication of CN115204535A publication Critical patent/CN115204535A/en
Application granted granted Critical
Publication of CN115204535B publication Critical patent/CN115204535B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G06F16/2474—Sequence data queries, e.g. querying versioned 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
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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
    • G06N3/08—Learning methods
    • 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
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
    • 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
    • G06Q10/00—Administration; Management
    • G06Q10/08—Logistics, e.g. warehousing, loading or distribution; Inventory or stock management
    • G06Q10/087—Inventory or stock management, e.g. order filling, procurement or balancing against orders

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • General Physics & Mathematics (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Development Economics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Business, Economics & Management (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Tourism & Hospitality (AREA)
  • Quality & Reliability (AREA)
  • Mathematical Physics (AREA)
  • Operations Research (AREA)
  • Marketing (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Educational Administration (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Game Theory and Decision Science (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Finance (AREA)
  • Fuzzy Systems (AREA)
  • Accounting & Taxation (AREA)
  • Probability & Statistics with Applications (AREA)
  • Databases & Information Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a purchasing business volume prediction method based on a dynamic multivariate time sequence and electronic equipment. The invention solves the problem of incomprehensive manual feature extraction in purchasing traffic prediction, and the time convolution module provided by the invention extracts the nonlinear feature of a static time sequence. The invention solves the problems of gradient disappearance and gradient explosion caused by overlong time interval when long-period time sequence prediction is carried out by purchasing traffic prediction, and the proposed multi-stage attention module adaptively selects important characteristics to model the dependency relationship between states of long time interval. The invention solves the problem that a nonlinear module cannot process the prediction of the dynamic time sequence in the purchasing traffic prediction, and the proposed autoregressive module processes the dynamic time sequence with unfixed period.

Description

Purchasing business volume prediction method based on dynamic multivariate time sequence and electronic equipment
Technical Field
The invention belongs to the field of artificial intelligence in computer science, relates to a purchasing traffic prediction method and electronic equipment, and particularly relates to a purchasing traffic prediction method and electronic equipment based on a dynamic multivariate time sequence in robot-oriented process automation.
Background
In recent years, purchasing management and supply chain management play an extremely important role in enterprise development and competition processes, a large amount of purchasing business volume prediction work exists in a traditional purchasing process, the purchasing business volume prediction means that the purchasing business volume of a next period is predicted based on large batch of purchasing data in the purchasing process, and efficient and reliable purchasing business volume prediction can help an enterprise to make a better purchasing decision in the next period, so that cost is saved, and purchasing efficiency is improved. However, the purchasing data required for purchasing traffic prediction is often of various types, most of the purchasing data are time series data and have long duration, and part of the data have no fixed period and belong to dynamic multivariate time series data. Time series refers to a series of data collected at regular intervals, and multivariate time series refers to a multivariate time series, i.e., a plurality of variables recorded over time. The dynamic multivariate time series data refers to multivariate time series data with an unfixed period and dynamically changing sequence length.
The traditional purchasing traffic prediction method is to manually select features and train a fully-connected neural network for prediction. However, the problem of incomplete feature extraction by manpower, the long-term dependence relationship of long-period time series cannot be modeled by a fully-connected neural network, and dynamic time series data cannot be processed, so that the search for purchasing traffic prediction of dynamic multivariate time series in robot process automation is one of the targets pursued by many enterprises.
The Robot Process Automation (RPA) is a new type artificial intelligent virtual process automation robot, which can simulate the operation of human on the computer interface by specific rules and automatically execute the corresponding process task according to the rules to replace or assist the human to complete the related computer operation. The demand for Robot Process Automation (RPA) has increased rapidly in recent years by various industries, and it is estimated that as many as 90% of medium and large organizations select a Robot Process Automation (RPA) solution by 2020. Robot Process Automation (RPA) has been widely used in various fields, and in recent years, with the advent and development of corporate financial sharing centers, robot Process Automation (RPA) has also been increasingly used in the fields of procurement management and supply chain management. The Robot Process Automation (RPA) has the advantages of high working efficiency, high accuracy, high expandability, compliance and safety, capability of automating simple tasks, capability of reducing procurement management cost and capability of freeing up space for more strategic activities.
Disclosure of Invention
The invention provides a purchasing traffic prediction method and electronic equipment based on dynamic multivariate time sequence in robot process automation, which are oriented to the problems that the characteristics are not comprehensively extracted manually in the purchasing traffic prediction, when the purchasing traffic prediction is carried out for long-period time sequence prediction, the problems of gradient disappearance and gradient explosion caused by overlong time interval and the demand for the purchasing traffic prediction of the dynamic time sequence.
The method adopts the technical scheme that: a purchasing traffic prediction method based on dynamic multivariate time sequence inputs data obtained by a robot flow automatic data acquisition network into a robot flow automatic purchasing traffic prediction network to predict purchasing traffic;
the robot process automation data acquisition network extracts index data required by service purchase quantity according to the keyword table, arranges the data according to time, and obtains time sequence data of N indexes as an external sequence
Figure DEST_PATH_IMAGE001
Wherein N represents the number of indexes, and T is the length of the external sequence; each index is a variable, the firstnThe time series data of each index is the secondnSingle univariate time series data
Figure DEST_PATH_IMAGE002
Is as followsnTime series data of the individual indices; the time sequence data of all indexes form dynamic multi-element time sequence data which are used as input variables X of the robot process automatic purchase business volume prediction network;
the robot process automatic purchasing business volume prediction network comprises a time convolution module, a multi-stage attention module, a full connection layer and an autoregressive module;
the time convolution module is characterized in that the time convolution module is composed of K convolution kernels, the size of each convolution kernel is F multiplied by M, F is the width of each convolution kernel, M is the depth of each convolution kernel, time sequence data of other N static indexes except the purchasing business volume index are used as an input external sequence X to be input into each convolution kernel, and finally an output characteristic sequence H with the size of P multiplied by K is obtained, wherein P is P
Figure 183407DEST_PATH_IMAGE003
Is also the length of the univariate time series;
the multi-stage attention module, comprising a sequenceA first stage attention layer, a second stage attention layer and a temporal attention layer connected; the output variable H passes through the first stage attention layer, an external sequence is selected in a self-adaptive mode for learning, the output external sequence of the first stage attention layer is connected with a target sequence corresponding to time through the second stage attention layer, finally the hidden states in the first two layers are combined through the time attention layer to learn the characteristics of longer time dependence, and the characteristics of longer time dependence are output
Figure DEST_PATH_IMAGE004
；
The autoregressive module takes the dynamic multivariate time sequence data as an input vector X and outputs the input vector X
Figure 897285DEST_PATH_IMAGE005
；
Figure DEST_PATH_IMAGE006
；
Wherein the content of the first and second substances,
Figure 37542DEST_PATH_IMAGE007
in order to be a super-parameter,
Figure DEST_PATH_IMAGE008
is a learnable weight vector, X is an input vector,Uin order to input the weight matrix, the weight matrix is input,bin order to be a vector of the offset,
Figure 229489DEST_PATH_IMAGE009
is the noise at the time t, and,
Figure DEST_PATH_IMAGE010
is variance, independent of time;
the robot process automation data acquisition network finally outputs the predicted value of the minimized purchasing business volume
Figure 902915DEST_PATH_IMAGE011
。
The technical scheme adopted by the electronic equipment is as follows: an electronic device for forecasting procurement traffic based on a dynamic multivariate time series, comprising:
one or more processors;
a storage device for storing one or more programs that, when executed by the one or more processors, cause the one or more processors to implement the dynamic multivariate time series-based procurement traffic prediction method.
Compared with the prior art, the invention has the advantages and positive effects mainly embodied in the following aspects:
(1) The invention provides a purchasing business volume prediction method for a dynamic multi-element time sequence in robot process automation. The invention solves the problem of incomplete feature extraction in the traditional purchasing traffic prediction process, and the time convolution module provided by the invention extracts the nonlinear feature of the static time sequence.
(2) The invention solves the problems of gradient disappearance and gradient explosion caused by overlong time interval when long-period time sequence prediction is carried out by purchasing traffic prediction, and the proposed multi-stage attention module adaptively selects important characteristics to model the dependency relationship between states of long time interval.
(3) The invention solves the problem that a nonlinear module cannot process the prediction of the dynamic time sequence required in the purchasing traffic prediction, and the proposed autoregressive module processes the dynamic time sequence with unfixed period.
Drawings
FIG. 1 is a general framework diagram of a method according to an embodiment of the invention;
FIG. 2 is a schematic loss diagram of a Purchases data set model according to an embodiment of the present invention;
FIG. 3 is a NASDAQ100 data set model loss diagram of an embodiment of the present invention;
FIG. 4 is a schematic diagram of model prediction results on Purchases data sets according to an embodiment of the present invention;
FIG. 5 is a graph illustrating the prediction of the model on the NASDAQ100 dataset according to an embodiment of the present invention;
fig. 6 is a schematic diagram of the prediction effect of different time steps according to an embodiment of the present invention.
Detailed Description
For the purpose of facilitating understanding and implementing the invention by those of ordinary skill in the art, the invention is described in further detail below with reference to the accompanying drawings and examples, it being understood that the examples described herein are for purposes of illustration and explanation only and are not intended to be limiting.
Referring to fig. 1, the method for forecasting the purchasing traffic based on the dynamic multivariate time sequence provided by the present invention inputs the data obtained through the robot process automation data acquisition network into the robot process automation purchasing traffic forecasting network to forecast the purchasing traffic;
according to the robot process automation data acquisition network, index data required by service purchase quantity are extracted according to the keyword table, the data are sorted according to time, and time sequence data of N indexes are obtained and serve as an external sequence
Figure DEST_PATH_IMAGE012
Wherein N represents the number of indexes, and T is the length of the external sequence; each index is a variable, the firstnThe time series data of each index is the secondnSingle univariate time series data
Figure 20651DEST_PATH_IMAGE013
Is as followsnTime series data of the individual indices; and the time sequence data of all indexes form dynamic multi-element time sequence data which are used as input variables X of the robot process automatic purchase traffic prediction network.
The keyword table adopted in the embodiment is shown in table 1 below;
TABLE 1
Figure DEST_PATH_IMAGE014
The robot flow automatic procurement business volume prediction network comprises a time convolution module, a multi-stage attention module, a full connection layer and an autoregression module;
the time convolution module of this embodiment has K convolution kernels, and the size of each convolution kernel is F × M, where F is the width of the convolution kernel and M is the depth of the convolution kernel.
In this embodiment, a convolution neural module is first constructed, convolution kernels of the convolution neural module are set, and the size of each convolution kernel is F × M by K convolution kernels, where F is the width of the convolution kernel and M is the depth of the convolution kernel. Inputting the input variable X into K convolution kernels in sequence, inputting X into the second convolution kernelkOutput variable obtained by convolution kernel
Figure 350001DEST_PATH_IMAGE015
Wherein W is k Weight parameter of k convolution kernel represents convolution operation, RELU is nonlinear activation function, and RELU function is defined as
Figure DEST_PATH_IMAGE016
；b k Is a bias vector.
In this embodiment, after the time series data of N static indicators other than the index of the purchase traffic is input as the input external sequence X into the convolution kernel, the output feature sequence H with the size of P × K is finally obtained, where P ish k Is also the length of the univariate time series.
The multi-stage attention module of the embodiment comprises a first stage attention layer, a second stage attention layer and a time attention layer which are connected in sequence; the output variable H passes through the first stage attention layer, an important external sequence is selected in a self-adaptive mode for learning, the output external sequence of the first stage attention layer is connected with a target sequence corresponding to time through the second stage attention layer, finally the hidden states in the first two layers are combined through the time attention layer to learn the characteristics of longer time dependence, and the characteristics of longer time dependence are output
Figure 407956DEST_PATH_IMAGE017
；
In the first stage attention layer of this embodiment, the output signature sequence H of the previous step is used as the input sequence X of this step, and the memory cell state S at the time of the long short term memory network (LSTM) t-1 is determined according to the target sequence y without considering the target sequence y t-1 And hidden layer state h at time t-1 t-1 Constructing the kth input sequence X k Attention value at time t
Figure DEST_PATH_IMAGE018
：
Figure 139151DEST_PATH_IMAGE019
；
Wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE020
is a parameter that needs to be learned,
Figure 696297DEST_PATH_IMAGE021
in the form of a matrix of state weights,uis a hidden layer state h t-1 I.e., the number of hidden layer nodes of the long-short term memory network (LSTM),U m in order to input the weight matrix, the weight matrix is input,b m for the bias vector, tanh () is a nonlinear activation function,X k is as followskThe number of the input sequences is one,
Figure DEST_PATH_IMAGE022
；
according totInput attention of time of day
Figure 247364DEST_PATH_IMAGE023
Calculating attention weights
Figure DEST_PATH_IMAGE024
Calculating first stage attention output
Figure 843430DEST_PATH_IMAGE025
(ii) a Andthidden state at time of day
Figure DEST_PATH_IMAGE026
In which
Figure 662088DEST_PATH_IMAGE027
Is a long short term memory network LSTM (LSTM) used as an encoder.
The second stage attention layer of this embodiment constructs an input sequence of the second attention stage at time t, the secondkAn input sequence
Figure DEST_PATH_IMAGE028
Wherein isyThe sequence of interest is a sequence of interest,
Figure 154250DEST_PATH_IMAGE029
representing a vertical concatenation of matrices;
computing intermediate states of the second stage attention tier
Figure DEST_PATH_IMAGE030
：
Figure 254930DEST_PATH_IMAGE031
；
Wherein
Figure DEST_PATH_IMAGE032
Is a parameter to be learned;
Figure 326791DEST_PATH_IMAGE033
in the form of a matrix of state weights,uis a hidden layer state h t-1 The number of hidden layer nodes of the long short term memory network (LSTM),U s in order to input the weight matrix, the weight matrix is input,b s for the bias vector, tanh () is a nonlinear activation function;
according to intermediate state
Figure DEST_PATH_IMAGE034
Computing second stage weights
Figure 534043DEST_PATH_IMAGE035
：
Figure DEST_PATH_IMAGE036
；
According to input attention
Figure 931527DEST_PATH_IMAGE037
And weight
Figure DEST_PATH_IMAGE038
Computing the second stage attention tier output
Figure 519503DEST_PATH_IMAGE039
：
Figure DEST_PATH_IMAGE040
；
Wherein
Figure 893590DEST_PATH_IMAGE041
In order to be a hyper-parameter,
Figure DEST_PATH_IMAGE042
for the purpose of the second stage attention weighting,
Figure 453885DEST_PATH_IMAGE043
is the second stage input sequence.
The first two phases focus more on the relationship between sequences in the time series, and for better prediction of the target value, it is necessary to learn the correlation to the time series data for a longer time interval. This phase is a temporal attention phase, with temporal attention layers, learning longer-time dependencies based on the hidden states of the first two phases and the hidden state with the predicted target.
The temporal attention layer of the present embodiment first calculates the intermediate states
Figure DEST_PATH_IMAGE044
：
Figure 84586DEST_PATH_IMAGE045
；
Wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE046
in the form of a matrix of state weights,
Figure 97541DEST_PATH_IMAGE047
in order to input the weight matrix, the weight matrix is input,
Figure DEST_PATH_IMAGE048
for the bias vector, the hidden layer state of the long short term memory network (LSTM) as decoder at time t-1
Figure 278249DEST_PATH_IMAGE049
，
Figure DEST_PATH_IMAGE050
The long-short term memory network (LSTM) as an encoder in the first two stages is formed by transversely splicing the states of memory units at the time t-1,
Figure 693050DEST_PATH_IMAGE051
the memory unit state of the long-short term memory network (LSTM) at the current stage at the moment k,
Figure DEST_PATH_IMAGE052
output for the second attention stage;
according to intermediate state
Figure 494653DEST_PATH_IMAGE053
Calculating temporal attention weights
Figure DEST_PATH_IMAGE054
：
Figure 493439DEST_PATH_IMAGE055
；
Attention according to timeWeight of
Figure DEST_PATH_IMAGE056
Computing context vectorsC t ：
Figure 710794DEST_PATH_IMAGE057
；
According to context vectorC t Calculating
Figure DEST_PATH_IMAGE058
，
Figure 42418DEST_PATH_IMAGE059
Calculating the time attention stagetHidden layer state of long-short term memory network (LSTM) as decoder at time of day
Figure DEST_PATH_IMAGE060
In which
Figure 218185DEST_PATH_IMAGE061
For long short term memory networks (LSTM) to be used as decoders,
Figure DEST_PATH_IMAGE062
in the form of a matrix of state weights,
Figure 441618DEST_PATH_IMAGE063
is a bias vector;
calculating final output of a multi-stage attention module
Figure DEST_PATH_IMAGE064
：
Figure 728243DEST_PATH_IMAGE065
；
Wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE066
as a state weight matrix，
Figure 648794DEST_PATH_IMAGE067
In order to be a vector of the offset,
Figure DEST_PATH_IMAGE068
is the hidden state of the decoder at time t,C t is a context vector.
The autoregressive module of this embodiment outputs the dynamic multivariate time series data as an input vector X
Figure DEST_PATH_IMAGE069
；
Figure DEST_PATH_IMAGE070
；
Wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE071
in order to be a hyper-parameter,
Figure DEST_PATH_IMAGE072
is a learnable weight vector, X is an input vector,Uin order to input the weight matrix, the weight matrix is input,bin order to be a vector of the offset,
Figure DEST_PATH_IMAGE073
is the noise at the time of the t-time,
Figure DEST_PATH_IMAGE074
is variance, independent of time;
the robot process automation data acquisition network of the embodiment finally outputs the predicted value of the minimized purchasing traffic
Figure DEST_PATH_IMAGE075
. And finally, the robot process automation stores the prediction result to a corresponding position.
The robot process automation data acquisition network of the embodiment is a trained robot process automation data acquisition network; training processUsing a small batch stochastic gradient descent method and an ADAM optimizer, the objective function is to minimize the predicted value of procurement trafficy t And true valueyMean square error between
Figure DEST_PATH_IMAGE076
。
The invention is further illustrated by the following experiments. To test the performance of the present invention, two data sets were used, as shown in table 2.
TABLE 2
Data set Test set Test set Description of data set
Purchases 40000 10000 Shopping data
NASDAQ
100 32000 8560 Stock price data
Shopping data sets (Purchases): shopping data sets (Purchases) are Kaggle-based "get valuable shopper" challenge data sets that contain the shopping history of thousands of individuals, with each user record containing over a year transactions, including many fields such as product name, chain of stores, quantity, and date of purchase. In the experiment, the data was processed over time to obtain a time series shopping data set (containing 50000 samples), of which 4000 were used as training set and the rest were used as test set.
Nasdak 100 dataset (NASDAQ 100): this data set collected the stock prices of 81 companies in the NASDAQ100 index as a driving time series. The NASDAQ100 index serves as the target sequence. In the experiment, the first 32000 samples were used as training set, and the last 8560 samples were used as test set.
In order to measure the effectiveness of the purchasing traffic prediction method based on the dynamic multivariate time sequence, two evaluation indexes are used in the experiment: root Mean Square Error (RMSE) and mean square error (MAE). Suppose that
Figure DEST_PATH_IMAGE077
Is the firstiAt a time of a sampletThe true value of the time of day,
Figure DEST_PATH_IMAGE078
is the firstiAt a time of a sampletThe predicted value of the time model, RMSE, is defined as:
Figure DEST_PATH_IMAGE079
MAE is defined as:
Figure DEST_PATH_IMAGE080
the smaller the two values are, the better the prediction effect is, wherein N is the number of samples. Parameters, such as time step T of a time sequence, need to be set in the purchasing business volume prediction method based on the dynamic multi-element time sequence. In order to reduce the experimental error, 10 experiments were repeated, and the average of 10 experiments was taken as the final result.
As shown in Table 3, the dynamic multivariate time series-based prediction method for the procurement traffic volume is about 0.3 in both RMSE and MAE on two data sets, and the effect is better than that of the traditional method (LSTM). The method has different predictive effects on different data sets because the data characteristics of different data sets are different.
TABLE 3 evaluation results
Figure DEST_PATH_IMAGE081
Please refer to fig. 2 and fig. 3, which are schematic diagrams of Purchases data set model loss and NASDAQ100 data set model loss, respectively; as can be seen from fig. 2 and 3, the dynamic multivariate time sequence-based purchasing traffic prediction method has a good training convergence effect on two different data sets, and proves that the method has good prediction performance on both periodic and non-periodic time sequences. As can be seen from fig. 2 and 3, the training loss of the model decreases with the increase of the number of training rounds, which proves the effectiveness of the model in predicting the time series.
In order to observe the model prediction effect, the prediction results of the model on two different data sets are plotted in fig. 4 and 5. As can be seen from fig. 4 and 5, when the training data is sufficient, the model prediction result is very close to the real result. In addition, as the training time is increased, the superposition effect of the predicted value and the actual value of the model on the two data sets is better and better. The solid line represents the predicted value and the dotted line represents the true value.
In order to further optimize the parameters of the model, the model performance at different time steps on different data sets was tested. The experimental result is shown in fig. 6, and it can be seen from fig. 6 that, on the Purchases data set, when the time step is 4, the performance of the model is the best; on the NASDAQ100 s dataset, the performance of the model was best when the time step was 9.
The invention provides a purchasing business volume prediction method based on a dynamic multivariate time sequence in robot-oriented process automation. The invention solves the problem of incomprehensive manual feature extraction in purchasing business volume prediction, and the time convolution module provided extracts the nonlinear features of the static time sequence. The invention solves the problems of gradient disappearance and gradient explosion caused by overlong time interval when long-period time sequence prediction is carried out by purchasing traffic prediction, and the proposed multi-stage attention module adaptively selects important characteristics to model the dependency relationship between states of long time interval. The invention solves the problem that a nonlinear module cannot process the prediction of the dynamic time sequence in the purchasing traffic prediction, and the proposed autoregressive module processes the dynamic time sequence with unfixed period.
The invention can provide more accurate and comprehensive purchasing business volume prediction method for users in more fields of artificial intelligence, purchasing management and the like.
It should be understood that the above description of the preferred embodiments is given for clarity and not for any purpose of limitation, and that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (7)

1. A purchasing business volume prediction method based on dynamic multivariate time sequence is characterized in that: inputting data obtained through the robot process automation data acquisition network into a robot process automation purchasing business volume prediction network to predict purchasing business volume;
the robot process automation data acquisition network extracts index data required by service purchase quantity according to the keyword table, arranges the data according to time, and obtains time sequence data of N indexes as an external sequence
Figure 413476DEST_PATH_IMAGE001
Wherein N represents the number of indexes, and T is the length of the external sequence; each index is a variable, the firstnThe time series data of each index isnSingle univariate time series data
Figure 74265DEST_PATH_IMAGE002
Is a firstnTime sequence of individual indexesColumn data; the time sequence data of all indexes form dynamic multi-element time sequence data which are used as input variables X of the robot process automatic purchase business volume prediction network;
the robot process automatic purchasing business volume prediction network comprises a time convolution module, a multi-stage attention module, a full connection layer and an autoregressive module;
the time convolution module is characterized in that K convolution kernels are used, the size of each convolution kernel is F multiplied by M, wherein F is the width of each convolution kernel, M is the depth of each convolution kernel, time sequence data of other N static indexes except the purchasing traffic indexes are used as an input external sequence X to be input into each convolution kernel, and finally an output characteristic sequence H with the size of P multiplied by K is obtained, wherein P is P
Figure 764003DEST_PATH_IMAGE003
Is also the length of the univariate time series;
the multi-stage attention module comprises a first stage attention layer, a second stage attention layer and a time attention layer which are connected in sequence; the output characteristic sequence H passes through the first stage attention layer, an external sequence is selected in a self-adaptive mode for learning, the output external sequence of the first stage attention layer is connected with a target sequence corresponding to time through the second stage attention layer, finally the hidden states in the first two layers are combined through the time attention layer to learn characteristics dependent for longer time, and the characteristics dependent for longer time are output
Figure 494455DEST_PATH_IMAGE004
；
The autoregressive module takes the dynamic multivariate time sequence data as an input vector X and outputs the input vector X
Figure 558226DEST_PATH_IMAGE005
；
Figure 898072DEST_PATH_IMAGE006
；
Wherein, the first and the second end of the pipe are connected with each other,
Figure 832530DEST_PATH_IMAGE007
in order to be a hyper-parameter,
Figure 763577DEST_PATH_IMAGE008
is a learnable weight vector, X is an input vector,Uin order to input the weight matrix, the weight matrix is input,bin order to be a vector of the offset,
Figure 49064DEST_PATH_IMAGE009
is the noise at the time t, and,
Figure 192601DEST_PATH_IMAGE010
is variance, independent of time;
the robot process automation data acquisition network finally outputs the predicted value of the minimized purchasing business volume
Figure 981565DEST_PATH_IMAGE011
。
2. The method of claim 1, wherein the method for forecasting procurement traffic based on dynamic multivariate time series comprises: in the time convolution module, input variable X is input into K convolution kernels in sequence, and X is input into the second convolution kernelkOutput variable obtained by convolution kernel
Figure 316470DEST_PATH_IMAGE012
WhereinW k Is a firstkWeight parameter of convolution kernel represents convolution operation, RELU is nonlinear activation function, and RELU function is defined as
Figure 89253DEST_PATH_IMAGE013
；b k Is a bias vector.
3. The method of claim 1 for forecasting procurement traffic based on dynamic multivariate time seriesThe method is characterized in that: in the attention layer of the first stage, the output characteristic sequence H of the previous stage is taken as the input sequence X of the first stage, and the target sequence y is not considered, and the state S of the memory unit at the moment of the long-short term memory network LSTMt-1 is determined according to the state of the memory unit t-1 And hidden layer state h at time t-1 t-1 Constructing the kth input sequence X k Attention value at time t
Figure 36481DEST_PATH_IMAGE014
：
Figure 211110DEST_PATH_IMAGE015
；
Wherein the content of the first and second substances,
Figure 952801DEST_PATH_IMAGE016
is a parameter that needs to be learned,
Figure 212881DEST_PATH_IMAGE017
in the form of a matrix of state weights,uis a hidden layer state h t-1 The dimension of (a), i.e. the number of hidden layer nodes of the long-short term memory network LSTM,U m in order to input the weight matrix, the weight matrix is input,b m in order to be a vector of the offset,
Figure 26117DEST_PATH_IMAGE018
in order to be a non-linear activation function,X k is as followskThe number of the input sequences is one,
Figure 727356DEST_PATH_IMAGE019
；
according totInput attention of time
Figure 967845DEST_PATH_IMAGE020
Calculating attention weights
Figure 103771DEST_PATH_IMAGE021
Calculating the first stageAttention output
Figure 517435DEST_PATH_IMAGE022
(ii) a Andthidden state at time of day
Figure 276443DEST_PATH_IMAGE023
Wherein
Figure 750150DEST_PATH_IMAGE024
Is a long short term memory network LSTM used as an encoder.
4. The method of claim 3, wherein the method for forecasting procurement traffic based on dynamic multivariate time series comprises: the second stage attention layer is constructed in a second attention stagetInput sequence of moments, nokAn input sequence
Figure 860188DEST_PATH_IMAGE025
Wherein isyThe sequence of the object is determined,
Figure 811964DEST_PATH_IMAGE026
representing a vertical concatenation of matrices;
computing an intermediate state of the second stage attention tier
Figure 425479DEST_PATH_IMAGE027
：
Figure 70087DEST_PATH_IMAGE028
；
Wherein
Figure 900377DEST_PATH_IMAGE029
Is a parameter to be learned;
Figure 655844DEST_PATH_IMAGE030
in the form of a matrix of state weights,uis a hidden layer state h t-1 The dimension of (a), i.e. the number of hidden layer nodes of the long-short term memory network LSTM,U s in order to input the weight matrix, the weight matrix is input,b s for the bias vector, tanh () is a nonlinear activation function;
according to the intermediate state
Figure 389445DEST_PATH_IMAGE031
Computing second stage weights
Figure 204954DEST_PATH_IMAGE032
：
Figure 351901DEST_PATH_IMAGE033
；
According to input attention
Figure 317583DEST_PATH_IMAGE034
And weight
Figure 233587DEST_PATH_IMAGE035
Computing the second stage attention tier output
Figure 95363DEST_PATH_IMAGE036
：
Figure 526345DEST_PATH_IMAGE037
；
Wherein
Figure 445DEST_PATH_IMAGE038
In order to be a hyper-parameter,
Figure 567692DEST_PATH_IMAGE039
for the purpose of the second stage attention weighting,
Figure 600371DEST_PATH_IMAGE040
is the second stage input sequence.
5. The method of claim 4, wherein the method for forecasting procurement traffic based on dynamic multivariate time series comprises the following steps: the temporal attention layer, first computing intermediate states
Figure 518648DEST_PATH_IMAGE041
：
Figure 29395DEST_PATH_IMAGE042
；
Wherein the content of the first and second substances,
Figure 716728DEST_PATH_IMAGE043
in the form of a matrix of state weights,
Figure 920308DEST_PATH_IMAGE044
in order to input the weight matrix, the weight matrix is input,
Figure 325881DEST_PATH_IMAGE045
for the bias vector, the long-short-term memory network LSTM as decoder has hidden layer state at t-1
Figure 404433DEST_PATH_IMAGE046
，
Figure 946273DEST_PATH_IMAGE047
The long-short term memory network LSTM as an encoder in the first two stages is formed by transversely splicing the states of memory units of the long-short term memory network LSTM at the time t-1,
Figure 320754DEST_PATH_IMAGE048
the memory unit state of the long-short term memory network LSTM at the moment k at the current stage,
Figure 948044DEST_PATH_IMAGE049
for the second attentionOutputting force stage;
according to the intermediate state
Figure 800594DEST_PATH_IMAGE050
Calculating temporal attention weights
Figure 462519DEST_PATH_IMAGE051
：
Figure 335797DEST_PATH_IMAGE052
；
Attention weighting according to time
Figure 325750DEST_PATH_IMAGE053
Computing context vectorsC t ：
Figure 106624DEST_PATH_IMAGE054
；
According to context vectorC t Calculating
Figure 999887DEST_PATH_IMAGE055
，
Figure 824493DEST_PATH_IMAGE056
Calculating the time attention stagetHidden layer state of long-short term memory network LSTM as decoder at time of day
Figure 426375DEST_PATH_IMAGE057
Wherein
Figure 886307DEST_PATH_IMAGE058
For the long short term memory network LSTM to be used as a decoder,
Figure 257245DEST_PATH_IMAGE059
is the status rightThe weight of the matrix is determined by the weight,
Figure 144430DEST_PATH_IMAGE060
is a bias vector;
calculating final output of a multi-stage attention module
Figure 968029DEST_PATH_IMAGE061
：
Figure 231651DEST_PATH_IMAGE062
；
Wherein, the first and the second end of the pipe are connected with each other,
Figure 974873DEST_PATH_IMAGE063
in the form of a matrix of state weights,
Figure 157592DEST_PATH_IMAGE064
in order to be a vector of the offset,
Figure 609433DEST_PATH_IMAGE065
is the hidden state of the decoder at time t,C t is a context vector.
6. The method for forecasting procurement traffic based on dynamic multivariate time series as claimed in any one of claims 1 to 5, characterized in that: the robot process automatic data acquisition network is a trained robot process automatic data acquisition network; the training process adopts a small batch random gradient descent method and an ADAM optimizer, and the objective function is the predicted value of the minimized purchasing trafficy t And true valueyMean square error of
Figure 535801DEST_PATH_IMAGE066
。
7. An electronic device for forecasting procurement traffic based on a dynamic multivariate time series, comprising:
one or more processors;
a storage device for storing one or more programs that, when executed by the one or more processors, cause the one or more processors to implement the dynamic multivariate time series based procurement traffic prediction method of any of claims 1-6.
CN202211126215.9A 2022-09-16 2022-09-16 Purchasing business volume prediction method based on dynamic multivariate time sequence and electronic equipment Active CN115204535B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211126215.9A CN115204535B (en) 2022-09-16 2022-09-16 Purchasing business volume prediction method based on dynamic multivariate time sequence and electronic equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211126215.9A CN115204535B (en) 2022-09-16 2022-09-16 Purchasing business volume prediction method based on dynamic multivariate time sequence and electronic equipment

Publications (2)

Publication Number Publication Date
CN115204535A CN115204535A (en) 2022-10-18
CN115204535B true CN115204535B (en) 2022-12-06

Family

ID=83572324

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211126215.9A Active CN115204535B (en) 2022-09-16 2022-09-16 Purchasing business volume prediction method based on dynamic multivariate time sequence and electronic equipment

Country Status (1)

Country Link
CN (1) CN115204535B (en)

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8032401B2 (en) * 2007-07-18 2011-10-04 General Electric Company System and method to calculate procurement of assets
CN110163401B (en) * 2018-02-12 2023-05-26 腾讯科技（深圳）有限公司 Time sequence prediction method, data prediction method and device
US11681914B2 (en) * 2020-05-08 2023-06-20 International Business Machines Corporation Determining multivariate time series data dependencies
WO2021257610A1 (en) * 2020-06-15 2021-12-23 Spartan Capital Intelligence, Llc Time series forecasting and visualization methods and systems
CN111880998B (en) * 2020-07-30 2022-09-02 平安科技（深圳）有限公司 Service system anomaly detection method and device, computer equipment and storage medium
CN114493014A (en) * 2022-01-28 2022-05-13 湖南大学 Multivariate time series prediction method, multivariate time series prediction system, computer product and storage medium
CN114881354A (en) * 2022-05-31 2022-08-09 支付宝(杭州)信息技术有限公司 Method and device for predicting multivariate time series
CN114818515A (en) * 2022-06-24 2022-07-29 中国海洋大学 Multidimensional time sequence prediction method based on self-attention mechanism and graph convolution network

Also Published As

Publication number Publication date
CN115204535A (en) 2022-10-18

Similar Documents

Publication Publication Date Title
Guan et al. An adaptive neuro-fuzzy inference system based approach to real estate property assessment
CN111160968A (en) SKU-level commodity sales prediction method and device
CN109191276B (en) P2P network lending institution risk assessment method based on reinforcement learning
Li et al. Multi-factor based stock price prediction using hybrid neural networks with attention mechanism
Adeyemo et al. Effects of normalization techniques on logistic regression in data science
Dat et al. Hybrid online model based multi seasonal decompose for short-term electricity load forecasting using ARIMA and online RNN
Goyal et al. A solution to forecast demand using long short-term memory recurrent neural networks for time series forecasting
Yarushev et al. Review of studies on time series forecasting based on hybrid methods, neural networks and multiple regression
Petkovic et al. Deep learning for spatio‐temporal supply and demand forecasting in natural gas transmission networks
Srivastava et al. Stock market prediction using RNN LSTM
Stach et al. Evolutionary development of fuzzy cognitive maps
CN108765137A (en) A kind of credit demand prediction technique and system, storage medium
Wu et al. A forecasting model based support vector machine and particle swarm optimization
CN115204535B (en) Purchasing business volume prediction method based on dynamic multivariate time sequence and electronic equipment
Pardeshi et al. Stock market price prediction: a hybrid LSTM and sequential self-attention based approach
Pradhan et al. Application of neural network in prediction financial viability
Singhania et al. Forecasting the United States unemployment rate by using recurrent neural networks with Google Trends data
CN111461815A (en) Order recognition model generation method, recognition method, system, device and medium
AmirAskari et al. A modified fuzzy relational model approach to prediction of foreign exchange rates
KR102409041B1 (en) portfolio asset allocation reinforcement learning method using actor critic model
Sun et al. Short-term stock price forecasting based on an svd-lstm model
Cai et al. XGBoost Analysis based on Consumer Behavior
M John et al. Predicting House Prices using Machine Learning and LightGBM.
Liu et al. A Novel Algorithm for Credit Default Prediction using TabNet
Chengwei Research on Performance Evaluation System of Manufacturing Listed Companies Based on CNN

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
GR01 Patent grant
GR01 Patent grant