KR20190013038A - System and method for trend predicting based on Multi-Sequences data Using multi feature extract technique - Google Patents

Abstract

The present invention relates to a multi-sequence data prediction system using a multi-feature extraction technique and a prediction method thereof. According to the present invention, the prediction method comprises: a step of analyzing an association rule pattern; a step of determining an association rule pattern with a high prediction possibility; a step of generating selected data as a prediction model; and a step of using the generated prediction model to predict a future trend upon generation of an associated pattern. Accordingly, a trend changed after generation of a feature pattern of sequential data is able to be precisely predicted.

Description

TECHNICAL FIELD The present invention relates to a system and method for predicting multiple time series data using multiple feature extraction techniques,

The present invention relates to a system and method for predicting trends of multiple time series data using a multiple feature extraction technique in the field of time series pattern analysis, and more particularly, to a system and method for predicting trends of multiple time series data by defining a time series- The present invention relates to a system and method for predicting trends of multiple time series data by extracting and patterning features, analyzing association rule patterns among multiple time series data by combining patterns, and generating prediction models.

Time series data analysis is an analysis that uses data mining technology and statistical analysis technology to build a model using past data without any prior knowledge of measurement from disordered and various time series data and to predict future using model.

Also, in forecasting through time series analysis, the observed historical data are analyzed, and regular patterns are found and modeled.

However, existing analytical methods have attempted to predict patterns by extracting patterns considering only absolute value and shape, and analyzing a model limited to a single time series rather than the inherent characteristics of a time series to be predicted.

As a result, conventional time-series data analysis has a limitation that it can be predicted only for regularly defined and regular patterns that are predefined, that is, frequently found, and can be predicted only for time series data having the same characteristics as a single time series analyzed .

Various kinds of time series data are generated through the development and industrialization of the industry, and the necessity of a complex correlation analysis is increasing for a wide variety of time series. In order to solve the limitation using the existing single time series analysis method, Efficient and time-series data analysis methods such as extraction and correlation analysis are needed.

According to an aspect of the present invention, there is provided a system and method for receiving various types of data having time series data attributes and extracting features representing characteristics of the corresponding time series and modeling and associating the association with other time series using the features .

According to another aspect of the present invention, there is provided a method for analyzing and extracting common attributes of time series data, including a feature extraction unit for defining features and a model for improving accuracy of association analysis between various time series data using a time lag And a method for predicting the same.

In order to accomplish the above object, the present invention provides a time-series data input unit for sequentially receiving time series desired to be analyzed; A data refining unit for reducing a difference between absolute values of the input time series data; A feature extraction unit that finds and defines a characteristic of the purified data; A similar feature pattern cluster unit that performs feature pattern clustering by characterizing a set of similarity patterns and a weight calculation using the appearance frequency in the time and feature patterns appearing in the time series data; A cluster storage unit for storing the feature pattern cluster data of the input single time series data on a cluster time basis; A multiple time series input unit for receiving and analyzing pattern information of the analyzed single time series for multiple time series analysis; A relevance analysis and association determination unit for performing association analysis using the AprioriAll algorithm for the input multiple time series pattern sets and determining whether there is an association according to the degree of association; A prediction model generation unit for modeling and storing a plurality of time series data of a plurality of time series data; And a trend predicting unit for predicting a trend by reflecting the generated multiple time series pattern model.

The data refiner may normalize the time series data information received from the time series data input unit to a value between 0 and 1 by removing an ideal value using a unit of a value and an ideal value by using the following equation.

(Equation 1 - outlier removal formula)

TS is time series data from which anomalies are removed, a is a sequence compression factor, t is a time series value in time units, and τ is an outlier removal time series period.

(Equation 2 - normalization formula)

는 정규화된 시계열 데이터, TS는 이상치가 제거된 시퀀스 데이터, μ는 이상치가 제거된 시퀀스 데이터의 평균이다.

Is the normalized time series data, TS is the sequence data from which the ideal value is removed, and is the average of the sequence data from which the ideal value is removed.

The feature extracting unit extracts features based on the following three criteria to find the unique characteristics possessed by the time series data. First, the time series feature extraction method is a feature extraction that can grasp the rate of change of the time series data and the degree of change scale, and it is calculated using the following equation.

(Equation 3 - Variation Rate and Scale Definition Equation)

VM defines the change rate and scale definition feature value,

는 시계열 데이터에서 t와 t-1간의차이 값들 중 0 미만의 값들의 평균값,

Is an average value of values less than 0 among the difference values between t and t-1 in time series data,

는 시계열 데이터에서 t와 t-1간의 차이 값들 중 0 이상의 값들의 평균값이다.

Is an average value of zero or more of the difference values between t and t-1 in time series data.

Second, the time series feature extraction method is feature extraction that can detect the volatility of time series data, and it is calculated by using the following formula.

(Equation 4 - Variability Feature Definition Equation)

는 시간 기준 t와 t-1의 변화율,

는

의 평균 값이다.VolTS is a volatility definition characteristic value, N is a time series data length for calculating volatility,

Is the rate of change of the time base t and t-1,

The

.

Third, the time series feature extraction method is a partial time series type feature extraction, and the feature values are the top 10 partial time series types having high frequency among the time series data separated into n days. The partial sequence extraction formula is as follows.

(Equation 5 - partial time series form extraction definition formula)

SubTS is the partial time series type definition feature value, i is the start value in time unit, and n is the end value in time unit.

On the other hand, the feature pattern cluster generating unit uses a full graph clustering technique by making three feature vectors extracted from the feature extracting unit into a three-dimensional vector, and clusters similar pattern sets among the three-dimensional pattern sets after clustering. The clustering criterion is cluster based on the frequency at which the feature occurred and the frequency at which the feature occurred. The feature that resulted from the frequency-based cluster is that the feature with high frequency has a high weight and is used in the correlation measurement part in the correlation analysis part among multiple time series.

In the present invention, the relevance analysis and relevance determining unit may collect the characteristics of data using the individual time series cluster data generated by the cluster generating unit, and associate the time lag and the association analysis between multiple time series data using the AprioriAll algorithm. It includes analysis.

As described above, according to the present invention, various time series data such as weather temperature, stock price, exchange rate, economic index, and the like are independently extracted as change rate, scale, variability, and time series shape, And a trend prediction model can be constructed by analyzing the correlation between the characteristic patterns of the time series data and the time series data.

In other words, it is possible to analyze the time series to other characteristics through the independent feature extraction, and it can be used for the correlation analysis by setting the weight on the feature set through the clustering based on the frequency. , And it has an advantage of improving the accuracy of prediction by analyzing the pre-action and post-action between time series using time lag.

Meanwhile, since the effect of the present invention is represented by a multi-time series data analysis method and system, it can be utilized in various industrial fields requiring various time series data analysis.

1 is a block diagram of a multiple time series data based trend prediction system in accordance with an embodiment of the present invention.
2 is an internal configuration diagram of the data refining unit.
3 is an internal configuration diagram of the feature extraction unit.
4 is a flowchart of a multiple time series data-based trend prediction method according to another embodiment of the present invention.
5 is a diagram showing a process of clustering and characterizing a set of similar feature patterns in a display unit.
6 is a diagram illustrating a procedure for analyzing association between multiple time series data using a time delay technique.
7 is a diagram showing an example of predicting using an association rule pattern included in a prediction model.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention, and how to accomplish them, will become apparent by reference to the embodiments described in detail below with reference to the accompanying drawings.

Further, the present invention is not limited to the embodiments disclosed below, but may be implemented in various other forms.

The present embodiments are provided so that the disclosure of the present invention is thoroughly disclosed and that those skilled in the art will fully understand the scope of the present invention.

And, the present invention is only defined by the claims.

Thus, in some embodiments, well known components, well known operations, and well-known techniques are not specifically described to avoid an undesirable interpretation of the present invention.

In addition, throughout the specification, like reference numerals refer to like elements, and the terms referred to herein are for describing the embodiments and are not intended to limit the present invention.

In this specification, the singular forms include plural forms unless the context clearly dictates otherwise, and components and acts referred to as " comprising (or collectively) " means that one or more other components and actions are performed by one or more other components and operations Or < / RTI >

Unless defined otherwise, all terms (including technical and scientific terms) used herein may be used in a sense commonly understood by one of ordinary skill in the art to which this invention belongs.

As such, commonly used predefined terms are not ideally or excessively interpreted unless defined otherwise.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

In the following description, the system and the analysis method according to the present invention are applied to a stock price trending system using financial data and weather data.

Referring to FIGS. 1 to 4, a multiple time series data-based trend prediction system 100 using a multiple feature extraction technique includes an individual time series feature cluster unit 200 for individually inputting time series feature extraction and feature clusters, And a multiple time series correlation analysis prediction unit 300 for generating a correlation analysis and a prediction model among multiple time series by combining the analysis results of the individual time series characteristic cluster unit.

Weather data and financial data (input data for forecasting trends using domestic and overseas stock indices, domestic and overseas economic indicators, etc.) are used to determine the model by analyzing the meaningful relationships between the data using defined time series data.

Here, the time series concept refers to data recorded at regular intervals according to the time. The individual time series feature clustering system 200 extracts features based on time, and clusters of similar features are subjected to multiple time series correlation analysis The predictor 300 analyzes the association and models it, and predicts future trend values using the model.

Here, the individual time-series feature clustering system 200 is a system for independently analyzing multiple time-series data to extract features, and collecting and clustering similar features. In addition, the apparatus includes a time series data input unit 210, a data refining unit 220, a feature extraction unit 230, a feature pattern cluster generation unit 240, and a cluster storage unit 250 to perform the above process.

The time series data input unit 210 receives the entire period of the time series data for analyzing the time series data as an input part and receives the input data according to the raw data of the database and file format.

Here, the weather data is input to the database, and the domestic stock index data is set to the file format, and the input format is set to {date, value}.

The data refinement unit 220 standardizes the time series data for comparison with other time series because the type of the extreme value of the input time series data is different. The standardization process is performed sequentially, and includes a DWT conversion unit 221 and a normalization unit 222.

2 is an internal configuration diagram of the data refining unit.

The data refinement unit 220 performs a DWT transformation 221 to remove an ideal value of the inputted time series data and perform smoothing. The DWT transform 221 is a high frequency filter using a wavelet transform and smoothes the sequence by eliminating the residual wave such as a high frequency. The smoothing operation can be performed using the following equation (1).

T denotes a time series value of the time series data for smoothing, a time series value of the time series data, and? Denotes a partial time series to be smoothed .

Next, the normalization unit 222 receives the smoothed data from which the ideal value is removed, and converts the absolute value of the time-series data into a normalized value between 0 and 1. The normalization operation can be performed using the following equation (2).

는 정규화된 시계열 데이터를 나타내며, TS는 이상치가 제거된 평활화된 시계열 데이터, μ는 평활화된 시계열 데이터의 평균값을 나타낸다.here,

TS denotes normalized time series data, TS denotes smoothed time series data from which an ideal value is removed, and μ denotes an average value of smoothed time series data.

Specifically, when the temperature data of the weather data and the time series refining period are input to the data refining unit 220, the DWT converting unit 221 removes the abnormal value of the time series data within the designated period, extracts the smoothed data, ), A value normalized to a value between 0 and 1 is output. Similarly, when the closing price data of the KOSPI index of the financial data and the refinement period are input, a value normalized to a value between 0 and 1 is output by a refining method similar to the data refining unit.

The feature extraction unit 230 receives the tabular data output from the data refinement unit 220, and extracts unique characteristics of the time series data from the change rate and scale feature extraction unit 231, the variance feature extraction unit 232, And extracts the time-series form extraction unit 233 by using the time-

Here, the characteristic inherent in the time series data is the property of the change that the time series data having the time attribute has by changing the time. In the feature extraction unit of the present invention, the characteristics of the change are defined by the change scale, the speed, the variability, and the changing time series.

3 is an internal configuration diagram of the feature extraction unit.

The change rate and scale feature extraction unit 231 defines a change rate and a change scale in order to extract characteristics of intensity of a change amount proportional to time among common attributes of time series data. The rate of change is the total amount on the average of the difference between the falling change amount and the rising change amount in the same period, and the change scale is the total amount changed within the period based on the change amount of all the time series data values between the start point and the end point of the partial time series period . The change rate and scale feature extraction unit 231 extracts the change rate and scale feature using the following equation (3).

는 시계열 데이터에서 t와 t-1간의차이 값들 중 0 미만의 값들의 평균값,

는 시계열 데이터에서 t와 t-1간의 차이 값들 중 0 이상의 값들의 평균값을 나타낸다.Here, VM represents the change rate and scale definition feature value,

Is an average value of values less than 0 among the difference values between t and t-1 in time series data,

Represents an average value of zero or more of the difference values between t and t-1 in the time-series data.

The feature extraction unit 230 receives the tabular data output from the data refinement unit 220, and extracts unique characteristics of the time series data from the change rate and scale feature extraction unit 231, the variance feature extraction unit 232, And extracts the time-series form extraction unit 233 by using the time-

Here, the characteristic inherent in the time series data is the property of the change that the time series data having the time attribute has by changing the time. In the feature extraction unit of the present invention, the characteristics of the change are defined by the change scale, the speed, the variability, and the changing time series.

3 is an internal configuration diagram of the feature extraction unit.

The change rate and scale feature extraction unit 231 defines a change rate and a change scale in order to extract characteristics of intensity of a change amount proportional to time among common attributes of time series data. The rate of change is the total amount on the average of the difference between the falling change amount and the rising change amount in the same period, and the change scale is the total amount changed within the period based on the change amount of all the time series data values between the start point and the end point of the partial time series period . The change rate and scale feature extraction unit 231 extracts the change rate and scale feature using the following equation (3).

Specifically, the change rate and scale feature extraction unit 231 divides the temperature time series data received from the data refinement unit 220 into partial time series divided by the start time and the end time, The feature value is extracted by calculating the average value of the variation amount difference value. In this case, in order to continuously extract the change rate and scale characteristics of the partial time series, the start time and end time of the partial time series can be changed and applied to the equation (3).

The feature values extracted as above are calculated as the ratio (%). The closer to 0 (%), the less or less the change rate and scale in the period, and the closer to 100 (%) the change rate and scale It means faster or more.

Next, the variance feature extracting unit 232 extracts the property and the degree of change in the property of the time series data in proportion to the time, as a feature extracting unit for extracting the volatility among the common attributes of the time series data. The nature and extent of the change are calculated using equation (4) as the degree of change extraction regardless of the direction of the time series data.

는 시간 기준 t와 t-1의 변화율,

는

의 평균값을 나타낸다.Volts denotes the value of the volatility definition characteristic, N denotes a time series data length for calculating the volatility,

Is the rate of change of the time base t and t-1,

The

.

Specifically, the volatility characteristic extracting unit 232 specifies a time series length for calculating the variability in the temperature time series data transmitted from the data refining unit 220, and then extracts the characteristic using the standard deviation and average of the time series.

Then, the extracted variability feature value defines the normalized value obtained by substituting the normal distribution sigma value as the final variance feature value.

Next, the partial time-series form extraction unit 233 extracts a time-series change form according to a change in time with a feature extraction unit for extracting a form among common attributes of time series data. The shape change of the time change is calculated by changing the start point by 1 day or N days by specifying the start point and the end point and extracting the shape by using Equation (5).

SubTS denotes a partial time series type definition feature value, i denotes a time unit value at the start time, n denotes a time unit value at the end time, and SubTS_Array denotes a partial time series type set of time series data.

Specifically, the partial time-series form extraction unit 233 extracts a start time unit value (i) for specifying the end point of the start point in order to extract the form of the partial time series from the temperature time-series data received from the data refinement unit 220, The endpoint unit value (n) and the unit value for the viewpoint change are specified, and the partial time series is extracted and the final partial time series type feature is calculated by the type set (SubTS_Array).

The characteristic pattern cluster generation unit 240 generates a characteristic pattern cluster based on a pattern parameter that reconfigures the characteristic transferred from the characteristic extraction unit 230 into a three-dimensional vector and patterned, for example, [change speed and scale characteristic data, Pattern data] is applied to the full graph clustering algorithm to cluster the feature patterns and characterize the clustered pattern sets.

The full graph clustering algorithm is a method of calculating using the graph clustering technique, and is well known to those skilled in the art, so a detailed description thereof will be omitted.

The clustering in the pattern cluster generating unit 240 clusters based on the time attribute values of the feature values delivered by the feature extracting unit 230. [ The characteristic of the time included in the characteristic value of the temperature data is the characteristic of the same time zone of another day, for example, the characteristic value from 3:00 pm to 6:00 pm on January 20, the characteristic from 3:00 pm to 6:00 pm on January 21 Clusters the values.

The criterion for clustering in the pattern cluster generation unit 240 clusters the feature values transmitted in the feature extraction 230 on the basis of the similar feature frequency. The feature having the similar feature frequency has a high weight and is used for the association measurement weight in the association determining unit 330 association between the multiple time series.

FIG. 5 shows a process of clustering and characterizing feature vectors in the feature pattern cluster generator 240.

The cluster storage unit 250 stores the set of feature patterns generated by the feature pattern cluster generation unit 240 in the cluster DB 261.

The multiple time series correlation analysis predicting unit 300 analyzes the association between multiple time series using each set of individual time series feature patterns classified in the time series feature cluster unit 200, A predictive model is created between time series, and the trend after the feature pattern is predicted using the predictive model. The relevance analyzing unit 320, the relevancy determining unit 330, the model generating unit 340, and the trend predicting unit 350 may be included in the apparatus of the present invention.

The multiple time series input unit 310 receives the input time series data from the cluster database 251 storing the individual time series data feature pattern sets extracted from the time series feature cluster unit 200 and transfers the input data to the relation analysis unit 320. At this time, the time sequence of the inputted time series data is arranged in ascending order.

The association analyzing unit 320 generates an association pattern using a time lag and an association rule extraction AprioriAll algorithm from the data received from the multiple time series input unit 310, And transmits it to the determination unit 330.

In addition, the association analyzing unit 320 analyzes the pattern received from the multiple time series input unit 310 using a time delay technique, and performs association analysis on a time basis using a combination of temperature feature set data and KOSPI feature feature set data at the same time And the time reference of data collected in the same temperature by the same data is changed to a day before, and when the time reference of the KOSPI index feature set data is changed from the same data to a day before, the time is different according to the number of transmitted time series The AprioriAll algorithm is used to create a set of cases and extract the support, confidence, sequence, and rule set of all association sets.

The AprioriAll algorithm calculates the support for all sets of items and outputs all association rules regardless of their support and confidence values. In addition, the AprioriAll algorithm is used in the association determining unit 330 to extract the unconstrained association rule from the existing Apriori algorithm.

Referring to FIG. 5, the association analyzer 320 will be described in detail.

The relevance analysis unit 320 arranges the feature pattern sets of the multiple time series data received from the multiple time series input unit 310 in chronological order and arranges them in sets of the same time series among multiple time series (P01). Then, through the process of extracting the association rule and frequency of the feature set through the time delay of one day and the process of extracting support and reliability and value according to the occurrence frequency (P02) . In addition, the association rule patterns P03 and P04 are generated with a time delay of 2 days and 5 days in the same manner as the time delay method.

The association determining unit 330 receives the association rule pattern generated by the association analyzing unit 320 and determines a meaningful association rule pattern as a prediction model.

Specifically, the association rule pattern determined by the association determining unit 330 may be an association rule pattern, a support rule, and an association rule, which are transmitted from the association analyzing unit 320 using the pattern set weight value calculated by the feature pattern cluster generating unit 240. [ ), Confidence values, and finally, an association rule pattern between multiple time series to be used as a prediction model is selected. Equation (6) includes a method of selecting an association rule pattern to be used as a prediction model based on weighted support and weighted reliability.

는 가중 지지도를 나타내며,

는 가중 신뢰도를 나타낸다. 또한, 계산된 가중 지지도는 최소 지지도 α 값 이상인 값만 수용하며, 미만인 값은 0이 되며, 가중 신뢰도는 최소 신뢰도 β값 이상인 값만 수용하며, 미만인 값은 0이 된다.

Represents the weighted support,

Represents the weighted reliability. Also, the calculated weighted support accepts only values that are greater than or equal to the minimum support α, while a value less than 0 accepts only the value that is greater than or equal to the minimum reliability β.

The prediction model generation unit 340 stores an association rule pattern determined by the association determination unit 330. [

The trend predicting unit 350 predicts the trend by comparing the current feature pattern of the time series data to be predicted using the association rule pattern prediction model stored in the prediction model generating unit 340.

The trend prediction unit 350 will be described in detail with reference to FIG.

If a characteristic pattern 'K' of time series data to be predicted is input to the trend predicting unit 350, a multiple time series association rule pattern is classified (P05) from a prediction model, and a future time series association rule Predict the trend (P06).

Referring to FIG. 4, a multi-time series data-based trend prediction method using the multi-feature extraction technique of the present invention will be described. A detailed description of the parts overlapping with those described above will be omitted.

First, several time series data having a time attribute for predictive analysis are selected (S10).

Next, the selected data is individually inputted through the time series data input unit 210 and the data is refined to standardize the range of the values (S20). In the time series data normalization, the normalization unit 222 removes the ideal value from the DWT conversion unit 221 and normalizes the value to a value between 0 and 1.

Next, the characteristics representing the time series characteristic of the purified data are extracted (S20).

Specifically, the feature extraction unit 230 defines a time series feature as a change rate and scale feature, a volatility feature, and a partial time series feature, and extracts a common defined feature.

Next, the features received in the time series data feature pattern extracting step (S20) are clustered on the basis of time and frequency, and are clustered in a similar feature pattern unit (S30).

In step S40, it is determined whether the feature pattern and the similar feature pattern comprehensive cluster extraction for all the selected multiple time series data are completed.

If feature extraction and clustering for all the multiple time series data are completed, the association rule patterns are classified using the AprioriAll algorithm for individual time series data clustered in the similar feature pattern unit (S50).

Finally, an association rule pattern having a high probability of prediction is selected using the weight support value and weight reliability value among the classified association rule patterns, and a prediction model is generated with the selected patterns (S60). Also, a trend prediction can be attempted when an association rule pattern occurs in multiple time series using the generated prediction model.

As described above, according to the present invention, when individual time series data are inputted, the association between the multiple time series data is analyzed using the association rule pattern for the pattern set by extracting and clustering the unique characteristic of the time series, , And it is possible to predict trends.

It will be apparent to those skilled in the art that many other modifications and applications are possible within the scope and spirit of the present invention.

100: Multiple Time Series Data Prediction System Using Multiple Feature Analysis
200: Dogubil time series data characteristic Cluster part
210: Time series data input unit
220: Data refining unit
230: Feature extraction unit
240: Feature pattern cluster generation unit
250: Cluster storage unit
300: Multiple time series correlation analysis prediction unit
310: Multiple time series input unit
320: Association Analysis Unit
330: Association determining section
340:
350:

Claims (8)

A time series data input unit sequentially receiving a plurality of time series data;
A data refinement unit for eliminating and standardizing a difference between different absolute values of the input time series data;
A feature extraction unit for defining and extracting three common characteristics of the refined time series data (a change rate, a change scale characteristic, a variability feature, and a partial time series shape feature);
A feature pattern cluster generation unit for performing a feature pattern cluster by characterizing a set of similarity patterns and a weight calculation using a frequency at which the extracted features appear in the time series data and an appearance frequency in the feature pattern;
A cluster storage unit for storing the feature pattern cluster data according to a feature pattern generation time reference and a frequency number;
A plurality of time series input units for receiving the feature pattern cluster information of one time series stored in the cluster storage unit as a plurality of time series feature pattern cluster data;
A relevance analyzer for inferring an association rule pattern using an association rule analysis (AprioriAll) algorithm for data received from the multiple time series input unit;
An association determining unit for determining an association rule pattern to be used as a prediction model by using an association rule pattern inferred by the association analyzing unit and a feature pattern frequency weight value in an association rule pattern;
A prediction model generation unit for storing a prediction model by combining the association rule patterns determined in the association determination unit;
A trend predicting unit for predicting a trend using the association rule pattern prediction model stored in the prediction model generation unit;
And a plurality of time series data trend prediction systems using multiple feature analysis. The method according to claim 1,
The data-
In order to reduce the anomalous value between the time series data transmitted to the time series data input unit and ensure the commonality of the time series, the time series data obtained by sequentially removing the calculation abnormal values by using the discrete wavelet transform and the normalization method is divided into 0 and 1 To the feature extraction unit. The system for estimating multiple time series data trends using multi-feature analysis. The method according to claim 1,
The feature extraction unit may extract,
Three types of feature extraction techniques, a volatility feature extraction technique, and a partial time series extraction technique, which extract characteristics of a time attribute, which is a time series characteristic common to the refined time series data received from the data refining section, And a time series characteristic is extracted using the feature analysis. The method according to claim 1,
Wherein the feature pattern cluster generating unit comprises:
The features extracted from the feature extraction unit (the change rate, the change scale feature, the variability feature, and the partial time series shape feature) are reconstructed into a three-dimensional vector, and the similar pattern weight value extraction through the clustering based on the same time reference and the clustering based on the frequency reference A full graph algorithm is applied and extracted, and a set of similar feature patterns extracted are characterized to generate a final similarity cluster, and a multiple time series data trend prediction system using multiple feature analysis. The method according to claim 1,
The association analyzing unit,
A plurality of feature sets of the individual time series data generated by the feature pattern cluster generation unit are received from the multiple time series data input unit, and an association rule extraction considering a time difference using a time lag technique, The association rule pattern is analyzed using the AprioriAll algorithm which can extract all the association rules between the association rule and the association rule together with the support value and the confidence value. Trend Forecasting System. The method according to claim 1,
The association determining unit may determine,
The relevance analyzing unit analyzes the association rule patterns of multiple time series and the similarity feature pattern weight values extracted from the feature pattern cluster generating unit and receives the similarity feature pattern weights and calculates the degree of support and reliability values of the association rule patterns delayed time- And determining a multiple time series association rule pattern to be used as a prediction model based on the value of the time series correlation rule pattern. The method according to claim 1,
Wherein the trend predicting unit predicts,
Wherein the trend is predicted in relation to multiple time series using an association rule pattern using the predictive model determined by the relevancy determination unit. A method for predicting multiple time series data trends using multiple feature analysis techniques,
Selecting multiple time series data with a time attribute to be predicted.
One time series of the selected multiple time series data is input and the abnormal values are removed and standardized to refine the data by using a DWT (Dscret Wavelet Transfer) technique and a normalization technique, and the change rate, the change scale characteristic value, The step of extracting three kinds of morphological features and defining the characteristics of time series.
(Cluster) a feature-based synthesis (clusters) and a frequency-based weightedness extraction for extracting similar feature patterns among the defined features. And
Checking whether the feature pattern for all the multiple time series data and the similar feature pattern comprehensive cluster extraction are completed.
Analyzing the association rule pattern using the AprioriAll algorithm and the time delay technique that can receive the completed multiple time series data similarity pattern cluster data and analyze all the association rule patterns among the association rule algorithms,
Determining an association rule pattern according to a predictive probability value using the extracted similarity feature pattern inter-similarity feature weight weight extraction data and the association rule pattern analysis step among the rule pattern analysis data received in the association rule pattern analysis step, Generating data as a prediction model,
And estimating a future trend when an association pattern is generated using the generated prediction model.

Images