KR102333271B1 - A method for selecting the number of hidden units using correlation of hidden units in an artificial neural network model and a method for predicting hydrological climate variables using the hidden unit - Google Patents

Abstract

The present invention relates to a method for selecting a number of concealment unit using the mutual information of the concealment unit in an artificial neural network model and a method for predicting a hydrological climate variable using the same. More specifically, the present invention relates to the method for selecting the number of concealment unit using the mutual information of the concealment unit in the artificial neural network model that selects the number of the concealment unit by using the mutual information (MI) of the concealment unit in the artificial neural network model, and allows the climate factor time series data to be analyzed to output the hydrological climate variable by using the artificial neural network model having the selected number of concealment unit; and the method for predicting the hydrological climate variable using the same. Therefore, the present invention is capable of having an effect wherein a calculation speed of the input data inputted to the artificial neural network model is increased thereby allowing the analysis result to be derived more quickly.

Description

{A method for selecting the number of hidden units using correlation of hidden units in an artificial neural network model and a method for predicting hydrological climate variables using the hidden unit}

The present invention relates to a method for selecting the number of hidden units using the correlation of hidden units in an artificial neural network model and a method for predicting hydroclimatic variables using the same. More specifically, the correlation of hidden units in an artificial neural network model (MI, Mutual Information) The number of hidden units is selected, and climate factor time series data is analyzed using an artificial neural network model having the selected number of hidden units to output hydrological climate variables. It relates to a selection method and a hydroclimatic variable prediction method using the same.

In the past, knowledge was mainly obtained from experts or literature, but as the scope of the problem became more complex and expanded, it became difficult to acquire knowledge by the traditional methods of the past. As an alternative to this and the progress of the 4th industrial revolution, methods for extracting knowledge by analyzing data to find patterns and rules are being proposed.

In other words, as a technique for extracting knowledge from data, traditionally, beyond statistical techniques, artificial intelligence (AI) techniques such as artificial neural networks, decision trees, genetic algorithms, case inference systems, and fuzzy systems are being used. .

In particular, artificial neural networks are used in various problem areas to solve classification and prediction problems. For example, artificial neural networks can predict results by analyzing and utilizing financial-related data such as insolvent company prediction models and bond rating evaluations. In addition, it can be used in water resource management, hydrological design, and many climate-related applications by evaluating and analyzing climate change from the past to the present and its impact.

However, the artificial neural network is not sensitive to the noise of the data and has a strong structure, but it is difficult for users to understand the results because the internal process of data learning is generated by a complex mathematical model, and It is pointed out that the biggest problem is that it takes a long time to analyze the data.

Related Document 1 relates to the operation method of a deep learning-based climate change prediction system. It has the advantage of being able to predict accurate climate change by detecting whether there is missing weather data, but it decomposes a lot of data without considering the number of hidden layers and Since the decomposed data is analyzed, there is a disadvantage that it is difficult to quickly analyze it.

KR 10-2020-0052806

The present invention is to solve the above problems, and the correlation between the number of hidden units and the hidden units can be omitted so that the process of dividing data in order to calculate the number of conventional hidden units and calculating the number of hidden units from other data can be omitted. (MI, Mutual Information) to select the most appropriate number of hidden units of the artificial neural network model to provide a method for selecting the number of hidden units using the correlation of hidden units in the artificial neural network model and a hydroclimatic variable prediction method using the same aim to

In addition, an object of the present invention is to solve the disadvantage of slowing down due to the complicated structure of the artificial neural network, and to increase the operation speed of the input data input to the artificial neural network model so that the analysis result can be derived more quickly. To provide a method for selecting the number of hidden units using the correlation of hidden units in an artificial neural network model to have the most appropriate number of hidden units and a method for predicting hydroclimatic variables using the same.

In addition, it is an object of the present invention to obtain climate factor (eg, PDO) time series data using an artificial neural network model having the number of hidden units selected from the number of hidden units selection step so that hydrologically meaningful prediction data can be quickly output. An object of the present invention is to provide a method for selecting the number of hidden units using the correlation of hidden units in an artificial neural network model that analyzes and outputs hydrological climate variables, and a method for predicting hydrological climate variables using the same.

In order to achieve the above object, a method for selecting the number of hidden units using the correlation of hidden units in an artificial neural network model of the present invention comprises: a data input step of inputting time series data into an artificial neural network model; an MI estimation step of estimating a correlation (MI, Mutual Information) of hidden units according to the number of hidden units in the time series data; an MI average value calculating step of calculating an average value of the correlation (MI, Mutual Information) of the hidden units for each number of the hidden units; A polynomial regression analysis step of deriving a second polynomial in which the independent variable is the number of hidden units and the dependent variable is the average value of the correlation (MI, Mutual Information) of the hidden units using polynomial regression; and 0 is substituted for the average value of the correlation (MI, Mutual Information) of the hidden units so that the second-order polynomial is first differentiated with respect to the number of hidden units, and the correlation between the hidden units is removed. It provides; a hidden unit number selection step in which the appropriate number of hidden units is selected.

In the method of selecting the number of hidden units using the correlation of the hidden units in the artificial neural network model of the present invention, the average value of the correlation (MI, Mutual Information) of the hidden units increases as the number of the hidden units increases. One characteristic is strengthened to decrease the average value of the correlation (MI, Mutual Information) of the hidden unit, and when the minimum point is passed, the average value of the correlation (MI, Mutual Information) of the hidden unit is maintained or increased again.

In the method for selecting the number of hidden units using the correlation of the hidden units in the artificial neural network model of the present invention, the polynomial regression analysis step is characterized in that a quadratic polynomial is calculated by the following [Equation 1].

[Equation 1]

,

,

는 상수이다.Here, y is the average value of the correlation (MI, Mutual Information) of the hidden units, x is the number of the hidden units,

,

,

is a constant.

In the method of selecting the number of hidden units using the correlation of hidden units in the artificial neural network model of the present invention, the step of selecting the number of hidden units is that the number (x) of the hidden units is rounded to one decimal place and derived as a positive integer. characterized in that

Next, in order to achieve the above object, the method for predicting hydrological climate variables using the method for selecting the number of hidden units of the present invention comprises: a data input step of inputting climate factor time series data into an artificial neural network model; an MI estimation step of estimating a correlation (MI, Mutual Information) of hidden units according to the number of hidden units in the climate factor time series data; an MI average value calculating step of calculating an average value of the correlation (MI, Mutual Information) of the hidden units for each number of the hidden units; A polynomial regression analysis step of deriving a second polynomial in which the independent variable is the number of hidden units and the dependent variable is the average value of the correlation (MI, Mutual Information) of the hidden units using polynomial regression; The second polynomial is first differentiated with respect to the number of hidden units, and 0 is substituted for the average value of the correlation (MI, Mutual Information) of the hidden units so that the correlation between the hidden units is removed, so that the most appropriate of the artificial neural network model a hidden unit number selection step in which the number of hidden units is selected; and a hydrological climate variable output step in which the climate factor time series data is analyzed and hydrological climate variables are output using an artificial neural network model having the number of hidden units selected in the step of selecting the number of hidden units.

In the hydrological climate variable prediction method using the method for selecting the number of hidden units of the present invention, the average value of the correlation (MI, Mutual Information) of the hidden units increases as the number of hidden units increases. It is characterized in that the average value of the correlation (MI, Mutual Information) of the hidden unit is strengthened, and when the minimum point is passed, the average value of the correlation (MI, Mutual Information) of the hidden unit is maintained or increased again.

In the hydroclimatic variable prediction method using the method for selecting the number of hidden units of the present invention, the climate factor time series data is PDO (Pacific Decadal Oscillation) time series data representing the 10-year cycle Pacific oscillation.

In the hydrological climate variable prediction method using the method for selecting the number of hidden units of the present invention, in the polynomial regression analysis step, a quadratic polynomial is calculated by the following [Equation 2]. Methods for predicting climate variables.

[Equation 2]

Here, y is the average value of the correlation (MI, Mutual Information) of the hidden units, and x is the number of the hidden units.

In the hydrological climate variable prediction method using the method for selecting the number of hidden units of the present invention, the step of selecting the number of hidden units is derived as a positive integer by rounding the number (x) of the hidden units to the first decimal place. do.

As described above, according to the present invention, the most appropriate number of hidden units of the artificial neural network model is selected using the correlation between the number of hidden units and the hidden units (Mutual Information), thereby calculating the number of conventional hidden units. There is an effect that the process of dividing the data and calculating the number of hidden units from other data can be omitted.

In addition, the present invention solves the disadvantage of slowing down due to the complicated structure of the artificial neural network by allowing the artificial neural network model to have the most appropriate number of hidden units, and increases the operation speed of input data input to the artificial neural network model. It has the effect of deriving analysis results more quickly.

In addition, the present invention uses an artificial neural network model having the number of hidden units selected from the number of hidden units selection step to analyze climate factor (eg, PDO) time series data to output hydroclimate variables, thereby predicting hydrologically significant It has the effect of printing out data quickly.

1 is a flowchart of a method for selecting the number of hidden units using the correlation of hidden units in an artificial neural network model according to the present invention.
2 is a flowchart of a hydroclimatic variable prediction method using a method for selecting the number of hidden units according to an embodiment of the present invention.
3 is a diagram illustrating PDO time series data according to an embodiment of the present invention.
4 is a view showing the state of each hidden unit of LSTM4 according to an embodiment of the present invention.
5 is a view showing the state of each hidden unit of the LSTM10 according to an embodiment of the present invention.
6 is a diagram illustrating the correlation between each hidden unit of LSTM6 and LSTM14 according to an embodiment of the present invention.
7 is a graph showing [Equation 2] calculated from the polynomial regression analysis step according to an embodiment of the present invention.
8 is a diagram illustrating the correlation between RMSE and hidden units according to the number of hidden units according to an embodiment of the present invention.

The terms used in this specification have been selected as currently widely used general terms as possible while considering the functions in the present invention, which may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technology, and the like. In addition, in a specific case, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the corresponding invention. Therefore, the term used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than the name of a simple term.

Unless defined otherwise, all terms used herein, including technical and scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Terms such as those defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present application. does not

Hereinafter, an embodiment according to the present invention will be described in detail with reference to the accompanying drawings. 1 is a flowchart of a method for selecting the number of hidden units using the correlation of hidden units in an artificial neural network model according to the present invention.

1, the method of selecting the number of hidden units using the correlation of hidden units in the artificial neural network model of the present invention is a data input step (S100), an MI estimation step (S200), an MI average value calculation step (S300), a polynomial regression analysis step (S400), and a step of selecting the number of hidden units (S500).

More specifically, in the data input step ( S100 ), time series data is input to the artificial neural network model.

In general, an artificial neural network includes an input layer, a hidden layer, and an output layer, and when input data is input to the input layer, the hidden layer analyzes the input data and provides output data to the output layer. The hidden layer may include a plurality of hidden units or hidden nodes. This specification is written on the premise that a plurality of hidden units are included in the hidden layer.

In addition, there are various types of artificial neural networks, among which there is a recurrent neural network (RNN) as a model that can consider time history. There is a neural network.

Most preferably, in the method of selecting the number of hidden units using the correlation of hidden units in the artificial neural network model of the present invention, a general artificial neural network, RNN, and LSTM (Long and Short Term Memory)-based recurrent neural networks are used to analyze long-term time series data. Therefore, all of the input data, the hidden unit, and the output data may be time series data.

And, the artificial neural network model may be named LSTM1, LSTM4, or LSTM10 according to the number of hidden units. For example, LSTM1 is an artificial neural network having one hidden unit in the hidden layer, and LSTM4 is an artificial neural network having four hidden units in the hidden layer.

Next, in the MI estimation step (S200), in the time series data, the correlation (MI, Mutual Information) of the hidden units according to the number of the hidden units is estimated.

Meanwhile, in the MI estimation step S200, the correlation (MI, Mutual Information) of the hidden unit is estimated using the bivariate empirical cumulative distribution function. Here, the bivariate empirical cumulative distribution function is a cumulative distribution function related to an empirical scale, and it can be derived as a step-like graph in which statistics are obtained for a probability sample and arranged in order from the smallest value.

Using LSTM4 as an example, there are 4 hidden units in the hidden layer of LSTM4. Each hidden unit can be expressed in a time series form, and the correlation (MI, Mutual Information) of the hidden units with each other can be obtained.

In this case, in the MI estimation step (S200), the correlation (MI, Mutual Information) of the four hidden units with each other can be estimated using a bivariate empirical cumulative distribution function, and from the first MI to the nth MI can be expressed as

Next, in the MI average value calculation step S300, the average value of the correlation (MI, Mutual Information) of the hidden units for each number of the hidden units is calculated.

Using LSTM4 as an example, if the nth MI from the first MI estimated in the MI estimation step S200 is averaged, the average value of the correlation (MI, Mutual Information) of the hidden units for the entire LSTM4 is finally calculated. can

That is, the MI average value calculation step (S300) is a method in which the average value of the correlation (MI, Mutual Information) of the hidden unit for LSTM4 was previously calculated, and the correlation (MI, The average value of mutual information) is calculated.

Here, as for the average value of the correlation (MI, Mutual Information) of the hidden units, as the number of the hidden units increases, the unique characteristic of each hidden unit is strengthened, so that the average value of the correlation (MI, Mutual Information) of the hidden units is It is characterized in that the average value of the correlation (MI, Mutual Information) of the hidden unit is maintained or increased again when the minimum point is passed.

Next, the polynomial regression analysis step (S400) uses polynomial regression, where the independent variable is the number of the hidden units, and the dependent variable is the secondary value of the correlation (MI, Mutual Information) of the hidden units. A polynomial is derived.

In general, polynomial regression is a method for analyzing an independent variable and a dependent variable having a nonlinear relationship, and a quadratic polynomial can be derived by the following [Equation 1].

,

,

상수를 구하면, 해당하는 데이터의 비선형적인 관계를 2차 다항식 그래프로 나타낼 수 있는 것이다.Here, y is the dependent variable and x is the independent variable. And by substituting the corresponding data,

,

,

If the constant is obtained, the nonlinear relationship of the corresponding data can be expressed as a quadratic polynomial graph.

,

,

는 상수이다. 그리고 상기 다항회귀분석단계(S400)는 다수 개의 데이터 중 임의의 (은닉단위의 개수(

), 은닉단위의 상관성의 평균값(

))과 (은닉단위의 개수(

), 은닉단위의 상관성의 평균값(

))를 대입하여 상기

,

,

을 구할 수 있다.That is, in the present invention, y is the average value of the correlation (MI, Mutual Information) of the hidden units, x is the number of the hidden units,

,

,

is a constant. And the polynomial regression analysis step (S400) is any of a plurality of data (the number of hidden units (

), the average value of the correlation of the hidden units (

)) and (the number of hidden units (

), the average value of the correlation of the hidden units (

)) by substituting

,

,

can be obtained

Next, in the step of selecting the number of hidden units ( S500 ), the second-order polynomial is first differentiated with respect to the number of hidden units, and the correlation between the hidden units is removed such that the correlation between the hidden units (MI, Mutual Information) is averaged. 0 is substituted in , and the most appropriate number of hidden units of the artificial neural network model is selected.

That is, in the step of selecting the number of hidden units (S500), when [Equation 1] is derived from the polynomial regression analysis step (S400), the [Equation 1] is It can be first differentiated as shown in [Equation 3] below.

The reason for performing the first differentiation is as mentioned above, as the number of the hidden units increases, the unique characteristics of each hidden unit are strengthened, the average value of the correlation (MI, Mutual Information) of the hidden units is reduced, and the minimum point After passing through, the average value of the correlation (MI, Mutual Information) of the hidden units is maintained or increased again, in order to obtain the minimum point.

And by substituting 0 to the average value of the correlation (MI, Mutual Information) of the hidden unit, which is the independent variable, as shown in Equation 3 below, the most appropriate number of hidden units of the artificial neural network model can be selected.

In this case, in the step of selecting the number of hidden units ( S500 ), the number (x) of the hidden units is rounded off from the first decimal place and is derived as a positive integer. This is because the number (x) of the hidden units in the artificial neural network model must be a positive integer.

Method for selecting the number of hidden units and predicting hydroclimatic variables using the same

2 is a flowchart of a hydroclimatic variable prediction method using a method for selecting the number of hidden units according to an embodiment of the present invention.

Referring to FIG. 2 , the method of predicting hydrological climate variables using the method for selecting the number of hidden units of the present invention includes a data input step (S110), an MI estimation step (S210), an MI mean value calculation step (S310), and a polynomial regression analysis step (S410). , a step of selecting the number of hidden units (S510) and a step of outputting a hydrological climate variable (S610).

More specifically, in the data input step ( S110 ), climate factor time series data is input to the artificial neural network model.

Most preferably, the climate factor time series data is PDO (Pacific Decadal Oscillation) time series data representing a 10-year cycle Pacific oscillation. The PDO time series data is data representing the Pacific climate change pattern, which is the most representative climate index, as a time series.

Also, most preferably, as the artificial neural network model, a Long and Short Term Memory (LSTM)-based recurrent neural network may be used to analyze long-term time series data. Accordingly, the PDO time series data as input data, the hidden unit, and the hydrological climate variable as output data may all be time series data.

And, the artificial neural network model may be named LSTM1, LSTM4, or LSTM10 according to the number of hidden units. For example, LSTM1 is an artificial neural network model having one hidden unit in the hidden layer, and LSTM4 is an artificial neural network model having four hidden units in the hidden layer.

Next, in the MI estimation step (S210), in the climate factor time series data, the correlation (MI, Mutual Information) of the hidden units according to the number of hidden units is estimated.

Meanwhile, in the MI estimation step S210, the correlation (MI, Mutual Information) of the hidden unit is estimated using the bivariate empirical cumulative distribution function. Here, the bivariate empirical cumulative distribution function is a cumulative distribution function related to an empirical scale, and it can be derived as a step-like graph in which statistics are obtained for a probability sample and arranged in order from the smallest value.

Using LSTM4 as an example, there are 4 hidden units in the hidden layer of LSTM4. Each hidden unit can be expressed in a time series form, and the correlation (MI, Mutual Information) of the hidden units with each other can be obtained.

In this case, in the MI estimation step (S210), the correlation (MI, Mutual Information) of the four hidden units with each other can be estimated using a bivariate empirical cumulative distribution function, and from the first MI to the nth MI. can be expressed as

Next, in the MI average value calculation step S310, the average value of the correlation (MI, Mutual Information) of the hidden units for each number of the hidden units is calculated.

For example, using LSTM4, if the nth MI from the first MI estimated in the MI estimation step S210 is averaged, the correlation (MI, Mutual Information) of the hidden units for the LSTM4 having four hidden units. ) can be finally calculated.

That is, the MI average value calculation step (S310) is a method in which the average value of the correlation (MI, Mutual Information) of the hidden unit for LSTM4 was previously calculated, and the correlation (MI, The average value of mutual information) is calculated.

At this time, as for the average value of the correlation (MI, Mutual Information) of the hidden units, the unique characteristic of each hidden unit is strengthened as the number of the hidden units increases, so that the average value of the correlation (MI, Mutual Information) of the hidden units is It is characterized in that the average value of the correlation (MI, Mutual Information) of the hidden unit is maintained or increased again when the minimum point is passed.

A simple experiment is conducted to prove this.

3 is a diagram illustrating PDO time series data according to an embodiment of the present invention. 4 is a view showing the state of each hidden unit of LSTM4 according to an embodiment of the present invention. 5 is a view showing the state of each hidden unit of the LSTM10 according to an embodiment of the present invention.

First, referring to FIG. 3 , in this experiment, PDO time series data is used as input data. The PDO time series data is data representing the Pacific climate change pattern, which is the most representative climate index, as a time series.

4, the PDO time series data is analyzed using an LSTM4 artificial neural network model having four as the number of hidden units, and the state of each hidden unit (black graph) is displayed overlaid on the PDO time series data graph. has been

More specifically, referring to the ht-1 graph of FIG. 3 , the first hidden unit has high-frequency variability. The first hidden unit is close to normal while representing a high frequency. In contrast, since the second to fourth hidden units are sensitive to only one positive or negative phase, it can be seen that the distortion is severe.

When looking at the ht-2 graph of FIG. 3 during periods of high frequency such as 1920-1940 and 1980-2000, the state of the second hidden unit obtains a high positive value, while looking at the ht-4 graph of FIG. 4 The state of the hidden unit gets a high negative value.

Also, the second hidden unit and the third hidden unit show almost the same appearance and are highly dependent on each other, and the first hidden unit and the fourth hidden unit hardly depend on each other.

Next, the LSTM10 model is further analyzed to characterize the structure with a relatively large number of hidden units.

5, the PDO time series data was analyzed using an LSTM10 artificial neural network model having 10 as the number of hidden units, and the state of each hidden unit (black graph) is displayed overlaid on the PDO time series data. have.

The variability in the first period, 1900-1930, is captured in the 9th to 10th hidden units, while the other hidden units consistently hold a value of -1 or 1 during the first period. That is, an important spectrum among the ten hidden units is the ninth to tenth hidden units in the first period.

And it seems that the 3rd hidden unit has a negative relationship with the 4th hidden unit, and has a positive relationship with the 7th hidden unit. The first hidden unit is very sensitive to the peak value of the PDO time series data as shown in ht-1 of FIG. 3, but the eighth unit is inversely sensitive to the negative peak. The sixth hidden unit represents a negative relationship with the PDO time series data in the long term except for the first period.

However, since other units do not depend on each other, it can be seen that the hidden unit in the LSTM10 model has a lower correlation than the hidden unit in the LSTM4 model.

In addition, FIG. 6 is a diagram showing the correlation of each hidden unit of LSTM6 and LSTM14 according to an embodiment of the present invention.

6 is a float matrix, and is a graph in which the hidden units of the LSTM are placed on a diagonal panel and the correlation is displayed. In other words, when a clear graph with less scatter is derived from overlapping panels in one hidden unit and the other, it can be said that the positive/negative correlation between the two hidden units is high.

Fig. 6(a) is a float matrix of LSTM6 having 6 hidden units, and it can be seen that a clear graph is derived with less dispersion and clearly showing positive or negative correlation in all panels. Fig. 5 (b) is a float matrix of LSTM14 having 14 hidden units, and it can be seen that an ambiguous graph with a lot of scatter is derived from all panels.

That is, in the small number of hidden units, each hidden unit has a strong correlation between the hidden units because the PDO time series data used as input data is not sufficiently separated. On the other hand, in a large number of hidden units, each hidden unit has a weak correlation between the hidden units because the PDO time series data used as input data is sufficiently separated.

However, when the average value of the correlation (MI, Mutual Information) of the hidden units exceeds the minimum number of hidden units, the average value of the correlation (MI, Mutual Information) of the hidden units may be maintained or increased again. This is explained in more detail below.

Next, in the polynomial regression analysis step (S410), the independent variable is the number of hidden units using polynomial regression, and the dependent variable is the average value of the correlation (MI, Mutual Information) of the hidden units. A second polynomial is derived.

Most preferably, in the polynomial regression analysis step (S410), when the climate factor time series data is Pacific Decadal Oscillation (PDO) time series data representing the 10-year cycle Pacific oscillation, the second polynomial is calculated by the following [Equation 2] characterized in that

Here, the dependent variable y is the average value of the correlation (MI, Mutual Information) of the hidden units, and the independent variable x is the number of the hidden units.

7 is a graph showing [Equation 2] calculated from the polynomial regression analysis step according to an embodiment of the present invention.

Referring to FIG. 7 , the coordinates of the average value of the correlation (MI, Mutual Information) of the hidden units calculated in the MI average value calculation step S310 are displayed according to the number of each hidden unit. And the graph showing [Equation 2] calculated from the polynomial regression analysis step (S410) is displayed in purple.

7, when the average value of the correlation (MI, Mutual Information) of the hidden units exceeds the minimum number of hidden units, it can be seen that the average value of the correlation (MI, Mutual Information) of the hidden units is increased again.

That is, when the number of hidden units increases by more than a certain amount, the average value of the correlation (MI, Mutual Information) of the hidden units increases again, which increases the analysis load, thereby slowing the speed. And it is very important to select an appropriate number of hidden units because the parameter increases exponentially according to the number of hidden units.

Next, in the step of selecting the number of hidden units ( S510 ), the second-order polynomial is first differentiated with respect to the number of hidden units, and the correlation between the hidden units is removed such that the correlation between the hidden units (MI, Mutual Information) is averaged. 0 is substituted in , and the most appropriate number of hidden units of the artificial neural network model is selected.

In addition, in the step of selecting the number of hidden units (S510), if the second polynomial is calculated by the following [Equation 2] from the polynomial regression analysis step (S410), the [Equation 2] is the number of hidden units (x) First differentiated

The reason for performing the first differentiation is as mentioned above, as the number of the hidden units increases, the unique characteristics of each hidden unit are strengthened, the correlation (MI, Mutual Information) of the hidden units is reduced, and when the minimum point is passed, Since the correlation (MI, Mutual Information) of the hidden unit is maintained or increased again, this is to obtain the minimum point.

And in the step of selecting the number of hidden units (S510), when 0 is substituted for the average value of the correlation (MI, Mutual Information) of the hidden units so that the correlation between the hidden units is removed, the number of hidden units of the artificial neural network model (x) 14.04 is yielded.

At this time, since the number (x) of the hidden unit should be derived as a positive integer, it is rounded up to 14 from the first decimal place.

That is, in the step of selecting the number of hidden units (S510), by differentiating the [Equation 2] derived from the polynomial regression analysis step (S410), the correlation (MI, Mutual Information) of the hidden units is the minimum hidden unit. It is easy to find the number of

Next, in the hydrological climate variable output step (S610), the PDO time series data is analyzed using an artificial neural network model having the number of hidden units selected in the number of hidden units selection step (S500), and the hydrological climate variable is output. .

That is, the PDO time series data is a value extracted from the past, and when it is input to the artificial neural network model, the hydrological climate variable output step S610 outputs meaningful hydrological data that may be generated in the future.

conventional comparison

8 is a diagram illustrating the correlation between RMSE and hidden units according to the number of hidden units according to an embodiment of the present invention.

In order to prove the accuracy and reliability of the present invention, the RMSE value according to the number of hidden units generally used and the average value of the correlation (MI, Mutual Information) of the hidden units according to the number of hidden units of the present invention are compared.

Referring to FIG. 8 , the blue graph is the RMSE value according to the number of commonly used hidden units, and the red graph is the average value of the correlation (MI, Mutual Information) of the hidden units according to the number of hidden units of the present invention.

The root mean square error (RMSE) may be derived through k-fold cross-validation. Here, in k-fold cross-validation, k is a positive integer, and the entire data is divided into sets of similar size as many as k.

That is, when k = 5, after dividing the entire data into 5 subsets, one fold for each division is used for testing and the remaining 4 folds are used for training. Repeat this process for each division. to measure accuracy.

Most preferably, FIG. 8 shows root mean square error (RMSE) derived according to the number of hidden units when K=10.

That is, referring to FIG. 8 , the average value of the correlation (MI, Mutual Information) of the hidden unit shows a significant decrease until similar to the RMSE value of LSTM5. As such, even if the two graphs do not perfectly match, it can be seen that the average value of the correlation (MI, Mutual Information) between the RMSE and the hidden unit operates similarly as the number of hidden units increases. Accordingly, the correlation (MI, Mutual Information) of the hidden unit can replace the conventionally used RMSE value.

As described above in detail a specific part of the present invention, for those of ordinary skill in the art, this specific description is only a preferred embodiment, and it is clear that the scope of the present invention is not limited thereto.

Accordingly, the substantial scope of the present invention will be defined by the appended claims and their equivalents.

Claims (9)

A data input step of inputting time series data to the artificial neural network model;
an MI estimation step of estimating a correlation (MI, Mutual Information) of hidden units according to the number of hidden units in the time series data;
an MI average value calculating step of calculating an average value of the correlation (MI, Mutual Information) of the hidden units for each number of the hidden units;
A polynomial regression analysis step of deriving a second-order polynomial in which the independent variable is the number of hidden units and the dependent variable is the average value of the correlation (MI, Mutual Information) of the hidden units using polynomial regression; and
The second polynomial is first differentiated with respect to the number of hidden units, and 0 is substituted for the average value of the correlation (MI, Mutual Information) of the hidden units so that the correlation between the hidden units is removed, so that the most appropriate of the artificial neural network model A method of selecting the number of hidden units using the correlation of hidden units in an artificial neural network model, including; The method of claim 1,
The average value of the correlation (MI, Mutual Information) of the hidden unit is,
As the number of the hidden units increases, the unique characteristics of each hidden unit are strengthened, so that the average value of the correlation (MI, Mutual Information) of the hidden units is reduced. ) is maintained or increased again. A method for selecting the number of hidden units using the correlation of hidden units in an artificial neural network model. The method of claim 1,
The polynomial regression analysis step is,
A method for selecting the number of hidden units using the correlation of hidden units in an artificial neural network model, characterized in that a quadratic polynomial is calculated by the following [Equation 1].
[Equation 1]

Here, y is the average value of the correlation (MI, Mutual Information) of the hidden units, x is the number of the hidden units,

,

,

is a constant. The method of claim 1,
The step of selecting the number of hidden units is
A method for selecting the number of hidden units using correlation of hidden units in an artificial neural network model, characterized in that the number (x) of the hidden units is rounded off from the first decimal place and derived as a positive integer. A data input step of inputting climate factor time series data into the artificial neural network model;
an MI estimation step of estimating a correlation (MI, Mutual Information) of hidden units according to the number of hidden units in the climate factor time series data;
an MI average value calculating step of calculating an average value of the correlation (MI, Mutual Information) of the hidden units for each number of the hidden units;
A polynomial regression analysis step of deriving a second-order polynomial in which the independent variable is the number of hidden units and the dependent variable is the average value of the correlation (MI, Mutual Information) of the hidden units using polynomial regression;
The second polynomial is first differentiated with respect to the number of hidden units, and 0 is substituted for the average value of the correlation (MI, Mutual Information) of the hidden units so that the correlation between the hidden units is removed, so that the most appropriate of the artificial neural network model a hidden unit number selection step in which the number of hidden units is selected; and
A method of selecting the number of hidden units including a; A method of predicting hydroclimate variables using 6. The method of claim 5,
The average value of the correlation (MI, Mutual Information) of the hidden unit is,
As the number of the hidden units increases, the unique characteristics of each hidden unit are strengthened, so that the average value of the correlation (MI, Mutual Information) of the hidden units is reduced. ) is maintained or increased again. 6. The method of claim 5,
The climate factor time series data is a hydrological climate variable prediction method using a method for selecting the number of hidden units, characterized in that the PDO (Pacific Decadal Oscillation) time series data representing the 10-year Pacific oscillation. 8. The method of claim 7,
The polynomial regression analysis step is,
A hydrological climate variable prediction method using a method for selecting the number of hidden units, characterized in that a quadratic polynomial is calculated by the following [Equation 2].
[Equation 2]

Here, y is the average value of the correlation (MI, Mutual Information) of the hidden units, and x is the number of the hidden units. 6. The method of claim 5,
The step of selecting the number of hidden units is
A method for predicting hydrological climate variables using a method for selecting the number of hidden units, characterized in that the number (x) of the hidden units is rounded off from the first decimal place and derived as a positive integer.

Images