KR20220109935A - Ansemble device for predicting the power - Google Patents

Abstract

An ensemble device of the present invention may include: a first prediction value generation unit calculating a first prediction value of a solar power generation amount by using a first parameter; a second prediction value generation unit calculating a second prediction value of a solar power generation amount by using a second parameter; and an ensemble unit combining the first prediction value and the second prediction value to calculate an ensemble prediction value, which is a future solar power generation amount.

Description

Ensemble device for predicting the power

The present invention relates to an ensemble device for predicting power generation by combining two different types of power prediction methods.

Recently, as issues such as de-nuclear power generation, fine dust, and abnormal climate have emerged in the power industry, DR (Demand Response), Renewable Energy, ESS (Energy Storage System), EMS (Energy Management System; Energy Management) device), the importance of clean energy and energy efficiency is increasing day by day.

Energy solution-related research such as DR capacity calculation, ESS optimal operation plan, and renewable energy generation prediction is being actively conducted.

The present invention predicts the amount of power generation in two different ways in order to predict the amount of solar power generation, and then assigns a weight dependent on the forgetting rate to each predicted amount to provide an ensemble device for power prediction that calculates an ensemble predicted value that is the amount of future power generation. can

The ensemble device of the present invention includes a first predicted value generator that calculates a first predicted value of the amount of solar power generation using a first parameter, a second predicted value generator that calculates a second predicted value of the amount of solar power that uses the second parameter; The ensemble unit may include an ensemble unit for calculating an ensemble predicted value that is a future solar power generation amount by combining the first predicted value and the second predicted value.

When the first parameter is singular and the second parameter is plural, the first parameter may not be included in the second parameter.

The first prediction value generator selects past or current generation data as a first parameter, decomposes the past or current generation data through transformation, calculates future data using the decomposed data, and performs inverse transformation of the future data It is possible to calculate the first predicted value, which is the amount of future power generation, by restoring it.

The first prediction value generator selects past or present power generation data as a first parameter, and a wavelet unit for decomposing the past or present power generation data through transformation is provided, and for transformation and decomposition, a discrete wavelet transform (DWT, Discrete Wavelet Transform) ) can be used, and the first prediction unit can use an AutorRegressive Moving Average (ARMA) model for future data calculation, and an inverse transform unit that calculates future data using the decomposed data and restores the future data through inverse transformation is provided. , the inverse transform unit may use an inverse discrete wavelet transform (IDWT) during the inverse transform.

The second parameter may be any one of date, temperature, wind speed, cloud amount, humidity, or dew point.

The second predicted value generating unit may display a past or present power generation amount according to the specified second parameter on a graph, calculate an expression indicating a trend of the displayed data, and calculate the second predicted value.

The second prediction value generating unit is provided with a graphing unit displaying the amount of power generation in the past or present according to the specified second parameter, and with respect to the data displayed by the graphing unit, an epsilon tube (ε-tube) corresponding to the noise range is provided. ), a noise setting unit for setting a support vector corresponding to the noise data is provided, and a second prediction unit for calculating the second prediction value using a support vector regression model (SVR) can be provided.

When the data displayed by the graphing unit is non-linear, a mapping unit for allowing the data to have a linear relationship through mapping is provided, and the mapping unit may use a kernel function for the mapping.

The ensemble unit may use a convex combination to optimize the ensemble prediction value.

A weighting unit may be provided in the ensemble unit to give weights to the first prediction value and the second prediction value, and the first prediction value and the second prediction value are set to have different weights contributing to the ensemble prediction value, and the weight is set to a rate of forgetting. of forgetting), and the forgetting rate may give weight to the first predicted value and the second predicted value based on the accuracy between the actual power generation data and the respective predicted values.

, a_n는 상기 제1 예측치의 상기 앙상블 예측치에 대한 기여도를 나타내는 제1 가중치이며, b_n는 상기 제2 예측치의 상기 앙상블 예측치에 대한 기여도를 나타내는 제2 가중치이고, 제1 예측치와 상기 제2 예측치의 예측에 사용된 과거 또는 현재 데이터가 x_n이며, 각 변수의 아래첨자는 앙상블 예측의 각 단계를 의미하고, 제1 가중치 및 상기 제2 가중치는 이전 단계의 제1 예측치, 이전 단계의 제2 예측치, 또는 망각률을 독립 변수로 가지는 종속 변수일 수 있다. When the first prediction value is y _n and the second prediction value is z _n , the ensemble prediction value w _n can be given by the following equation,

, a _n is a first weight indicating the contribution of the first prediction value to the ensemble prediction value, b _n is a second weight indicating the contribution of the second prediction value to the ensemble prediction value, the first prediction value and the second The past or present data used for prediction of the prediction value is x _n , the subscripts of each variable mean each stage of the ensemble prediction, and the first weight and the second weight are the first prediction value of the previous stage, the second weight of the previous stage 2 It may be a predictive value or a dependent variable having the rate of forgetting as an independent variable.

When the past or current data is a value between the first prediction value and the second prediction value, the first weight and the second weight may be given by the following equations,

When the first prediction value and the second prediction value are the same, the first weight and the second weight may be given by the following equation,

When past or present data is greater than or equal to the first prediction value, and the first prediction value is greater than the second prediction value, or when past or present data is less than or equal to the first prediction value, the first prediction value is the When it is smaller than the second predicted value, the first weight and the second weight may be given by the following equation.

When past or present data is greater than or equal to the second prediction value, and the second prediction value is greater than the first prediction value, or when past or present data is less than or equal to the second prediction value, the second prediction value is the When it is smaller than the first predicted value, the first weight and the second weight may be given by the following equation.

The present invention is an ensemble device for estimating the amount of electricity that ensembles the prediction values of two methods of predicting the amount of solar power generation, the first method can use a univariate model, and the past or present solar power generation can be selected as a univariate, , the second method is a multivariate model, and may be exogenous parameters such as temperature, wind speed, and relative humidity.

Univariate models are more economical, perform faster, and can be effective for forecasting over a short period of time, while multivariate models are more cost-effective than exogenous factors such as date, temperature, wind speed, cloud cover, precipitation, humidity, and dew point. It can be effective in predicting

To ensemble the two predictions, we can choose a convex combination for convex optimization, which simplifies the problem to compute the ensemble predictions.

1 is a block diagram showing the structure of an ensemble device of the present invention.
2 illustrates a wavelet portion of the present invention.
3 illustrates a first predicted value calculated by the first predicted value generator of the present invention.
4 illustrates embodiments of a kernel function that may be used to calculate a second prediction value.
5 is a diagram illustrating a Discrete Wavelet Transform (DWT) by the wavelet unit of the present invention.
6 shows actual data mapped by the mapping unit of the present invention.
7 illustrates an embodiment in which an error or noise is included in FIG. 6 .
FIG. 8 illustrates that, when past or present power generation data is non-linear, linear data is generated in a high dimension using a kernel function.

The present invention relates to an wattage prediction device that ensembles prediction values of two methods for estimating the amount of solar power generation.

For relatively short time-interval forecasts, it can be prohibitively expensive to set up and maintain a meteorological station that can collect and store information on solar radiation, solar duration, temperature, wind speed, and relative humidity. Satellite images or numerical weather forecasts may also be limited. In addition, it may be difficult to expect a third party that acts as an intermediate bridge to provide information about the data that can be used to predict the amount of electricity for a reasonably short time interval.

A first method of predicting power may be a univariate model, and a first prediction value P1 for ensemble combination may be generated through the first prediction value P1 generator 100 .

Univariate models can be economical, computationally fast, and realistic for very short time predictions. The univariate model may be a function of a current or past power time series value, not a time series value of other factors such as temperature, wind speed, relative humidity, etc., may use a first parameter R1 as a variable, and may use a first parameter R1 ) may be the past or present power generation (P0).

The second method of predicting power may be a multivariate model, and the second prediction value P2 for the ensemble combination may be generated through the second prediction value P2 generator 200 , and as a variable The second parameter R2 may be used, and the second parameter R2 may include at least one of date, temperature, wind speed, cloud amount, humidity, and dew point. Most multivariate models can require a wide range of input data. The limited use of input data in these multivariate models can lead to low accuracy. In addition, inaccuracies in the input data are propagated to the output, and these errors can accumulate. Univariate models lack interpretation of exogenous inputs, such as temperature, wind speed, relative humidity, etc. may be difficult to incorporate or consider. Thus, this can lead to ensemble combinations of algorithms or prediction methods that are complementary to each other.

Evolution of convex combination parameters occurring over the coarsest time step mesh uses a convex optimizer ensemble to generate a convex combination of predictors. can The evolution of the convex combinatorial parameters with a rate of forgetting can weight the performance of each predictor based on past accuracy. This enables the localized linear combination of the two predictors to be applied to the finer time step mesh for the two predictors with the most recent distinct values in the mesh.

In this way, the ensemble prediction of the present invention can extract additional accuracy from the univariate model at the intermediate gap between time steps in the multivariate model, enabling short interval estimation. Because the underlying two models have complementary advantages, and because multivariate models are particularly robust to the occurrence of changes in external inputs, the convex optimization ensemble approach with oblivion is useful when multivariate models perform poorly. You can have the best of both models by allowing the model to supplement and disappear into the background.

The first prediction value (P1) generating unit 100 includes a wavelet unit 120 that decomposes a series of data for a specific parameter into a desired basic unit block, such as a wavelet, and a first prediction that predicts a future value using the wavelet. It may include a unit 140 and an inverse transform unit 160 that recombines the values decomposed and predicted by the wavelet unit 120 and the first prediction unit 140 .

The predicted value of the univariate model can first be obtained through wavelet decomposition through the discrete wavelet transform (DWT), and the parameters can be predicted as an independent time series using the ARMA model (AutorRegressive Moving Average), and finally the inverse Predictions for the parametric time series can be reassembled using the Inverse Discrete Wavelet Transform (IDWT). The predicted value may be a specific physical quantity for a time series that the user wants to predict, for example, it may mean the amount of power generated in solar power generation. In analyzing specific parameter data according to time series, various methods for a series of processes of decomposition, prediction, and combination may exist, but since there are coding package programs for performing DWT and IDWT, it is easy to perform a series of processes can Numerous structural changes in the model such as decomposition levels and wavelet types by such a package program can be performed through simple parameter manipulation of functions.

A wavelet transform is similar to a Fourier transform or a Windowed Fourier transform, but the merit function can be very different. The merit function may indicate the correspondence between the actual data and the prediction value of the prediction model according to the selection of a specific parameter, and the more the merit function matches each other, the smaller the merit function may be. By adjusting the parameters, one can find the combination that minimizes the merit function.

The wavelet transform may be dense by increasing the temporal resolution for a signal of a high-frequency component corresponding to a high scale, and may be coarse by decreasing the frequency resolution. However, for a signal of a low frequency component corresponding to a low scale, the frequency resolution may be increased to become coarser, and the temporal resolution may be decreased to become coarser.

The Fourier transform can decompose a signal into sines and cosines. In other words, the function can be localized in Fourier space. Conversely, wavelet transforms can use functions localized in both real space and Fourier space. That is, different basis functions may be used, the basis functions of the Fourier transform may be sine and core, and the basis function of the wavelet transform may be a wavelet. The fundamental function of the wavelet may vary according to each frequency domain.

4 may show a representative scaling function having mutual orthogonality and a wavelet family thereof. 4A may be a diagram illustrating a Haar scaling function. 4 (b) and (c) may represent a Debussy (Daubechies) scaling function. The Debussy scaling function is expressed as Daubechies 4, Daubechies 6, and the like, and the numbers 4 and 6 may mean a non-zero number of a _k in Equation 1 above. FIG. 4(b) may show Daubechies 4, and FIG. 4(c) may show Daubechies 20. FIG. It can be seen that the graph becomes smoother as the number of wavelet functions increases.

Wavelet transform can include discrete wavelet transform and continuous wavelet transform (CWT), and compared to CWT, DWT can be simpler because it can impart orthogonality between elementary functions. Therefore, the DWT can be fast enough for the prediction calculation time.

The wavelet transform scales the wavelet (expands and contracts with respect to frequency) and can analyze the signal by function translation (translation with respect to the time axis).

A wavelet can be created from a scaling function that describes the properties of the scale. The scaling function φ may have a constraint that it must be orthogonal to its discrete translation, for example, this constraint may satisfy the dilation equation of Equation 1.

S is a scaling factor, and may be mainly 2, and the function may be normalized. A scaling function may need to be orthogonal to its integer translation. This can be expressed as Equation (2).

That is, the signal is divided into wavelets, and the wavelet may be constructed from a scaling function, and the basic function may be moved and scaled to construct a wavelet. Countless scaling and function shifts may be required for calculations in DWT. A wavelet boundary has upper and lower bounds on the axis, and in order to include all spectra with it, if the remaining empty space of the spectrum becomes small enough, the space can be replaced with another function, a scaling function. That is, the scaling function may be a filter of a low frequency spectrum.

The signal can be expressed as the sum of the scaling function part and the wavelet function. A coefficient multiplied by each scaling function before the scaling function is added may be referred to as a scaling coefficient, and a coefficient multiplied by each wavelet function before the wavelet function is added may be referred to as a wavelet coefficient. This may be equivalent to adding detail, which is the difference between the signal and the coarse approximation and the signal coarse approximation. This may be the same as expressing the signal as the average signal and the difference between the signal and the average. Thus, the approximation may be a low frequency component, and the detail may be a high frequency component. The coefficient may mean a value that has passed through the filters of each band, the scaling coefficient may have passed through a low frequency band, and the wavelet coefficient may have passed through a high frequency band. As a result, the wavelet transform can decompose the signal into a low-frequency component containing the overall content and a high-frequency component containing the details.

Referring to FIG. 2 , according to an embodiment, the y-axis may represent data on the amount of solar power generation according to time, and the x-axis may represent time. When the first wavelet transform is performed on the power generation data (blue), it can be divided into approximate coefficients and detail coefficients. . Since the high-frequency band contains the smallest detail component of interest, the decomposition is stopped, and a second wavelet transform is performed on the approximate coefficients that have passed the low-frequency band, and can be further divided into a low-frequency band and a high-frequency band.

Figure 2 shows this decomposition up to three levels. For example, a three-level decomposition can yield one approximate coefficient result and three detailed coefficient results. The decomposition level may be determined by the user.

Based on the coefficient data obtained by decomposing the wavelet by the wavelet unit 120 , the first prediction unit 140 may generate prediction values for coefficients. For example, the first predictor 140 may use an Autoregressive Moving Average (ARMA) model.

The ARMA model may be a combination of an autocorrelation (AR) model and a moving average (MA) model. The AR model may reflect the influence of a previous value of a specific variable on a subsequent value with respect to a specific variable. That is, the predicted value may be given as the sum of the influence of the previous value and the part affected by the white noise. The MA model reflects the change trend of the average value of a specific variable in the predicted value, and is a method of obtaining the predicted value from the previous error value. Therefore, in the combined ARMA model, the predicted value may include the previous own value, the previous error value, and white noise, and may be expressed as follows. Equation 3 may be an ARMA(1,1) model.

Accordingly, future coefficient prediction values can be obtained from the past coefficients obtained by the wavelet unit 120 using ARMA, which is a statistical technique for predicting the future based on own past data.

The inverse transform unit 160 may perform inverse discrete wavelet transform (IDWT) again on the coefficient predicted value generated by the first predictor 140 to return it to the original time-generation space before decomposition. Therefore, the data restored through the IDWT from the coefficient prediction may be the amount of solar power generation.

Referring to FIG. 3 , it can be seen that the future generation amount (red) is predicted based on the past generation amount (blue color).

The second predicted value (P2) generating unit 200 includes a graphing unit 220 that displays actual data on a graph, and a noise setting unit 240 that generates an epsilon (ε) tube and finds a supporter vector (V) (Support vector). ), and a second prediction unit 260 that generates a second prediction value P2 that is a prediction value from the real data.

The second predicted value P2 generator 200 may generate the second predicted value P2 using the multivariate model. The multivariate model may be a support vector regression model (SVR) based on a support vector machine (SVM) with a radial-based kernel function, and may implement a regression model for various parameters. For example, the regression is It may relate to seven-dimensional metrics for seven parameters, the seven parameters may be days of the year, temperature, wind speed, sky cover, humidity, dew point. The parameters can be normalized to have zero mean and unit variance before applying the SVR technique, and the number of input dimensions can be adjusted according to the usefulness of the set dimension.

The algorithms and key features of SVM can also be used for SVR. That is, SVR can use the basic principles of classification of SVM except for some differences from SVM. First of all, it can be difficult to predict information with infinite possibilities because the output is a number.

An important point of SVR is that it minimizes errors and individualizes a hyperplane that maximizes margins, and some of the errors may be tolerated.

SVM is to solve the classification problem, and for this purpose, a hyperplane can be used to distinguish margins from other types.

SVR can be said to find a line that can explain the data well given the data. The graphing unit 220 may display actual data on a graph, an x-axis may be a parameter as an independent variable, and a y-axis may be a target value, for example, solar power generation power. For example, when the x-axis is selected as the temperature, the graphing unit 220 may display the amount of power generation according to the temperature on the graph, and the goal of the SVR is to find a straight line or curve that can most appropriately represent the relationship between them. can The straight line or curve may be referred to as an SVR regression equation.

An epsilon (ε) tube may be set from the hyperplane, data more distant than the epsilon tube may correspond to noise, and the far data may not be used to obtain an SVR regression equation. The far data may be referred to as a support vector (V). That is, the most appropriate SVR regression equation can be derived only from the data in the epsilon tube excluding the supporter vector (V). In the drawing, one independent variable and one dependent variable are expressed in two dimensions, but the independent variable as a parameter may include, for example, date, wind speed, cloud amount, humidity, dew point, etc., so the hyperplane and supporter vector (V) are It can be expressed at a higher level.

로 주어질 수 있다.

은 벡터 노름(vector norm)이고 다음과 같이 정의될 수 있고, SVM은 2/

을 최대화하는 문제로 귀결될 수 있다. 문제 해결에는 노름의 루트때문에 제곱을 하는것이 더 편하고 같은 결과로 유도되기 때문에, 2/

를 최대화하는 문제와 같을 수 있다. SVM may aim to maximize margin and minimize error. For example, the hyperplane may be given by the equation of a straight line, and may be expressed as y=wx+b. The distance between the decision boundaries representing the epsilon tube may depend on the slope of the straight line, w, 2/

can be given as

is the vector norm and can be defined as: SVM is 2/

can lead to the problem of maximizing Because it is more convenient to square the root of the norm for solving the problem and leads to the same result, 2/

It can be the same as the problem of maximizing .

/2 를 최소화하는 문제로 바뀔 수 있고,

/2 는 회귀식의 평편도를 나타내는 것일 수 있다. 따라서, SVR의 목적은 회귀 계수의 크기를 작게하여 회귀식을 평평하게 만들지만 실제 데이터와 예측치의 차이를 되도록 작게 하는 것일 수 있다. In SVR, since the purpose is different from that of SVM, the formula and form of SVM are the same but can be used in different meanings. ), it is necessary to flatten the regression coefficient w by lowering it. Therefore, in SVR

can be turned into a problem of minimizing /2,

/2 may indicate the flatness of the regression equation. Therefore, the purpose of the SVR may be to make the regression equation flat by reducing the size of the regression coefficient, but to make the difference between the actual data and the predicted value as small as possible.

The epsilon tube may be set to a value specified by a user, which may mean a desired level of noise. The xi (ξ) means a distance out of the epsilon tube, and a penalty may be given to an actual value outside the tube by a multiplier of the weight. That is, the SVR assumes that there is noise in the data, and may allow a difference between the actual data and the predicted value within the epsilon tube (2ε).

)에 대한 제한식이고, 후부의 제한 조건은 튜브 아랫쪽의 가장 가까운 서포터 벡터(V)까지의 거리(

)에 대한 제한식일 수 있다. A constraint condition may occur between the regression equation and the epsilon tube, and the upper constraint condition is the distance to the nearest supporter vector (V) above the tube (

), and the rear limiting condition is the distance to the nearest supporter vector (V) at the bottom of the tube (

) can be a limiting expression for

Therefore, the problem of SVR can be concluded as a problem of solving Equation 5 under the constraint condition of Equation 6.

By solving the above equation using a Lagrangian multiplier or the like, b can be obtained, thereby obtaining a desired SVR regression equation.

Accordingly, the second predicted value P2 generator 200 may display the actual data on the graph in the graphing unit 220 and set the epsilon tube in the noise setting unit to set the supporter vector V, and the second The prediction unit 260 may obtain an SVR regression equation corresponding to the predicted value of the power generation, so that the second predicted value P2 generator 200 may generate a second predicted value P2 that is a second predicted value.

The above case may be a case in which the SVR regression equation is given linearly, and the SVR regression equation that is the actual data and its estimated prediction value may be given non-linearly. In this case, the mapping unit 230 may use a kernel function to convert nonlinear data into linear data. When the actual data is given as a curve, the kernel function can obtain a linear expression by mapping the curve into a feature space. Therefore, the kernel SVR can describe the data well by mapping the data in the original space to a higher dimension using the mapping function (φ). The kernel function K may be defined as an inner product of the mapping function φ. Even if you do not know the mapping function exactly in the form of the expression, the same effect can be obtained internally between the mapping functions, so it has the advantage of being able to solve it without knowing exactly the mapping function itself.

Accordingly, when the actual data is given non-linearly in the time-generation graph, the mapping unit may map the data in a high dimension to be linear.

6 may be for a linear SVR, and FIG. 7 may be for a non-linear SVR including an error.

)는 라그랑지안 멀티플라이어(Lagrangian multiplier)일 수 있고, 파이(

)는 맵핑 함수이며, K는 커널 함수일 수 있다. Equation 7 may be an equation for a linear SVR, and Equation 8 may be an equation for a non-linear SVR. Alpha(

) may be a Lagrangian multiplier, and pi (

) may be a mapping function, and K may be a kernel function.

Since the kernel function may vary, and the shape space characteristics may vary depending on the selected kernel, the kernel function may be selected according to the characteristics of the data. The kernel function may include, for example, a Radial Basis Function (RBF) kernel, a sigmoid kernel, and a polynomial kernel.

The ensemble unit 300 combines the first predicted value P1 and the second predicted value P2, which are the prediction values of the first predicted value P1 generator 100 and the second predicted value P2 generator 200 , to obtain a final predicted value. can be calculated.

When combining the first prediction value P1 and the second prediction value P2 , the ensemble unit 300 may perform convex optimization using a convex combination.

A convex combination may be a linear combination in which all coefficients of the parameters are non-negative and sum to one. Geometrically, all parts of a line connecting two points in a set may belong to itself. That is, it may be a concave part or an assembly without a hole therein.

First, the optimization problem may mean finding a parameter combination that optimizes (maximizes or minimizes) a function value of an objective function. This could be a problem of finding a maximum or a minimum, and both could be the same problem by taking the minus of the objective function.

The optimization problem of finding the minimum value can be like finding the direction in which the value of a function decreases from the current position and going down, and it can be repeated until it reaches a point where it can no longer go down. In this case, there may be various methods depending on which direction to descend and how much to descend at a time.

Among the optimization methods, the problem can be solved much simpler if it results in a problem of convex optimization. This is because the local minimum of the convex optimization becomes the global minimum. In addition, if the optimization problem results in convex optimization, it can be easily solved without performing complex calculations directly because there is a package called CVX (CVXPY), which is a convex optimization solver.

Accordingly, the ensemble unit 300 can solve the convex optimization problem by convexly combining the two prediction values.

One of the two prediction values may be calculated by the first prediction value P1 generator 100 , and the other may be calculated by the second prediction value P2 generator 200 .

The ensemble prediction value P3 to be finally predicted by the ensemble unit 300 may be referred to as w _n , the first prediction value P1 may be referred to as y _n , and the second prediction value P2 may be referred to as z _n , and wn is It can be given as the sum of yn and zn. That is, w _n = a _n y _n + b _n z _n may be.

The weighting unit 320 may set a weight to the coefficient of each prediction value depending on the forgetting rate. As described below, the weight for each prediction can have the forgetting rate (q) as an independent variable, and when the actual data (xn) is closer to any one of the predictions, the predictions other than the predictions close to the actual data are more You can give it less weight. In this case, the degree of each weight can be adjusted by the forgetting rate. For example, if the first predictor gives a closer prediction in a certain interval, the second predictor gives a higher rate of forgetting, and the first predicts gives a lower rate of forgetting, so that the first predictive value in that interval is given as a result. has a higher weight, allowing the second prediction to have a lower weight.

Coefficients a and b may represent weights of respective prediction values. For example, if the coefficient is large, the contribution of each predictor to the ensemble predictive value P3 may increase. If a is greater than b, the first predictive value P1 is greater than the second predictive value P2 It can mean contributing to (P3).

This contribution can be controlled by the forgetting rate.

If y and z are in different time meshes, we can find a coarse time mesh to which the two predictions can be aligned. That is, when a time interval or a time window is used for calculation, it may be selected that both y and z may be included. For example, if two parameters are different in seconds and minutes, you can select the minutes to calculate.

The coefficients a _n and b _n may depend on the predictor variables y and z, the actual data x and the rate of forgetting q. That is, the coefficients a and b can use the predicted values and data values of the previous stage to predict the a and b of the current stage, and it is possible to determine which variables are to be gradually forgotten by the forgetting rate. The forgetting rate q may be a number greater than 0 and less than 1.

The ensemble prediction value P3 may depend on the first prediction value P1 and the second prediction value P2 , and the degree of forgetting of each prediction value may be determined according to the forgetting rate.

The ensemble prediction P3 is determined by the coefficients a and b, and can be specified by giving specific evolution rules. The expansion rule can be given by Equations 10 to 13. The initial values of a and b are both 1/2, so you can start with the same weight.

In order to know whether Equations 10 to 13 have the characteristics of the desired ensemble prediction value P3, looking at each ensemble prediction value P3 and the dominant factor determining the coefficient, Equation 10 is the actual data In the case where x is a value between the predicted values y and z, a2 may be dominated by (x1-z1), b2 may be dominated by -(x1-y1), and a3 may be dominated by (x2-z2) dominant, and b3 may be dominated by -(x2-y2).

Accordingly, when x is a value closer to y than z, (x-y) may become small and (x-z) may become large. That is, a2 may be greater than b2, and a3 may be greater than b3. The fact that x is closer to y than z means that the actual data is closer to the first predicted value P1 than to the second predicted value P2. Therefore, a weight must be added to the first predicted value (P1)(y) when the ensemble is developed, and the second It can be seen that the predicted values P2 (z) can produce an ensemble prediction value P3 that is better to be forgotten, and a is larger than b, so it can be seen that the property is satisfied.

In Equation 11, yn and zn are the same, and all yn and zn may have the same value as the initial value by 1/2. That is, the same yn and zn means that the distance between the first predicted value P1 and the second predicted value P2 and the actual data is the same.

yn > zn or xn

yn < zn 인 경우로서, a2는 1-(q/2)이고, b2는 q/2이며, a3는 1-(

/2)이고, b3는

/2일 수 있다. a2에서 b2를 빼면 1-q이고, a3에서 b3를 빼면 1-

일 수 있다. 또한, q는 0보다 크고 1보다 작으므로, a2는 b2보다 크고, a3는 b3보다 클 수 있다. xn

yn > zn or xn

yn < zn 의 조건은 x가 z보다 y에 가까운 값이므로, 제1 예측치(P1)에 가중치가 주어지고, 제2 예측치(P2)는 점점 잊혀져야 할 수 있다. a가 b보다 큰 값이므로, 성질을 만족함을 알 수 있다. Equation 12 is xn

yn > zn or xn

When yn < zn, a2 is 1-(q/2), b2 is q/2, and a3 is 1-(

/2), and b3 is

It can be /2. a2 minus b2 is 1-q, a3 minus b3 is 1-

can be Also, since q is greater than 0 and less than 1, a2 may be greater than b2 and a3 may be greater than b3. xn

yn > zn or xn

Since the condition of yn < zn is that x is closer to y than z, a weight is given to the first predicted value P1 and the second predicted value P2 may have to be gradually forgotten. Since a is a larger value than b, it can be seen that the property is satisfied.

zn > yn or xn

zn < yn 인 경우로서, b2는 1-(q/2)이고, a2는 q/2이며, b3는 1-(

/2)이고, a3는

/2일 수 있다. b2에서 a2를 빼면 1-q이고, b3에서 a3를 빼면 1-

일 수 있다. 또한, q는 0보다 크고 1보다 작으므로, b2는 a2보다 크고, b3는 a3보다 클 수 있다. xn

zn > yn or xn

zn < yn의 조건은 x가 y보다 z에 가까운 값이므로, 제2 예측치(P2)에 가중치가 주어지고, 제1 예측치(P1)는 점점 잊혀져야 할 수 있다. b가 a보다 큰 값이므로, 성질을 만족함을 알 수 있다. Equation 13 is xn

zn > yn or xn

When zn < yn, b2 is 1-(q/2), a2 is q/2, and b3 is 1-(

/2), and a3 is

It can be /2. If b2 minus a2 is 1-q, b3 minus a3 is 1-q.

can be Also, since q is greater than 0 and less than 1, b2 may be greater than a2 and b3 may be greater than a3. xn

zn > yn or xn

Since the condition of zn < yn is that x is closer to z than y, a weight is given to the second prediction value P2, and the first prediction value P1 may have to be gradually forgotten. Since b is a larger value than a, it can be seen that the property is satisfied.

When the ensemble predicted value P3 is obtained in the above manner, the residual corrector 400 may perform a residual analysis to discover a nonlinear pattern from an error. Extracting meaningful nonlinear patterns from errors can compensate for wavelet ARMA limitations of univariate models and advance prediction algorithms. In this case, the residual corrector 400 may use NARX (Nonlinear AutoRegressive with eXternal input) for this purpose.

In the case of using NARX, the residual analysis may be performed on the ensemble predictive value P3 in which the two predictive values are combined, rather than performing the analysis on each predictive value. NARX is a recursive dynamic neural network and may be more suitable than other recursive structures in predicting time series-based data. The NARX error model can be expressed as a nonlinear function for the past error as a feedback object and a time series change of power as a non-feedback input for every hour.

100... First predicted value generator 120... Wavelet unit
140... First prediction unit 160... Inverse transform unit
200... Second prediction value generating unit 220... Graphing unit
230... mapping unit 240... noise setting unit
260... Second prediction part 300... Ensemble part
320... Weight part 400... Residual correction part
P0... Current or past power generation P1... First forecast
P2... Second Prediction P3... Ensemble Prediction
V... supporter vector R1... first parameter
R2... second parameter

Claims (18)

a first predicted value generator for calculating a first predicted value of the amount of solar power generation by using the first parameter;
a second predicted value generator for calculating a second predicted value of the amount of solar power generation by using the second parameter;
an ensemble unit calculating an ensemble predicted value that is a future solar power generation amount by combining the first predicted value and the second predicted value; An ensemble device comprising a.
The method of claim 1,
When the first parameter is singular and the second parameter is plural,
The first parameter is not included in the second parameter.
The method of claim 1,
The first predicted value generator,
Selecting past or current power generation data as the first parameter,
An ensemble device for decomposing the past or present power generation data through transformation, calculating future data using the decomposed data, and calculating a first predicted value that is a future power generation amount by restoring the future data through inverse transformation.
The method of claim 1,
The first predicted value generator,
Selecting past or current power generation data as the first parameter,
A wavelet unit for decomposing the past or present power generation data through transformation is provided,
The wavelet unit is an ensemble device using a Discrete Wavelet Transform (DWT) that decomposes the past or present power generation data into wavelets.
The method of claim 1,
The first predicted value generator,
Selecting past or current power generation data as the first parameter,
A first prediction unit for decomposing the past or present power generation data through transformation and calculating future data using the decomposed data is provided,
The first prediction unit is an ensemble device using an AutorRegressive Moving Average (ARMA) model that calculates the future data, which is a dependent variable, using a first parameter of a previous step, an error of a previous step, and white noise as independent variables.
The method of claim 1,
The first predicted value generator,
Selecting past or current power generation data as the first parameter,
The past or present power generation data is decomposed through transformation, and future data is calculated using the decomposed data,
An inverse transform unit for restoring data through inverse transform is provided,
The inverse transform unit is an ensemble that calculates the first predicted value from the past or present power generation data and the future data through an inverse transform corresponding to an inverse of a transform process that the wavelet unit decomposes using wavelets. Device.
The method of claim 1,
The second parameter is an ensemble device including at least one of date, temperature, wind speed, cloud amount, humidity, and dew point.
The method of claim 1,
The second predicted value generator,
An ensemble device that displays a past or present power generation amount according to the specified second parameter on a graph, calculates an equation representing a trend of the displayed data, and calculates the second predicted value.
The method of claim 1,
In the second prediction value generator,
A graphing unit for displaying the amount of power generation in the past or present according to the specified second parameter is provided,
A noise setting unit is provided for setting an epsilon tube (ε-tube) corresponding to the noise range with respect to the data displayed by the graphing unit to set a support vector corresponding to the noise data,
and a second prediction unit for calculating the second prediction value using a support vector regression model (SVR) for deriving a regression equation representing the trend of the past or present generation amount.
10. The method of claim 9,
When the data displayed by the graphing unit is non-linear,
A mapping unit that allows data to have a linear relationship through mapping is provided,
The mapping unit uses a kernel function that converts a nonlinear data relationship into a linear data relationship in a high dimension for the mapping.
The method of claim 1,
The ensemble prediction value is a convex combination in which a sum of coefficients is 1 and each coefficient has a value greater than 0 for convex optimization of the ensemble prediction value.
The method of claim 1,
The ensemble unit is provided with a weighting unit for giving weights to the first predicted value and the second predicted value.
The method of claim 1,
The first prediction value and the second prediction value are set to have different weights contributing to the ensemble prediction value,
The weight is determined by the rate of forgetting,
The forgetting rate gives a first weight and a second weight to the first prediction value and the second prediction value, respectively, based on the accuracy between the actual power generation data and the respective prediction values,
Each weight has the forgetting rate as an independent variable,
When the first predicted value is a value closer to the actual power generation data than the second predicted value, the first weight having a forgetting rate lower than the forgetting rate of the second weight is an ensemble having a larger value than the second weight Device.
The method of claim 1,
When the first prediction value is y _n and the second prediction value is z _n ,
The ensemble prediction value (w _n ) is given by the following equation,

wherein a _n is a first weight indicating the contribution of the first prediction value to the ensemble prediction value,
b _n is a second weight indicating the contribution of the second prediction value to the ensemble prediction value,
Past or present data used for prediction of the first prediction value and the second prediction value is x _n ,
The weight is determined by the forgetting rate (q),
The subscripts of each variable refer to each step of the prediction of the ensemble prediction,
The first weight and the second weight are dependent variables having a first prediction value of a previous stage, a second prediction value of a previous stage, or a forgetting rate as an independent variable.
15. The method of claim 14,
When the past or present data is a value between the first prediction value and the second prediction value,
The first weight and the second weight are given by the following equation.

15. The method of claim 14,
When the first prediction value and the second prediction value are the same,
The first weight and the second weight are given by the following equation.

15. The method of claim 14,
When the past or present data is greater than or equal to the first prediction value, and the first prediction value is greater than the second prediction value; or
When the past or present data is less than or equal to the first prediction value, and the first prediction value is less than the second prediction value,
The first weight and the second weight are given by the following equation.

15. The method of claim 14,
When the past or present data is greater than or equal to the second prediction value, and the second prediction value is greater than the first prediction value; or
When the past or present data is less than or equal to the second prediction value and the second prediction value is less than the first prediction value,
The first weight and the second weight are given by the following equation.

Images