KR20210156654A - Stacking Ensemble Type Short-term Power Demand Prediction Method and Apparatus - Google Patents

Abstract

According to the present invention, a stacking ensemble type short-term power demand predicting device comprises: a data collection unit for collecting date/time information, weather information, and power usage data of a predetermined period; a data pre-processing unit for extracting features of the collected data and performing pre-processing including maximum or minimum normalization; and a predictive modeling unit that applies preprocessed input data to unit prediction models having a different number of layers, which is two or more, based on a deep neural network and calculates a final prediction value from prediction values derived from each of the unit prediction models.

Description

Stacking Ensemble Type Short-term Power Demand Prediction Method and Apparatus

The present invention relates to a short-term power demand prediction method and apparatus to which a stacking ensemble technique is applied, and to a short-term power demand prediction method and apparatus for accurately predicting short-term or ultra-short-term power demand of less than a day by using relatively low-spec hardware. .

Electric Load Forecasting is the basis for securing necessary power in advance for efficient operation of the power system. used for

Electricity demand forecasting is largely divided into short-term, medium-term, and long-term forecasts according to the forecast period. Among them, Short-Term Load Forecasting (STLF) includes daily power demand forecasting, daily peak power demand forecasting, and ultra-short-term power demand forecasting that predicts units of less than an hour per day.

The smart grid establishes an operation plan based on short-term power demand forecasts related to CCHP (Combined Cooling, Heat, and Power), ESS (Energy Storage System), and new and renewable energy. Accurate short-term electricity demand forecasting is essential to build a smart grid system aimed at optimizing efficiency.

Existing power demand forecasting methods used statistical approaches, machine learning, and deep learning algorithms to construct predictive models, and recent deep learning can effectively learn nonlinear external factors, It shows better prediction performance than approaches and machine learning.

However, deep learning algorithms require a lot of time because they need to consider several hyper-parameters that determine the performance of the model, and have disadvantages in that they require high-performance computer specifications to retrain the model in real time.

In addition, the existing prediction method performed a method of predicting a point one day later from the present time using point prediction, but this has a problem in that it cannot prepare for uncertainty, and there are few research cases on ultra-short-term power demand prediction in units of 15 minutes. do.

Republic of Korea Publication No. 10-2013-0035577

Lab 11 - PCR and PLS Regression in Python, March 23, 2016

An object of the present invention is to provide a short-term power demand prediction apparatus and method capable of accurately predicting an ultra-short-term power demand while having a low calculation cost.

A stacking ensemble type short-term power demand prediction apparatus according to an aspect of the present invention includes: a data collecting unit for collecting date/time information, weather information, and power usage data of a predetermined period; a data pre-processing unit performing pre-processing including feature extraction of the collected data and maximum/minimum normalization; and a predictive modeling unit that applies the pre-processed input data to unit prediction models having two or more different number of layers based on a deep neural network, and calculates a final prediction value from the prediction values derived by each of the unit prediction models. .

Here, the predictive modeling unit may include: a basic unit prediction model learning unit configured to learn a basic unit prediction model having a minimum hidden layer; a multi-prediction model constructing unit that derives one or more additional unit prediction models with an increased hidden layer from the basic unit prediction model that has been trained, and generates a multi-prediction model composed of the basic unit prediction model and the additional unit prediction models; and an ensemble prediction unit that calculates a final prediction value by performing stacking ensemble processing on individual prediction values of the basic unit prediction model and the additional unit prediction models.

Here, the basic unit prediction model and the additional unit prediction models may be a deep neural network model.

Here, the ensemble prediction unit may perform principal component regression analysis, which extracts principal components for each individual prediction values of the additional unit prediction models, and then creates a regression model using the principal components based on a sliding window. .

Here, the ensemble prediction unit may perform stacking ensemble processing by applying a sliding window so that the weights for the basic unit prediction model and the additional unit prediction models are corrected to the values learned continuously during the power demand prediction operation.

Here, the data preprocessor may augment the one-dimensional time information into a two-dimensional rotational coordinate system, but the day and holiday information may be expressed as 1 on a day corresponding to a public holiday, and 0 otherwise.

Here, the data preprocessor may apply a Pearson correlation coefficient to determine a correlation between the power usage at the prediction time and the power usage at the past time at the prediction time.

Here, it may further include a storage unit in which the collected data or pre-processed data and learning result values are stored.

Here, the weights for the basic unit prediction model and the additional unit prediction models may be corrected to values continuously learned during the operation of predicting power demand for the target building.

A stacking ensemble type short-term power demand forecasting method according to another aspect of the present invention comprises the steps of: collecting weather information and power usage data for a predetermined period; performing preprocessing including feature extraction of the collected data and maximum minimum normalization; performing learning on a basic unit prediction model having a minimum hidden layer; deriving one or more additional unit prediction models with an increased hidden layer from the basic unit prediction model for which learning is completed, and constructing a multi-prediction model composed of the basic unit prediction model and the additional unit prediction models; The method may include calculating a final predicted value by performing stacking ensemble processing on individual prediction values of the basic unit prediction model and the additional unit prediction models.

Here, the basic unit prediction model and the additional unit prediction models may be a deep neural network model.

Here, in the step of calculating the final predicted value, principal component regression analysis is performed to create a regression model using the principal components after extracting principal components from each of the individual prediction values of the additional unit prediction models based on a sliding window. can be done

Here, in the step of calculating the final predicted value, a sliding window may be applied so that the weights for the basic unit prediction model and the additional unit prediction models are corrected to the values learned continuously during the power demand prediction operation.

Here, in the step of performing the pre-processing, one-dimensional time information is augmented with a two-dimensional rotational coordinate system, but Sunday and public holidays information may be expressed as 1 on a day corresponding to a public holiday, and 0 otherwise.

Here, in the step of performing the preprocessing, a Pearson correlation coefficient may be applied to determine a correlation between the power usage at the prediction time and the power usage at the past time at the prediction time.

Here, the weights for the basic unit prediction model and the additional unit prediction models may be corrected to values continuously learned during the operation of predicting power demand for the target building.

When the short-term power demand prediction apparatus and method according to the spirit of the present invention having the above configuration are implemented, there is an advantage in that the calculation cost is low and the ultra-short-term power demand can be accurately predicted.

The short-term power demand prediction apparatus and method of the present invention have the advantage of being able to construct a deep neural network model in a scikit-learn environment of Python that does not require high-performance computer specifications.

The short-term power demand prediction apparatus and method of the present invention has the advantage that it takes very little learning time and evaluation time and shows better prediction performance than a single deep neural network model as well as various machine learning algorithms.

1 is a block diagram illustrating an embodiment of an apparatus for predicting short-term power demand according to the spirit of the present invention.
FIG. 2 is a block diagram illustrating a data collection unit and an input variable configuration unit for a prediction model in the short-term power demand prediction apparatus of FIG. 1 .
3 is a block diagram illustrating a prediction model constructing unit for performing a stacking ensemble type multiple power demand prediction method in the short-term power demand prediction apparatus of FIG. 1 .
4 is a flowchart illustrating an embodiment of a method for predicting short-term power demand according to the spirit of the present invention.

In describing the present invention, terms such as first, second, etc. may be used to describe various components, but the components may not be limited by the terms. The terms are only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component.

When a component is referred to as being connected or connected to another component, it may be directly connected or connected to the other component, but it can be understood that other components may exist in between. .

The terminology used herein is used only to describe specific embodiments, and is not intended to limit the present invention. The singular expression may include the plural expression unless the context clearly dictates otherwise.

In this specification, the terms include or include are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, and includes one or more other features or numbers, It may be understood that the existence or addition of steps, operations, components, parts, or combinations thereof is not precluded in advance.

In addition, shapes and sizes of elements in the drawings may be exaggerated for clearer description.

The present invention provides a predictive model input variable in consideration of only date/time information, weather information, and past power usage that can be collected anywhere, that is, easy to collect, for power demand prediction applicable to any building regardless of region/use. make up

Next, in performing preprocessing on the collected input variables, data interpolation and transformation of continuous time information into a rotational coordinate system are performed.

Next, based on the deep neural network, which is one of the deep learning models, several prediction models are constructed according to the number of hidden layers, and then principal component regression through the sliding window method using the prediction values derived from the prediction model. Construct an analysis-based stacking ensemble model.

Since the developed stacking ensemble model can effectively reflect the recent power demand pattern, it derives more sophisticated forecasting results than various previously proposed forecasting models. In addition, this provides a basis for effectively preparing for uncertainties in smart grid operation in the future, as it can perform accurate power demand forecasting even in multiple ultra-short-term power demand forecasts (a total of 96 time points from 15 minutes later to one day later).

1 is a block diagram illustrating an embodiment of an apparatus for predicting short-term power demand according to the spirit of the present invention.

The apparatus 100 for predicting short-term power demand according to an embodiment of the present invention as shown includes a data collection unit 110 for collecting date/time information, weather information, and power usage data of a predetermined period; a data pre-processing unit 120 performing feature extraction and maximum/minimum normalization on input data; and a prediction model constructing unit 160 that applies the preprocessed input data to two or more unit prediction models based on a deep neural network, and calculates a final prediction value by stacking ensemble processing the prediction values derived from each of the unit prediction models. can do.

The illustrated short-term power demand prediction apparatus 100 includes an input variable configuration unit including the data collection unit 110 and the data pre-processing unit 120 ; The predictive model component can be divided into two parts.

FIG. 2 is a block diagram illustrating a data collection unit 110 and an input variable configuration unit for a prediction model in the short-term power demand prediction apparatus of FIG. 1 .

FIG. 3 is a block diagram illustrating detailed configurations of the prediction model configuration unit 160 for performing the stacking ensemble type multiple power demand prediction method in the short-term power demand prediction apparatus of FIG. 1 .

It can be seen that the storage unit 140 is further expressed in comparison with the block diagram of FIG. 1 in FIGS. 2 and 3 .

The input variable configuration unit shown in FIG. 2 , specifically, the data collection unit 110 collects building power usage, weather information, and date/time data pertaining to a predetermined period of time.

In order to construct the power demand prediction model, it is proposed to first collect data having three characteristics: time information (time stamp corresponding to the collected power usage), weather information, and past power usage. The reason only this data is considered is that it is information that can be easily collected in the environment where the actual smart grid and microgrid are installed.

The data preprocessing process 126 for power usage information and date/time data performed by the data preprocessor 120 is exemplified as follows.

Power demand forecasting is a process of deriving a forecast value according to the information when the forecast model learns past information and inputs future information about the forecast time, so it is very important to know future information with high actual reliability.

Time information consists of month, day, hour, minute, day of the week, and public holidays, which is information that can be fully understood even in the future. Month, day, hour, and minute have periodicity, but if the collected values are applied as they are, there is a problem that periodicity cannot be reflected. For example, 11:45 and 12:00 (midnight) of the next day are adjacent times, but if the numerical value is expressed as it is, a difference of 95 will occur.

To solve this problem, spatial characteristics were expressed by augmenting one-dimensional temporal information with a two-dimensional rotational coordinate system. Through this, the one-dimensional data is converted into data having the characteristics of a watch. That is, the periodicity of month, day, hour, and minute data is reflected by applying Equations (1) to (8) below in date/time data. Through this, the predictive model can learn date/time information more effectively because input variables are constructed by reflecting close numbers instead of 23, which is the interval between 11 pm and midnight the next day.

은 해당 월의 마지막 일을 나타낸다. 예를 들어

은 31이며,

은 30이다. In Equations 3 and 4 above

represents the last day of the month. For example

is 31,

is 30.

Since the information on the day of the week and public holidays has the characteristic of a nominal scale, data can be expressed by indicating 1 on the corresponding day and 0 otherwise. For example, when the date of the prediction time is Monday and weekdays, the value of the variable corresponding to Monday is 1, and the values of the remaining variables, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday, and holidays, are expressed as 0.

In general, high power consumption is shown during working hours, so reflecting this can lead to an accurate power demand forecast. Since the information on the day of the week and holidays is an indicator that can indicate working hours, it is represented by a total of eight dummy variables. Here, public holidays consist of Saturdays, Sundays, and public holidays. Thus, a total of 16 date/time information is configured as input variables of the predictive model.

The pre-processing process 122 for the weather information performed by the data pre-processing unit 120 is exemplified as follows.

In the present invention, weather information of the Meteorological Administration's neighborhood forecast is collected so that it can be effectively applied in an actual smart grid environment. This is because a typical forecasting model predicts power demand based on the weather forecast of the Meteorological Administration's neighborhood forecast, so if the predictive model learns the actual weather information, it may be difficult to accurately predict the electric power demand due to the error between the forecasted value and the actual value. . The collected meteorological information is temperature, humidity, and wind speed, which are highly correlated with power consumption. Since they are collected every 3 hours, the 3-hour unit is subdivided into 15-minute units through linear interpolation.

This takes into account that the weather data was collected with a 3-hour resolution, so it is not appropriate to apply the weather data directly, and due to the difference from the forecast period, the forecast model. To obtain the same weather data, we estimate the weather data at 15-minute resolution using linear interpolation with the temporal resolution of electrical energy consumption.

Weather information consists of temperature, humidity, and wind speed predicted by the local weather forecast of the Korea Meteorological Administration. Reflecting only these three pieces of information is a factor that has a big impact on actual power consumption, considering that the information provided in 3-hour units from the Meteorological Administration's neighborhood forecast. However, since the neighborhood forecast forecasts in units of 3 hours, there is a disadvantage in that it is difficult to accurately predict because the units do not match when actually predicting the power demand in units of 15 minutes. weather information can be adjusted/reflected in 15-minute increments.

,

에서의 데이터 값이 각각

,

일 때,

,

사이의 임의의 지점

에서의 데이터 값

를 계산할 수 있다.

은

에서

까지의 거리,

는

에서

까지의 거리를 말합니다. 즉, 오후 3시의 기온이 15도이고, 오후 6시의 기온이 18도라면, 오후 4시와 5시의 기온은 16도와 17도라는 것으로 유추할 수 있다. usually two points

,

Each data value in

,

when,

,

any point between

data value in

can be calculated.

silver

at

distance to,

Is

at

refers to the distance to That is, if the temperature at 3 PM is 15 degrees and the temperature at 6 PM is 18 degrees, it can be inferred that the temperatures at 4 PM and 5 PM are 16 degrees and 17 degrees.

The pre-processing process 124 for the power usage information performed by the data pre-processing unit 120 is exemplified as follows.

For the past power usage, it is advantageous to use the actual power usage one day before and one week before the forecast time. It can also reflect other historical power usage, that is, power usage from 2 days to 6 days ago, but it does not show a high correlation with the power usage at the time of prediction, so the deep learning technique learns incorrectly (overfitting) when training the prediction model. problem) because

To prove this, the Pearson correlation coefficient may be used to determine the correlation between the power usage at the prediction time and the power usage at the past time at the prediction time, and specifically, Equation 10 below.

The Pearson correlation coefficient, which measures the linear correlation between two variables x and y, is the value of the covariance of these variables divided by the product of the standard deviation of equal or proportional scale data, and is a value between +1 and -1 according to the Cauchy-Schwarz inequality. can have where +1 is a perfect positive linear correlation, 0 indicates no linear correlation, and -1 is a perfect negative linear correlation. In general, the numerical value of the Pearson correlation coefficient can be interpreted as follows.

If r is between -1.0 and -0.7, then there is a strong negative linear relationship

If r is between -0.7 and -0.3, then there is a pronounced negative linear relationship

If r is between -0.3 and -0.1, a weakly negative linear relationship

If r is between -0.1 and +0.1, then a linear relationship that can be almost neglected

If r is between +0.1 and +0.3, then there is a weak quantitative linear relationship

If r is between +0.3 and +0.7, then there is a clear quantitative linear relationship

If r is between +0.7 and +1.0, then there is a strong quantitative linear relationship

),...,(

)}에 대해 표본을 기반으로 공분산과 분산을 대체하는 추정값인

를 산출할 수 있다. 여기서,

은 샘플 크기를 나타내며,

와

는 시간

로 인덱스된 개별 샘플 시점이다.

와

는 각각

와

의 표본 평균이다. In Equation 10, paired data consisting of n pairs {(

),...,(

), which is an estimate that substitutes for covariance and variance based on the sample for

can be calculated. here,

represents the sample size,

Wow

is the time

Individual sample time points indexed by .

Wow

is each

Wow

is the sample mean of

When the Pearson correlation coefficient is calculated between the power consumption at the prediction time and the power usage from 1 to 7 days in the past from the data set you have, the power usage at the prediction time and the power usage one day and one week before the prediction time The Pearson's correlation coefficient was calculated to be 0.7 or higher, indicating that it was a variable with high correlation, and the present invention also considered this as an input variable.

In summary, historical electric load is a variable that can reflect the recent electricity usage trend, but reflecting all of them can cause overfitting of the forecasting model. The Pearson correlation coefficient was calculated from the same time point one day before to the same time point one week ago, and it was confirmed that the day before and one week before the prediction time showed a high correlation. So, this was reflected as an input variable of the predictive model.

In the predictive model constructing units 162, 164, and 166 shown in FIG. 3, for effective deep neural network model learning, the input variables configured in the input variable constructing unit are subjected to Min-Max normalization between 0 and 1 change to the values of Next, the data collection period is divided into a training set, a validation set, and an evaluation set at a constant rate.

In FIG. 3, a normalization unit 146 that performs min-max normalization on the input information recorded in the storage unit 140; And it is expressed as further comprising a data set configuration unit 148 constituting a basic unit prediction model training data set, an extended prediction model configuration data set, an actual prediction data set, and the like.

In addition, in FIG. 3, the prediction modeling unit 160 of FIG. 1 is a deep neural network (DNN) model having a minimum hidden layer, and learning (determining weights for each node and each branch) is performed on a basic unit prediction model. a basic unit prediction model learning unit 162; From the basic unit prediction model that has been trained, one or more additional unit prediction models are derived as a deep neural network (DNN) model with an increased hidden layer, and a multiple prediction model composed of the basic unit prediction model and the additional unit prediction models is generated. a multi-prediction model configuration unit 169; and an ensemble prediction unit 166 that calculates a final prediction value by performing stacking ensemble processing on individual prediction values of the basic unit prediction model and the additional unit prediction models.

Table 1 below illustrates input variables that may constitute the input data set of each learning model.

A total of 21 input variables are configured in Table 1, and for effective learning of the deep neural network model, these variables may be subjected to maximum-minimum normalization as shown in Equation 11 below by the regularizer 146 .

는 변수

에 속해 있는

라는 변숫값의 정규화를 수행한 값이며,

와

은 데이터 셋

에서 가장 큰 값과 작은 값을 나타낸다. 그에 따라, 데이터 셋의 전체 변수는 0에서 1 사이의 값을 가지게 되며, 이를 심층 신경망 모델에 적용한다.in the above formula

is a variable

belonging to

It is a value obtained by normalizing the variable value of

Wow

silver data set

represents the largest and smallest values in Accordingly, all variables in the data set have a value between 0 and 1, and this is applied to the deep neural network model.

A general artificial neural network derives a predicted value through a process as in Equation 12 below.

는 비선형 함수이고,

는 연속형 또는 이산형 변수들의 벡터로 입력 변수이며,

는 연속형 또는 이산형 변수들의 벡터로 출력 노드이다.

는

의 함수로

는 은닉 노드를 가리키며,

는 활성화 함수이며,

는 가중치이다. 여기서 활성화 함수는 임계치를 적용해 의미 없는 데이터를 사전에 필터링하고 미분을 편하게 하기 위해서 사용하는 함수이다. in the above formula

is a nonlinear function,

is the input variable as a vector of continuous or discrete variables,

is the output node as a vector of continuous or discrete variables.

Is

as a function of

points to the hidden node,

is the activation function,

is the weight. Here, the activation function is a function used to filter meaningless data in advance by applying a threshold value and to facilitate differentiation.

The important thing in a neural network is to adjust the weight, and the accuracy of the neural network model varies depending on how the weight is adjusted, so the learning result can also be reflected as a weight. Feedforward is a method of calculating the input layer, hidden layer, and output layer in order. Backpropagation is a representative supervised learning algorithm used when training a feedforward neural network with multiple hidden layers with labeled training data. to be. That is, it is a forward phase + backward phase, and it goes from the input layer to the hidden layer and the output layer in order, then goes back from the output layer to the hidden layer to the input layer, correcting the weights and repeating the above process to get the best result.

A deep neural network (DNN) is an artificial neural network composed of several hidden layers between an input layer and an output layer. Deep neural networks can model complex non-linear relationships like general artificial neural networks. The advantage of deep neural networks is that analysis is possible regardless of continuous and categorical variables, and nonlinear combinations between input variables (when input variables actually have high correlation with power consumption but low Pearson correlation coefficient) are possible. In addition, prediction performance is often relatively superior to other machine learning techniques, and since feature extraction is performed automatically, the hassle of variable selection can be reduced, and performance continues to improve as the amount of data increases.

However, when the neural network is complex, it takes a long time to operate, requiring a computer equipped with a GPU and a high-end computer, and it takes a long time to detect the optimal model because various hyperparameters must be considered. Combinations that are time- and computationally expensive do not guarantee accuracy in other data sets.

As hyperparameters of neural networks, various combinations such as number of hidden layers, number of hidden nodes, activation function, epoch, learning rate, optimizer, and batch size should be considered. It is very important to determine the optimal number of hidden layers because it greatly affects the model performance.

In general, learning rate or optimizer can roughly figure out the optimal value by referring to other research cases, but the number of hidden layers is not. The present invention creates several deep neural network models irrespective of the number of hidden layers, and uses the power demand forecast value primarily predicted from the deep neural networks to more accurately predict the power demand through the stacking ensemble technique.

The stacking ensemble technique is a technique for deriving a final prediction by effectively combining multiple results of various prediction models. This is a technique performed under the assumption that the prior information is sufficiently reflected in the prediction value because each prediction model has already learned and predicted sufficient information in advance.

For this purpose, principal component regression analysis can be used. Principal component regression analysis is a technique for creating a regression model using the principal components after extracting the principal components of the input variables. That is, in principal component analysis + multiple linear regression analysis, since the principal components are orthogonal to each other, multicollinearity (difficulty in selecting weights during model training because input variables are very similar) does not occur.

Therefore, multiple linear regression models can be created with confidence. In addition, if only the top few variables among the principal component variables transformed using principal component analysis are used, a standardization effect can be given, and the problem of overfitting the model (mislearning) can also be alleviated.

However, when constructing a general prediction model, the training data set and the evaluation data set are classified, the prediction model is constructed from the training data set, and it is applied to the evaluation data set as it is. In doing so, when the interval between the time of constructing the model and the time of prediction becomes large (the end of the evaluation data set), there is a disadvantage that the prediction performance deteriorates.

To compensate, predictive models must reflect recent trends and patterns in real time. Thus, several deep neural network models (deep neural networks with 4 to 7 layers each) use the predicted values from the evaluation data set. Here, we do not apply all forecasts, only the forecasts for the past week.

For example, assuming that the electricity usage (96 points in 15-minute increments) is predicted for 45 minutes from 1:00 PM on April 1 to midnight on April 2, from 1:15 PM on March 25th to April 1 PM A total of 7 days up to 12:45 are considered. A prediction model based on principal component regression analysis is constructed using the predicted values of each prediction model (4 deep neural network models), and the prediction model based on principal component regression is constructed from 1:00 PM on April 1st to midnight 45 minutes on April 2nd of the four deep neural network models. When the predicted power usage of , is input to the principal component regression model, the principal component regression model is finally constructed in such a way that it predicts the power usage at 96 points in time.

In general, the hyperparameters of the neural network, particularly epoch, are determined by learning and determining optimal values by the multi-prediction model configuration unit at the time of designing an electric power demand prediction device for a specific building (site), for the specific building (site). During the power demand forecasting operation, the optimal values for the above-described parameters are not changed.

On the other hand, the weights for the basic unit prediction model and the additional unit prediction model constituting the multi-prediction model are corrected with values continuously learned during the operation of predicting the power demand for the specific building. To this end, a sliding window as shown inside the ensemble prediction unit 166 of FIG. 3 may be applied.

In addition to the weight, some of the hyperparameters may be corrected to values that are continuously learned during the operation of predicting power demand for the specific building.

4 is a flowchart illustrating an embodiment of a method for predicting short-term power demand according to the spirit of the present invention.

The illustrated short-term power demand prediction method includes the steps of collecting weather information and power usage data for a predetermined period (S110); performing pre-processing including feature extraction of the collected data and maximum minimum normalization (S120); performing learning (determining weights for each node and each branch) on a basic unit prediction model having a minimum hidden layer (S160); deriving one or more additional unit prediction models with an increased hidden layer from the basic unit prediction model for which learning is completed, and constructing a multi-prediction model including the basic unit prediction model and the additional unit prediction models (S170); and performing stacking ensemble processing on individual prediction values of the basic unit prediction model and the additional unit prediction models to calculate a final prediction value ( 180 ).

The multi-prediction model prediction method applying the stacking ensemble technique according to the flowchart shown is to construct a deep neural network model with two hidden layers using the data set of the training set to solve the overfitting problem of the prediction model, and optimize the model in the validation set. Epoch is detected. Next, four deep neural network models with different number of hidden layers are developed using the data sets of the training set and the validation set by applying the detected Epoch, and then applied to the evaluation set to derive predicted values.

For example, as shown in Equation 13 below, a stacking ensemble model using sliding window-based principal component regression may be configured through the predicted values of four deep neural network models. Principal component regression analysis can account for most data variability with only a few principal components, and has advantages such as multicollinearity between input variables and prevention of overfitting.

For example, more specifically, a stacking ensemble model with a sliding window size of one week and one main component is constructed, and using this, the ultra-short-term power demand prediction is performed in 15-minute units from 15 minutes after the present time to one day later. . The present invention can be applied not only in units of 15 minutes, but also in units of 30 minutes and 1 hour.

Those skilled in the art to which the present invention pertains should understand that the present invention can be embodied in other specific forms without changing the technical spirit or essential characteristics thereof, so the embodiments described above are illustrative in all respects and not restrictive. only do The scope of the present invention is indicated by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. .

110: data collection unit
120: preprocessor
160: predictive modeling unit

Claims (16)

a data collection unit for collecting date/time information, weather information, and power usage data of a predetermined period;
a data pre-processing unit performing pre-processing including feature extraction of the collected data and maximum/minimum normalization; and
A predictive modeling unit that applies the preprocessed input data to unit prediction models having two or more different number of layers based on a deep neural network, and calculates a final prediction value from the prediction values derived by each of the unit prediction models
A stacking ensemble type short-term power demand forecasting device comprising a.
According to claim 1,
The predictive modeling unit,
a basic unit prediction model learning unit that performs learning on a basic unit prediction model having a minimum hidden layer;
a multi-prediction model construction unit that derives one or more additional unit prediction models with an increased hidden layer from the basic unit prediction model that has been trained, and generates a multi-prediction model composed of the basic unit prediction model and the additional unit prediction models; and
An ensemble prediction unit that calculates a final prediction value by performing stacking ensemble processing on individual prediction values of the basic unit prediction model and the additional unit prediction models
Short-term power demand forecasting device comprising a.

3. The method of claim 2,
The basic unit prediction model and the additional unit prediction models are a short-term power demand prediction device that is a deep neural network model.
4. The method of claim 3,
The ensemble prediction unit,
A short-term power demand prediction apparatus for performing principal component regression analysis to create a regression model using the principal components after extracting principal components from each of the individual prediction values of the additional unit prediction models based on a sliding window.
3. The method of claim 2,
The ensemble prediction unit,
A short-term power demand prediction apparatus for stacking ensemble processing by applying a sliding window so that the weights for the basic unit prediction model and the additional unit prediction models are corrected to the values learned continuously during the power demand prediction operation.
According to claim 1,
The data preprocessor,
Augmenting one-dimensional time information into a two-dimensional rotational coordinate system,
A short-term power demand forecasting device that indicates the day of the week and public holidays as 1 on the day corresponding to the holiday, and 0 otherwise.
According to claim 1,
The data preprocessor,
A short-term power demand forecasting device that applies the Pearson correlation coefficient to determine the correlation between the power usage at the time of prediction and the power usage at the time of the past at the time of prediction.
According to claim 1,
A storage unit in which collected data or pre-processed data and learning results are stored
Short-term power demand forecasting device further comprising a.
According to claim 1,
The short-term power demand prediction device in which the weights for the basic unit prediction model and the additional unit prediction models are corrected to values that are continuously learned during the operation of predicting power demand for a target building.
collecting weather information and power usage data for a predetermined period;
performing preprocessing including feature extraction of the collected data and maximum minimum normalization;
performing learning on a basic unit prediction model having a minimum hidden layer;
deriving one or more additional unit prediction models with an increased hidden layer from the basic unit prediction model for which learning is completed, and constructing a multi-prediction model composed of the basic unit prediction model and the additional unit prediction models;
Calculating a final predicted value by performing stacking ensemble processing on individual prediction values of the basic unit prediction model and the additional unit prediction models
Stacking ensemble type short-term power demand forecasting method comprising a.
11. The method of claim 10,
The basic unit prediction model and the additional unit prediction models are short-term power demand prediction methods that are deep neural network models.
11. The method of claim 10,
In the step of calculating the final predicted value,
A short-term power demand prediction method for performing principal component regression analysis to create a regression model using the principal components after extracting principal components from each of the individual prediction values of the additional unit prediction models based on a sliding window.
11. The method of claim 10,
In the step of calculating the final predicted value,
A short-term power demand prediction method for stacking ensemble processing by applying a sliding window so that the weights for the basic unit prediction model and the additional unit prediction models are corrected to the values learned continuously during the power demand prediction operation.
11. The method of claim 10,
In the step of performing the pre-processing,
Augmenting one-dimensional time information into a two-dimensional rotational coordinate system,
Short-term electricity demand forecasting method in which information on days of the week and holidays is marked with 1 on the day corresponding to a public holiday and 0 otherwise.
11. The method of claim 10,
In the step of performing the pre-processing,
A short-term power demand forecasting method that applies the Pearson correlation coefficient to determine the correlation between the power consumption at the time of prediction and the power usage at the time of the past at the time of prediction.
11. The method of claim 10,
The short-term power demand prediction method in which the weights for the basic unit prediction model and the additional unit prediction models are corrected to values that are continuously learned during the operation of predicting power demand for a target building.

Images