KR102518629B1 - Method for forecasting electric power demand using convolutional neural network, recording medium and device for performing the method - Google Patents

Abstract

A power demand prediction method using CNN includes the steps of collecting data of meteorological factors for power demand prediction from API, which is online public data; Compensating for missing data of the collected data; Selecting a predictor to be reflected in the prediction of power demand for each section by analyzing the correlation between predictors through CNN, which is a convolutional neural network that takes supplemented data as input; and selecting a final power demand with higher accuracy for each section through a parallel structure of CNN-LSTM and CNN-GRU. Accordingly, it is possible to accurately predict power demand.

Description

Method for predicting power demand using CNN, recording medium and device for performing the same

The present invention relates to a method for predicting power demand using CNN, and to a recording medium and apparatus for performing the same, and more particularly, to reflect power demand prediction based on a CNN algorithm capable of identifying a power demand pattern through image processing. It is about the technology of selecting elements.

It was announced by the International Energy Agency (IEA) that energy savings in equipment and facilities accounted for about 36% of the policy measures implemented to reduce greenhouse gas emissions by 2050. Accordingly, as energy consumption continuously increases, efficient energy management is becoming more important.

We are expanding businesses related to energy forecasting, management and monitoring at home and abroad, including large power generation companies, and are developing independent energy demand forecasting models and establishing monitoring systems for efficient energy management.

In addition, in the process of shifting the government's awareness of future-oriented new energy industries from supply-oriented to demand-side management, technologies for increasing the efficiency of energy management for energy-consuming buildings and integrated management technologies are needed.

The change to a low-energy consumption social structure and the diversification of energy consumption and production are increasing the demand for efficient energy use and integrated management. In addition, it is necessary to derive improvement directions through performance management for energy production and reduction through integrated energy management, verification of energy self-sufficiency before and after improvement, and analysis of measurement data.

In particular, when analyzing energy solutions, it is necessary to derive reliable predictive data for various types of energy consumption patterns through a power demand forecasting model using AI techniques.

For example, interest in reducing energy consumption and improving performance in the building sector, such as smart grids, is continuously increasing. Therefore, by using EMS (Energy Management System) technology to collect and visualize data such as customer operation status and energy consumption, it is a data-based management system that is not a building management method that relied on the experience or ability of existing building operators and managers. is operating

Therefore, it is necessary to analyze the energy consumption pattern in order to operate the data-based management system and to manage and improve the energy efficiency of consumers.

KR 10-2020-0131928 A KR 10-2020-0123310 A US 2015/0046221 A1

Accordingly, the technical problem of the present invention is conceived in this respect, and an object of the present invention is to provide a power demand prediction method using CNN.

Another object of the present invention is to provide a recording medium on which a computer program for performing the power demand prediction method using the CNN is recorded.

Another object of the present invention is to provide an apparatus for performing the power demand prediction method using the CNN.

A method for predicting power demand using CNN according to an embodiment for realizing the object of the present invention includes the steps of collecting data of meteorological factors for predicting power demand from API, which is online public data; Compensating for missing data of the collected data; Selecting a predictor to be reflected in the prediction of power demand for each section by analyzing the correlation between predictors through CNN, which is a convolutional neural network that takes supplemented data as input; and selecting a final power demand with higher accuracy for each section through a parallel structure of CNN-LSTM and CNN-GRU.

In an embodiment of the present invention, the data of the meteorological elements may include at least one of weather data according to time and interval, daily weather index data, and past power demand data.

In an embodiment of the present invention, the step of collecting the data of the meteorological factors may include normalizing the collected data of the meteorological factors in order to reduce the scale difference of the data.

In an embodiment of the present invention, the supplementing of missing data of the collected data may include determining whether or not there is missing data through an overlap ratio of data patterns; and complementing missing data using mean interpolation when the number of missing data is 2 or less, and supplementing missing data through an ARIMA time series prediction model when the number of missing data is 3 or more. there is.

In an embodiment of the present invention, the step of selecting a predictor to be reflected in the power demand forecast for each section includes image processing of a pattern of power demand and a predictor for each section, and selecting factors to be reflected in the prediction of power demand. It can be used for forecasting electricity demand.

In an embodiment of the present invention, the power demand prediction method using the CNN may further include analyzing errors of power demand values predicted N days ago through MAPE and RMSE functions.

In an embodiment of the present invention, the power demand prediction method using the CNN may further include classifying and resampling data to integrate sampling rates of selected predictors.

A computer program for performing the power demand prediction method using the CNN is recorded in a computer readable storage medium according to an embodiment for realizing the above object of the present invention.

An apparatus for predicting power demand using CNN according to an embodiment for realizing another object of the present invention described above includes a data collection unit that collects data of meteorological factors for predicting power demand from API, which is online public data; a data complementation unit that compensates for missing data of the collected data; a predictor selection unit that analyzes the correlation between predictors through CNN, a convolutional neural network that takes the supplemented data as an input, and selects a predictor to be reflected in the electricity demand forecast for each section; and a power demand prediction unit that selects a final power demand with higher accuracy for each section through a parallel structure of CNN-LSTM and CNN-GRU.

In an embodiment of the present invention, the data supplementation unit determines whether there is missing data through an overlap ratio of data patterns, and when the number of missing data is 2 or less, compensates for missing data using mean interpolation, If the number of missing data is three or more, the missing data can be supplemented through the ARIMA time series prediction model.

According to the power demand prediction method using such a CNN, it is possible to predict the final power demand by adopting an algorithm having higher accuracy for each time and section by applying an algorithm that distinguishes time and section.

Accordingly, it can contribute to economic benefits and energy efficiency improvement through the development of an AI-based power demand prediction algorithm and optimal operation scheduling for each customer. In addition, it is possible to improve the accuracy and objectivity of demand forecasting by utilizing key variables that affect electricity demand, such as meteorological factors, regional population density, and daily weather index.

1 is a diagram for explaining a power demand prediction algorithm proposed in the present invention.
2 is a block diagram of an apparatus for predicting power demand using a CNN according to an embodiment of the present invention.
FIG. 3 is an example of meteorological data used for correlation analysis in the data collection unit of FIG. 2 .
FIG. 4 is an example of past power demand data from 1 hour to 6 hours ago used for correlation analysis in the data collection unit of FIG. 2 .
FIG. 5 is an example of season and cycle, holidays, afternoon and workinghour data used for correlation analysis in the data collection unit of FIG. 2 .
FIG. 6 is a diagram for explaining a Moving Window technique used in the data complementation unit of FIG. 2 .
7 is an example of data interpolation using an ARIMA predictive model in the data complementation unit of FIG. 2 .
8 is a diagram for explaining the CNN-LSTM or GRU combined model structure used in the present invention.
9 is a diagram showing the basic structure of an LSTM network.
10 is a diagram showing the basic structure of a GRU.
11 is a flowchart of a power demand prediction method using CNN according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The detailed description of the present invention which follows refers to the accompanying drawings which illustrate, by way of illustration, specific embodiments in which the present invention may be practiced. These embodiments are described in sufficient detail to enable one skilled in the art to practice the present invention. It should be understood that the various embodiments of the present invention are different from each other but are not necessarily mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in one embodiment in another embodiment without departing from the spirit and scope of the invention. Additionally, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the invention. Accordingly, the detailed description set forth below is not to be taken in a limiting sense, and the scope of the present invention, if properly described, is limited only by the appended claims, along with all equivalents as claimed by those claims. Like reference numbers in the drawings indicate the same or similar function throughout the various aspects.

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

1 is a diagram for explaining a power demand prediction algorithm proposed in the present invention. 2 is a block diagram of an apparatus for predicting power demand using a CNN according to an embodiment of the present invention.

An apparatus for predicting power demand (10, hereinafter) using a CNN according to the present invention is for accurately predicting power demand. Various factors can be used to predict electricity demand. Meteorological data (temperature, humidity, wind speed, cloudiness, etc.) according to time (n days ago, 1 week ago, etc.) and intervals (season, quarter, etc.), daily weather index (unpleasant index, perceived temperature, etc.), and past power demand data.

These data are collected through API, which is online public data, and include measures to analyze data and reconstruct damaged data that is intermittently found.

In the present invention, CNN, which is a representative algorithm capable of identifying power demand patterns through image processing, is used to select factors to be reflected in power demand prediction.

Referring to FIG. 1, the algorithm for predicting actual power demand in the present invention is divided into two types. First, LSTM is a deep learning algorithm having high performance in predicting time series data. Second, GRU is a simplified version of LSTM, a deep learning algorithm that reduces three gates to two. Compared to GRU, LSTM tends to have better prediction performance than GRU as the amount of computation is larger and the prediction data size is larger.

In the present invention, an algorithm for predicting power demand by constructing an ensemble having a parallel structure of CNN-LSTM and CNN-GRU is proposed. The present invention is an algorithm that predicts the final power demand by adopting an algorithm having higher accuracy for each time and section by applying an algorithm that distinguishes time and section.

Referring to FIG. 2 , the apparatus 10 according to the present invention includes a data collection unit 110, a data supplementation unit 130, a predictor selection unit 150, and a power demand estimation unit 170.

In the apparatus 10 of the present invention, software (application) for predicting power demand using CNN may be installed and executed, and the data collection unit 110, the data complement unit 130, and the predictor selection Configurations of the unit 150 and the power demand prediction unit 170 may be controlled by software for performing power demand prediction using the CNN, which is executed in the device 10 .

The device 10 may be a separate terminal or a part of a module of the terminal. In addition, the configuration of the data collection unit 110, the data supplementation unit 130, the predictor selection unit 150, and the power demand estimation unit 170 may be formed as an integrated module or composed of one or more modules. can However, on the contrary, each component may be composed of a separate module.

The device 10 may be mobile or stationary. The apparatus 10 may be in the form of a server or engine, and may be a device, an apparatus, a terminal, a user equipment (UE), a mobile station (MS), or a wireless device. It can be called by other terms such as wireless device, handheld device, etc.

The device 10 may execute or manufacture various software based on an operating system (OS), that is, a system. The operating system is a system program for enabling software to use the hardware of the device, and is a mobile computer operating system such as Android OS, iOS, Windows mobile OS, Bada OS, Symbian OS, Blackberry OS, and Windows-based, Linux-based, Unix-based, It can include all computer operating systems such as MAC, AIX, and HP-UX.

The data collection unit 110 collects data of meteorological factors for predicting power demand from API, which is online public data.

In one embodiment of the present invention, meteorological factors necessary for predicting power demand can be analyzed through a CNN algorithm. Data such as analyzed temperature, humidity, atmospheric pressure, and wind speed can be collected through APIs such as the meteorological data open portal of the Korea Meteorological Administration. For the flexibility and sensitivity of algorithm execution, the data collection period can be set for a long period of time, such as the last 1 year or 3 years. FIG. 3 is an example of meteorological data used for correlation analysis in the data collection unit of FIG. 2 .

Various factors can be used to predict electricity demand. Meteorological data (temperature, humidity, wind speed, cloudiness, etc.) according to time (n days ago, 1 week ago, etc.) and intervals (season, quarter, etc.), daily weather index (unpleasant index, perceived temperature, etc.), and past power demand data.

These data are collected through API, which is online public data, and the present invention includes a method of analyzing data and reconstructing damaged data that is intermittently found.

Short-term electricity demand forecast is directly influenced by meteorological factors along with past electricity demand data. Therefore, in order to accurately predict electricity demand, necessary data must be analyzed through correlation analysis with various meteorological factors. FIG. 4 is an example of past power demand data from 1 hour to 6 hours ago used for correlation analysis in the data collection unit of FIG. 2 .

In addition, the analysis can be performed by considering the perceived temperature and discomfort index provided by the Korea Meteorological Administration. Equations for the discomfort index and the perceived temperature are shown in Equation 1 and Equation 2 below, respectively.

[Equation 1]

[Equation 2]

In addition, the power demand shows a shape in which the pattern changes periodically. Since the power demand pattern is different depending on the four seasons, weekdays, weekends, and holidays, analysis can be performed considering not only meteorological factors but also seasonal changes, past history data, weekdays, weekends, and shared days to forecast power demand. . FIG. 5 is an example of season and cycle, holidays, afternoon and workinghour data used for correlation analysis in the data collection unit of FIG. 2 .

에서

으로 표시하였으며 계절에 해당하는 변수는

으로 선정한다. 또한, 각 월, 요일, 각 주중에 해당하는 변수는

,

,

로 선정하고, 주중 및 주말에 해당하는 변수는

이며 공휴일 유무, 그리고 24시간 중 오후 시간대 및 근무시간에 해당하는 값을 각각

로 선정한다.First, referring to FIG. 4 , in order to analyze the past power demand correlation, analysis of past data from a total of 1 hour to 6 hours ago is performed. Variables corresponding to each time period are

at

, and the variable corresponding to the season is

select as In addition, the variables corresponding to each month, day of the week, and weekday are

,

,

, and the variables corresponding to weekdays and weekends are

, whether there is a holiday, and the value corresponding to the afternoon time zone and working hours among 24 hours

select as

Adoption of a predictor for predicting electricity demand is generally calculated as shown in Equation 3 below based on the Pearson's correlation analysis equation and a scatter plot graph.

[Equation 3]

는 피어슨 상관계수를 의미하며 Y는 전력수요 데이터, X는 상관분석을 위해 사용된 인자들을 의미한다.

은 데이터 개수를 의미하며, 각 데이터간의 공분산값과 표준편차 값을 이용하여 구해진다. 도 5는 시간적 특징을 나타내는 플롯(plot)이다. LSTM, GRU 같은 경우 기상요소, 과거 데이터뿐만 아니라 시계열 특징을 나타내는 데이터도 중요한 요인으로 작용할 수 있기 때문에 함께 분석한다.here,

is the Pearson correlation coefficient, Y is the power demand data, and X is the factors used for correlation analysis.

Means the number of data, and is obtained using the covariance value and standard deviation value between each data. 5 is a plot showing temporal characteristics. In the case of LSTM and GRU, not only meteorological factors and past data, but also data representing time series characteristics can act as important factors, so they are analyzed together.

The data collection unit 110 may first process and pre-process data in order to predict power demand.

If the scale difference of the data itself is severe, such as meteorological element data, it acts as a cause of deteriorating prediction performance. Therefore, normalization of each data may be performed to improve prediction performance and prevent overfitting.

Data normalization methods include the Min-Max Normalization method using the maximum and minimum values of each data, and the Z-Score Normalization method using the standard deviation and average of the data. Each normalization equation can be expressed as Equation 4 and Equation 5 below.

[Equation 4]

[Equation 5]

The data compensating unit 130 supplements missing data of the collected data. First, the presence or absence of missing data can be determined using a moving window technique.

Domestic power demand, meteorological factors, and power demand data by site may have missing data or abnormal data when actually stored in the database. Such data adversely affects the accuracy of electricity demand forecasting, and is particularly applied as a factor that increases inaccuracy in deep learning-based algorithms such as LSTM. Therefore, additional work is required to handle missing values.

Referring to FIG. 6, the moving window technique is a method used in many fields to automatically observe data as a method of determining the presence or absence of missing data through an overlap ratio of data patterns.

In one embodiment of the present invention, when missing data is observed through a moving window technique, a processing method is changed according to the number of missing data.

For example, if the number of missing data is 2 or less, the missing data can be supplemented using Mean Interpolation by analyzing data before and after the missing point. Mean Interpolation is a technique that replaces missing parts with the mean of each variable. When the number of missing data is small, the prediction performance is not degraded even if the data are supplemented through mean interpolation.

On the other hand, if the number of missing data is three or more, the missing data can be supplemented through the ARIMA time series prediction model.

In particular, electricity demand and meteorological factors do not cause large changes in a short time, so even using time-series-based interpolation, high accuracy is shown. An example of data interpolation using the ARIMA predictive model can be shown as shown in FIG.

The predictor selection unit 150 may select a predictor to be reflected in the prediction of power demand for each section by analyzing a correlation between predictors through CNN, which is a convolutional neural network that takes supplemented data as an input.

In addition, the correlation coefficient analysis for each predictor is automatically selected based on the high correlation coefficient within the power demand prediction algorithm, and can be implemented to select the optimal input variable whenever data is updated.

Correlation analysis is automatically updated through a convolutional neural network (CNN) to select predictive factors to be reflected in the electricity demand forecast for each section. In the present invention, in order to save time and prevent overfitting, correlations between different predictor factors for each section are analyzed, and factors having a correlation coefficient greater than or equal to a preset threshold value are selected and used for predicting power demand.

In the present invention, factors having high correlation with power demand patterns are used as input data through CNN.

The power demand prediction unit 170 selects the final power demand with higher accuracy for each section through a parallel structure of CNN-LSTM and CNN-GRU.

In the present invention, the algorithm for predicting actual power demand is divided into two types. First, LSTM is a deep learning algorithm that has high performance in predicting time series data. Second, GRU is a simplified version of LSTM, a deep learning algorithm that reduces three gates to two. Compared to GRU, LSTM tends to have better prediction performance than GRU as the amount of computation is larger and the prediction data size is larger.

Referring to FIG. 8, as a new method for predicting power demand, the present invention performs power demand prediction by combining a CNN, which is a convolutional neural network, with an LSTM or GRU model.

Existing Pearson's correlation analysis alone has a disadvantage in that it cannot accurately explain the relationship between power demand data and predictors. This is because Pearson's correlation coefficient does not consider seasonal cycles or data patterns during the day.

Using CNN, a convolutional neural network, it is possible to extract highly correlated data by considering patterns between power demand patterns and other factors. This is because it is possible to automatically extract features of data by applying a convolution layer in CNN. Therefore, prediction accuracy can be improved by developing a hybrid model that can utilize the advantages of both deep learning technologies.

Long Short Term Memory (LSTM) is a part of Recurrent Neural Network (RNN) and is a technique used when learning or performing data for a long time. When predicting time series data, RNN has a disadvantage that its learning ability is greatly reduced when the predicted data size increases. A method devised to overcome this problem is a structure in which cell-state is added to the hidden state of LSTM RNN.

Since it is currently used in most time series predictions and shows high accuracy in predicting power demand, the present invention uses LSTM to predict power demand.

Referring to FIG. 9, the general structure of LSTM is composed of Cell, Input Gate Layer, Forget Gate Layer, and Output Gate Layer. Cells store values at time intervals, and three gate layers adjust data weights in and out of cells.

,

를 연결한 부분을 셀 스테이트(Cell state)라고 한다. 이는 마이너한 선형 연산만을 거치고 전체 체인을 관통한다. As the most important structure, with the horizontal line at the top,

,

The part connected is called a cell state. It goes through only minor linear operations and goes through the entire chain.

Compared to RNN networks, LSTM networks have a more refined gate structure, which is a device that allows information to flow selectively. Each consists of a sigmoid neural network layer and a point-by-point multiplication operation. The sigmoid layer is output with a value of 0,1.

The output data indicates whether the component affects the predicted data. In the case of 0, it means that the component does not affect the predicted data, and in the case of 1, it means that it affects the predicted data.

와

를 입력값으로 받아 시그모이드 레이어에 저장하며, 첫 번째, 시그모이드 레이어는 업데이트할 값을 결정하는 역할을 한다.The first step of an LSTM is a value between 0 and 1.

and

is received as an input value and stored in the sigmoid layer. First, the sigmoid layer plays a role in determining the value to be updated.

The tanh layer plays a role of generating candidate values that can be added to the cell state, and is combined with the value of the sigmoid layer to affect the next state. The input gate function expression is expressed as Equation 6 below.

[Equation 6]

,

를 로 업데이트 하는 과정은 전 단계에서 결정하였으므로, 새로운 후보 값이 기존 값에 영향을 주게 한다. 망각 게이트의 함수식은 아래의 수학식 7 및 수학식 8과 같으며, tanh은 -1에서 1사이의 값을 가진다.

,

Since the process of updating to is determined in the previous step, the new candidate value affects the existing value. The function formula of the forget gate is as shown in Equations 7 and 8 below, and tanh has a value between -1 and 1.

[Equation 7]

[Equation 8]

A function expression in which a new candidate value affects an existing value is shown in Equation 9 below.

[Equation 9]

Finally, when determining the output value, it is derived through the sigmoid layer. At this time, the function expression of the output gate is as shown in Equation 10 below, and the expression reflecting the final result value is as shown in Equation 11.

[Equation 10]

[Equation 11]

Existing RNN techniques have a problem of data disappearing when processing time series data, so data is classified and predicted through LSTM technique. The LSTM technique is insensitive compared to RNN even when the interval length is long, so it does not cause problems, so it is widely used for learning and predicting data that occurs over a long period of time.

The GRU (Gate Recurrent Unit) cell is a simplified version of the LSTM cell that compensates for the long LSTM operation speed. 10 is a diagram showing the basic structure of a GRU. The basic concept of GRU is very similar to LSTM, which can be seen in Equation 12 below.

[Equation 12]

GRU has two gates, Reset Gate and Update Gate. As the name of the gate suggests, the Reset Gate determines how new inputs are merged with the old memory, and the Update Gate determines how much of the old memory is remembered.

If all reset gate values are set to 1 and all update gate values are set to 0, the basic RNN structure becomes. GRU has the same basic idea as LSTM, but has the following differences.

1) There are 2 GRU Gates and 3 LSTM Gates.

)이 외부에서 보게 되는 Hidden state 동일함. 이는 LSTM에 있는 출력 게이트가 없기 때문임.2) GRU is the internal memory value (

) is the same as the hidden state seen from the outside. This is because there is no output gate in LSTM.

3) Input Gate and Forget Gate are merged into Update Gate, and Reset Gate is applied directly to the previous hidden state value.

4) No additional non-linear function is applied when calculating the output value.

GRU has fewer parameters to tune than LSTM, so it takes shorter learning time than LSTM and can learn with less data. However, since LSTM can show excellent learning ability when the number of training data is sufficient, in this study, both algorithms are compared and verified before use.

In the present invention, an algorithm for predicting power demand by constructing an ensemble having a parallel structure of CNN-LSTM and CNN-GRU is proposed. The present invention is an algorithm that predicts the final power demand by adopting an algorithm having higher accuracy for each time and section by applying an algorithm that distinguishes time and section.

Power demand prediction for each section is performed through LSTM/GRU, and the predicted value calculated through a technique with higher accuracy for each section is selected. In addition, the values predicted N days ago can be analyzed for errors through the MAPE and RMSE functions of Equations 13 and 14 below.

[Equation 13]

[Equation 14]

RMSE is an abbreviation for root mean square error, which expresses the precision of measured values by squaring the difference between the actual value and the predicted value and taking the square root again. MAPE expresses the precision of the measured value as a percentage value between 0 and 100% by adding all the ratios of relative error to the actual value and then dividing by the number of data.

The present invention can predict the final power demand by adopting an algorithm having higher accuracy for each time and section through the application of an algorithm that distinguishes time and section.

Accordingly, it can contribute to economic benefits and energy efficiency improvement through the development of an AI-based power demand prediction algorithm and optimal operation scheduling for each customer. In addition, it is possible to improve the accuracy and objectivity of demand forecasting by utilizing key variables that affect electricity demand, such as meteorological factors, regional population density, and daily weather index.

11 is a flowchart of a power demand prediction method using CNN according to an embodiment of the present invention.

The power demand prediction method using CNN according to this embodiment may be performed in substantially the same configuration as the device 10 of FIG. 1 . Accordingly, components identical to those of the apparatus 10 of FIG. 1 are given the same reference numerals, and repeated descriptions are omitted.

In addition, the power demand prediction method using CNN according to this embodiment may be executed by software (application) for performing power demand prediction using CNN.

Referring to FIG. 11 , in the power demand prediction method using CNN according to the present embodiment, data of meteorological factors for power demand prediction is collected from API, which is online public data (step S10). In this case, a step of normalizing the data of the collected meteorological factors may be further included in order to reduce the scale difference of the data.

The data of the meteorological elements includes at least one of weather data according to time and interval, daily weather index data, and past power demand data.

Various factors can be used to predict electricity demand. For example, weather data (temperature, humidity, wind speed, cloudiness, etc.) There are weather index (discomfort index, perceived temperature, etc.), past power demand data, etc.

In one embodiment of the present invention, meteorological factors necessary for predicting power demand can be analyzed through a CNN algorithm. Data such as analyzed temperature, humidity, atmospheric pressure, and wind speed can be collected through APIs such as the meteorological data open portal of the Korea Meteorological Administration. For the flexibility and sensitivity of algorithm execution, the data collection period can be set for a long period of time, such as the last 1 year or 3 years.

Missing data of the collected data is complemented (step S20).

Domestic power demand, meteorological factors, and power demand data by site may have missing data or abnormal data when actually stored in the database. Such data adversely affects the accuracy of electricity demand forecasting, and is particularly applied as a factor that increases inaccuracy in deep learning-based algorithms such as LSTM. Therefore, additional work is required to handle missing values.

In the step of compensating for missing data of the collected data, the presence or absence of missing data is determined through an overlap ratio of data patterns. As a result, when the number of missing data is 2 or less, the missing data can be supplemented using mean interpolation, and when the number of missing data is 3 or more, the missing data can be supplemented through the ARIMA time series prediction model.

Correlation between predictors is analyzed through CNN, which is a convolutional neural network that takes the supplemented data as input, and predictors to be reflected in the prediction of electricity demand are selected for each section (step S30). In this case, a step of classifying and resampling the data to integrate the sampling rate of the selected predictor may be further included.

Correlation coefficient analysis for each predictor is automatically selected based on high correlation coefficient within the power demand prediction algorithm, and can be implemented to select the optimal input variable whenever data is updated.

Through the parallel structure of CNN-LSTM and CNN-GRU, the final power demand with higher accuracy is selected for each section (step S40). In addition, it is possible to analyze the error of the power demand values predicted N days ago through the MAPE and RMSE functions.

In the present invention, the algorithm for predicting actual power demand is divided into two types. First, LSTM is a deep learning algorithm that has high performance in predicting time series data. Second, GRU is a simplified version of LSTM, a deep learning algorithm that reduces three gates to two. Compared to GRU, LSTM tends to have better prediction performance than GRU as the amount of computation is larger and the prediction data size is larger.

In the present invention, an algorithm for predicting power demand by constructing an ensemble having a parallel structure of CNN-LSTM and CNN-GRU is proposed. The present invention is an algorithm that predicts the final power demand by adopting an algorithm having higher accuracy for each time and section by applying an algorithm that distinguishes time and section.

Therefore, it is possible to predict the final power demand by adopting an algorithm having higher accuracy for each time and section by applying an algorithm that distinguishes time and section.

Such a power demand prediction method using CNN may be implemented as an application or implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer readable recording medium may include program instructions, data files, data structures, etc. alone or in combination.

Program instructions recorded on the computer-readable recording medium may be those specially designed and configured for the present invention, or those known and usable to those skilled in the art of computer software.

Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. media), and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like.

Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter or the like as well as machine language codes such as those produced by a compiler. The hardware device may be configured to act as one or more software modules to perform processing according to the present invention and vice versa.

Although the above has been described with reference to embodiments, those skilled in the art can variously modify and change the present invention without departing from the spirit and scope of the present invention described in the claims below. You will understand.

The present invention can be used for the development of a power demand prediction algorithm and correlation analysis by domestic and customer. Therefore, it can be used as a useful reference for energy consumption behavior analysis, which should be preceded in order to provide energy efficiency services for each customer in the future, and it is expected that an optimal scheduling technique suitable for the type of customer can be presented through the present invention.

10: Power demand prediction device using CNN
110: data collection unit
130: data supplementation unit
150: predictor selection unit
170: power demand prediction unit

Claims (10)

Collecting data of meteorological factors for electricity demand prediction from API, which is online public data;
Compensating for missing data of the collected data;
Selecting a predictor to be reflected in the prediction of power demand for each section by analyzing the correlation between predictors through CNN, which is a convolutional neural network that takes supplemented data as input; and
Through the parallel structure of CNN-LSTM and CNN-GRU, power demand prediction is performed for each section, and the predicted value calculated through a technique with higher accuracy for each section is selected among the CNN-LSTM and CNN-GRU, and the final Including; selecting the power demand;
In the step of selecting a predictor to be reflected in the electricity demand forecast for each section,
An electric power demand forecasting method using CNN, which image-processes the electric power demand pattern and the predictor pattern for each section, selects factors to be reflected in electric power demand forecasting, and uses them for electric power demand forecasting.
According to claim 1,
The data of the meteorological elements includes at least one of weather data according to time and interval, daily weather index data, and past power demand data, power demand prediction method using CNN.
The method of claim 1, wherein the step of collecting data of the meteorological factors,
A power demand prediction method using a CNN, comprising: normalizing data of collected meteorological factors in order to reduce the scale difference of the data.
The method of claim 1, wherein the supplementing of missing data of the collected data comprises:
determining whether there is missing data through an overlap ratio of data patterns; and
Compensating for missing data using mean interpolation when the number of missing data is 2 or less, and supplementing for missing data through an ARIMA time series prediction model when the number of missing data is 3 or more; including, CNN. Electric power demand forecasting method using .
delete According to claim 1,
Analyzing the error of the power demand values predicted N days ago through the MAPE and RMSE functions; further comprising a power demand prediction method using CNN.
According to claim 1,
Classifying and resampling the data to integrate the sampling rate of the selected predictor; further comprising a power demand prediction method using CNN.
A computer-readable storage medium on which a computer program for performing the method of predicting power demand using the CNN according to claim 1 is recorded.
A data collection unit that collects data of meteorological factors for predicting power demand from API, which is online public data;
a data complementation unit that compensates for missing data of the collected data;
a predictor selection unit that analyzes the correlation between predictors through CNN, a convolutional neural network that takes the supplemented data as an input, and selects a predictor to be reflected in the electricity demand forecast for each section; and
Through the parallel structure of CNN-LSTM and CNN-GRU, power demand prediction is performed for each section, and the predicted value calculated through a technique with higher accuracy for each section is selected among the CNN-LSTM and CNN-GRU, and the final Including; power demand forecasting unit for selecting power demand;
The predictor selection unit,
An apparatus for predicting power demand using CNN, which selects factors to be reflected in prediction of power demand by image-processing the pattern of power demand and the pattern of predictors for each section to be used in prediction of power demand.
10. The method of claim 9, wherein the data complementation unit,
The presence or absence of missing data is determined through the overlap ratio of data patterns, and if the number of missing data is 2 or less, the missing data is complemented using mean interpolation, and if the number of missing data is 3 or more, ARIMA time series prediction model A device for predicting power demand using CNN, supplementing missing data through

Images