KR102324783B1 - Apparatus and method for Deep neural network based power demand prediction - Google Patents

Abstract

A deep neural network-based power demand prediction device and method for predicting power demand by performing preprocessing on missing data, configuring an independent prediction model for each data for prediction, and constructing a prediction model that merges independent prediction models it's about
A deep neural network-based power demand prediction apparatus according to an embodiment of the present invention includes a data pre-processing unit for pre-processing data for predicting power demand, and an independent first prediction model for each data using the data pre-processed in the data pre-processing unit. A predictive model generation unit generating a second prediction model by generating a plurality of first prediction models by merging the plurality of first prediction models, and predicting the amount of power demand by using the first prediction model and the second prediction model generated by the prediction model generation unit Prediction evaluating the validity of the first prediction model and the second prediction model based on the prediction accuracy according to the power demand prediction unit and the result of preprocessing the data in the data preprocessing unit and the power demand prediction unit in the power demand prediction unit It includes an accuracy analysis unit.

Description

Apparatus and method for Deep neural network based power demand prediction

The present invention relates to an apparatus and method for predicting power demand based on a deep neural network, and in more detail, a prediction model that performs preprocessing on missing data, configures an independent prediction model for each data for prediction, and merges independent prediction models It relates to a deep neural network-based power demand prediction apparatus and method for predicting power demand by configuring

In order to operate the power system stably, it is important to predict the power demand. Furthermore, in recent years, as the electric power environment rapidly changes, such as the increase in renewable power sources, the supply and self-consumption of home photovoltaic power generation facilities, and the expansion of electric vehicles, it is becoming more important to accurately predict the electric power demand.

However, the existing power demand forecasting method applies statistical methods based on past load performance and weather information. This power demand forecasting technique is a statistical technique, and there is a problem in that it is difficult to use various types of data to be used for prediction and that large-capacity data cannot be reflected in the prediction. In addition, there is a problem in that the final prediction result is greatly changed according to the correction of the prediction result by the expert's experience.

The present invention is to solve the problems described above, by performing pre-processing on missing data, configuring an independent prediction model for each data for prediction, and configuring a prediction model merging the independent prediction models to predict the power demand An object of the present invention is to provide a deep neural network-based power demand forecasting apparatus and method for

In addition to the technical problems of the present invention mentioned above, other features and advantages of the present invention will be described below or will be clearly understood by those skilled in the art from such description and description.

A deep neural network-based power demand prediction apparatus according to an embodiment of the present invention for achieving the above-described object includes a data pre-processing unit for pre-processing data for predicting power demand, and data pre-processed by the data pre-processing unit for each data Using a predictive model generator that generates an independent first predictive model and generates a second predictive model by merging a plurality of first predictive models, and the first and second predictive models generated by the predictive model generator Validity of the first prediction model and the second prediction model based on the prediction accuracy according to the result of preprocessing the data by the power demand prediction unit for predicting the power demand and the data preprocessing unit and the power demand prediction unit for the power demand prediction unit It may include a prediction accuracy analysis unit to evaluate the.

On the other hand, the deep neural network-based power demand prediction method according to an embodiment of the present invention for achieving the above-described object includes the steps of preprocessing data for predicting the amount of power demand, and using the preprocessed data to make independent first prediction for each data Generating a model, generating a second prediction model by merging a plurality of first prediction models, predicting the amount of power demand by using the generated first prediction model and the second prediction model, and preprocessing the data It may include evaluating the effectiveness of the first prediction model and the second prediction model based on the prediction accuracy according to the result and the result of predicting the power demand.

A deep neural network-based power demand prediction apparatus and method according to an embodiment of the present invention automatically generates a prediction model for predicting power demand and periodically evaluates the effectiveness of the prediction model to provide a prediction model reflecting the latest data at all times. can

In addition, it is possible to reduce the burden of the existing work that relied on the experience of the person in charge of demand forecasting by predicting the power demand through the predictive model.

In addition, as a prediction model is generated for each data for prediction, it is possible to predict the power demand by selecting a prediction target according to the situation.

In addition, other features and advantages of the present invention may be newly recognized through embodiments of the present invention.

1 is a diagram illustrating a system for predicting power demand according to an embodiment of the present invention.
2 is a diagram illustrating an apparatus for predicting power demand according to an embodiment of the present invention.
3 is a diagram illustrating an example of data pre-processed by a data pre-processing unit according to an embodiment of the present invention.
4 is a diagram illustrating generation of a prediction model according to an embodiment of the present invention.
5 is a diagram illustrating a method for predicting power demand based on a deep neural network according to an embodiment of the present invention.
6 is a diagram illustrating a method of determining validity of a first prediction model and a second prediction model according to an embodiment of the present invention.

Although not defined otherwise, all terms including technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the present invention belongs. Commonly used terms defined in the dictionary are additionally interpreted as having a meaning consistent with the related technical literature and the presently disclosed content, and unless defined, they are not interpreted in an ideal or very formal meaning.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement them. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein.

1 is a diagram illustrating a system for predicting power demand according to an embodiment of the present invention.

Referring to FIG. 1 , a power demand prediction system 1000 according to an embodiment of the present invention may include a power demand device 100 , a power system 200 , and a consumer 300 .

The power system 200 may provide power to the consumer 300 . Here, the power system 200 may provide power to the consumers 300 according to the demand for power of the consumers 300 . Accordingly, the power demand prediction apparatus 100 may predict the power demand amount of the consumer 300 so that the power system 200 can supply power based on the power demand amount.

The power demand prediction apparatus 100 according to an embodiment of the present invention may predict the power demand by using a prediction model through a deep neural network. The power demand prediction apparatus 100 may pre-process the missing data and generate a prediction model using the pre-processed missing data. This is because when forecasting the power demand with weather data missing, the weather data has a continuous nature with the passage of time due to the nature of the weather data, so if the weather data of a nearby date is not utilized, the prediction accuracy may decrease. . In order to solve this problem, the power demand prediction apparatus 100 according to an embodiment of the present invention may pre-process and use the missing data so that the data for predicting the power demand is not omitted.

In addition, the power demand prediction apparatus 100 may predict the amount of power demand by using a plurality of prediction data. In this case, since the data ranges of the plurality of data are all different, the plurality of data are normalized to have the same data range. can do. Due to this, a prediction model in which the influence on the plurality of data is evenly reflected may be generated, and prediction accuracy may be increased as the power demand is predicted through the generated prediction model.

2 is a diagram illustrating an apparatus for predicting power demand according to an embodiment of the present invention, and FIG. 3 is a diagram illustrating an example of data preprocessed by the data preprocessor according to an embodiment of the present invention.

2 and 3 , the power demand prediction apparatus 100 according to an embodiment of the present invention includes a data preprocessing unit 110 , a prediction model generation unit 120 , a power demand prediction unit 130 , and prediction accuracy analysis. It may include a unit 140 .

The data preprocessor 110 may preprocess data for predicting the amount of power demand. Data for predicting power demand are weather observation (D1), weather forecast (D2), transmission load (D3), AMI (Advanced Metering Infrastructure) power usage (D4), power generation end load (D5), and social statistical indicators (D6) ) may be included.

The weather observation data (D1) may be minute-by-minute measurement data provided by the Korea Meteorological Administration (AWS), and the minute-by-minute measurement data includes temperature, humidity, wind speed, hourly cumulative precipitation, daily cumulative precipitation, and daily maximum It may include wind speed.

The weather forecast data D2 may be neighborhood weather forecast data in units of three hours.

Transmission load data (D3) is collected by SCADA (Supervisory Control And Data Acquisition) system in 5-minute increments for each 15 regional headquarters, and load analysis in 5 minutes by regional headquarters provided by SOMAS (Substation Operating Results Management System). Load data may be included. For example, the SCADA system may be a system for monitoring control and data acquisition operated by Korea Electric Power Corporation, and the SOMAS may be a system for managing substation operation performance.

The AMI power usage data D4 may be 15-minute power usage data of high voltage customers nationwide.

The power generation stage load data D5 may be current load data in units of 5 minutes provided by the Power Exchange.

The social statistical index D6 may include a perceived temperature, an unpleasant index, air pollution information, and consumption psychology statistics.

In addition, the data preprocessor 110 may correct the missing data to more accurately predict the amount of power demand. Specifically, the data preprocessor 110 may divide the weather observation data D1 and the weather forecast data D2 at regular time intervals. For example, the predetermined time interval may be 1 hour, and the data preprocessor 110 may select a representative value from among data divided at a predetermined time interval. That is, the data preprocessor 110 may select a representative value from among data included in the first time interval, and select a representative value from among data included in a second time interval that is a time interval following the first time interval. . The data preprocessor 110 may define the selected representative value as a value of missing data. In this case, the data preprocessor 110 may exclude data having an outlier or a missing value among the weather observation data D1 and the weather forecast data D2 .

When the weather observation data D1 is data related to temperature, humidity, and wind speed, the data preprocessor 110 may calculate an average value of the data for each temperature, humidity, and wind speed, and select the calculated average value as a representative value. That is, the data preprocessor 110 may calculate an average value of data included in each time interval and select it as a representative value. For example, since the temperature does not change significantly in a short time, the data pre-processing unit 110 calculates the average value of normal data for a specific time period and selects it as a representative value, so that the accuracy of the power demand prediction is reduced due to missing data. can be prevented The data preprocessor 110 may define the selected representative value as a value of missing data.

In addition, when the weather observation data D1 is data related to time accumulated precipitation amount, daily accumulated precipitation amount, and one instantaneous maximum wind speed, the data preprocessor 110 may select data having a maximum value as a representative value. Here, that is, the data preprocessor 110 may select data having a maximum value among data included in each time interval as a representative value. The data preprocessor 110 may define the selected representative value as a value of missing data.

In addition, the data preprocessor 110 may select the representative value by averaging the weather forecast data D2 divided by the time interval of 3 hours with the average of the previous day and the value predicted immediately before or recently. The data preprocessor 110 may define the selected representative value as a value of missing data.

That is, according to an embodiment of the present invention, by correcting the data by calculating or selecting values of the missing data, it is possible to prevent the accuracy of the prediction of the power demand amount from being reduced due to the missing data.

The data preprocessor 110 may calculate a national weather weighted average by weighted averaging the weather observation and weather forecast data of the nearest point through the power usage of the AMI customer, and the calculated national weather weighted average is a prediction for power demand prediction It can be used to create models. The data preprocessor 110 may calculate a national weather weighted average through Equation (1).

[Equation 1]

은 전국 기상 가중 평균이고,

은 AMI 전체 고객의 전력 사용량이고,

는 AMI 고객 각각의 전력 사용량이고,

는 AMI 고객 각각의 기상 관측 데이터(D1) 및 기상 예보 데이터(D2)일 수 있다.here,

is the national weather weighted average,

is the power consumption of all AMI customers,

is the power consumption of each AMI customer,

may be weather observation data D1 and weather forecast data D2 of each AMI customer.

The data preprocessor 110 may generate data in units of one hour by adding up the data in units of 5 minutes of the SCADA system. Here, the power demand prediction apparatus 100 according to an embodiment of the present invention predicts the amount of power demand by using the data of the SCADA system, but when omission occurs in the data of the SCADA system, the data of the SOMAS system can be used for correction. . In this case, the data preprocessor 110 may correct the data of the same section among the data of the SOMAS system based on the maximum and minimum values among the data of the SCADA system for each time period. Also, the data preprocessor 110 may normalize the data of the SOMAS system to be a value between 0 and 1, and project and simulate the minimum and maximum values among the data of the SCADA system.

The data preprocessor 110 may preprocess the AMI power usage data by classifying the AMI customers by region, and calculating the summation and average value of the power usage data classified by region by industry and contract type.

In addition, the data pre-processing unit 110 may pre-process by changing daily, monthly, and yearly statistics in the social statistics index data D6 into hourly data. That is, the data pre-processing unit 110 may allow the corresponding value to be applied equally at all times in the case of weather-related information such as air pollution information in the social statistics indicator data D6. In addition, the data pre-processing unit 110 assumes that the previous statistical value is the lowest value and the current statistical value is the maximum value in the case of other information except for weather-related information in the social statistical indicator data D6, and the intermediate time unit data can be simulated. Here, the data preprocessor 110 may simulate the increase or decrease of data in the form of a linear, normal distribution, exponential distribution, or log normal distribution according to the characteristics of the data.

The prediction model generator 120 may generate a first prediction model and a second prediction model. The first prediction model is a model for predicting the maximum power demand of the next day, and may have one output value. That is, the first prediction model may output only one maximum power demand of the next day as an output value.

In addition, the second prediction model is a model for predicting the maximum power demand for each time zone of the next day, and may have a plurality of output values. That is, since the second prediction model outputs the maximum power demand for each time period from 12:00 pm on the same day to 24:00 the next day as an output value, the number of output values may be 36.

Specifically, the prediction model generator 120 may generate an independent first prediction model for each data for prediction. That is, since the prediction model generator 120 generates an independent first prediction model for each data for prediction, a plurality of first prediction models may be generated. Here, the first prediction model may be composed of a long-short-term memory deep neural network.

Also, the predictive model generator 120 may generate a second predictive model by merging a plurality of first predictive models. Here, the second prediction model may be configured as a dense deep neural network. The prediction model generator 120 may normalize the data for prediction and convert it into a value having a distribution of 0 to 1.

Also, the predictive model generator 120 may learn the first predictive model and the second predictive model by using other data except for data of the last month among the data preprocessed by the data preprocessor 110 . Also, the predictive model generator 120 may verify the first predictive model and the second predictive model by using data of the last month among the data preprocessed by the data preprocessor 110 .

Here, the predictive model generator 120 may train the first predictive model and the second predictive model in the order of the past data to the most recent data according to the order of time on the data preprocessed by the data preprocessor 110 . The predictive model generator 120 may train the first predictive model and the second predictive model several times by using the preprocessed data.

In addition, the predictive model generator 120 compares the error with the verification data whenever the first predictive model and the second predictive model are trained using the preprocessed data, and ends the learning when the error becomes smaller than a specific value. can Here, the verification data may be data of the last month among preprocessed data.

The power demand prediction unit 130 may predict the amount of power demand by using the prediction model generated every predetermined time by the prediction model generation unit 120 . Here, the power demand prediction unit 130 reflects the result of predicting the maximum power demand for the next day using the first prediction model to the result of predicting the maximum power demand for each time period using the second prediction model, the value of the maximum power demand can be corrected. That is, the power demand prediction unit 130 may correct the maximum power demand by projecting the maximum power demand for each time period to become the maximum power demand for the next day.

The prediction accuracy analysis unit 140 performs the first prediction model and the second prediction based on the prediction accuracy according to the result of preprocessing the data by the data preprocessing unit 110 and the result of predicting the power demand by the power demand prediction unit 130 . The validity of the model can be evaluated. As a result of evaluating the validity of the first prediction model and the second prediction model by the prediction accuracy analysis unit 140, when the validity of the first prediction model and the second prediction model do not meet the criteria, the prediction model generation unit 120 returns the second prediction model. A signal may be transmitted to regenerate the first predictive model and the second predictive model. Here, the criterion for evaluating the validity may be preset, and when the prediction accuracy of the first prediction model and the second prediction model becomes smaller than the preset criterion, the prediction accuracy analyzer 140 performs the first prediction model and the second prediction model. 2The validity of the predictive model can be evaluated. Also, the prediction accuracy analyzer 140 may determine the validity of the first prediction model and the second prediction model by setting weights for each data for prediction.

For example, the prediction accuracy analyzer 140 may determine the validity of the first prediction model and the second prediction model when the recent rate of change of the weather observation and weather forecast data preprocessed by the data preprocessor 110 exceeds a threshold. Also, when the recent rate of change changes from negative to positive or from positive to negative, the prediction accuracy analyzer 140 may determine the validity of the first prediction model and the second prediction model. Also, when the average of the prediction accuracy of the first prediction model and the second prediction model decreases for a specific day continuously, the prediction accuracy analyzer 140 determines that the time when the first prediction model and the second prediction model were last generated is one week before. In the case of , the validity of the first prediction model and the second prediction model may be determined.

In addition, the power demand prediction apparatus 100 may predict the power demand for a specific region, and according to an embodiment of the present invention, predicting the power demand nationwide is to predict the power demand for a specific region. It can be implemented by changing

For example, in the case of a power plant load, the power demand can be predicted using the load information of a specific area provided by the SCADA system. Power demand can be predicted by weighted average of power usage.

4 is a diagram illustrating generation of a prediction model according to an embodiment of the present invention.

Referring to FIG. 4 , the predictive model generator 120 may generate the first predictive model 122 for each of a plurality of data by using the data preprocessed by the data preprocessor 110 . The first prediction model is a model for predicting the maximum power demand for the next day, and may be composed of a long-short term memory deep neural network.

That is, the prediction model generator 120 may generate the first prediction model 122 by using the weather observation data D1 among the data preprocessed by the data preprocessor 110 . Here, the first prediction model 122 may predict the maximum power demand of the next day based on the weather observation data D1 , and may output the predicted maximum power demand of the next day as an output value.

Also, the predictive model generator 120 may generate the first predictive model 122 by using the weather forecast data D2 among the data preprocessed by the data preprocessor 110 . Here, the first prediction model 122 may predict the maximum power demand of the next day based on the weather forecast data D2 and may output the predicted maximum power demand of the next day as an output value.

Also, the predictive model generator 120 may generate the first predictive model 122 by using the power transmission load data D3 among the data preprocessed by the data preprocessor 110 . Here, the first prediction model 122 may predict the maximum power demand of the next day based on the power transmission load data D3 and may output the predicted maximum power demand of the next day as an output value.

Also, the predictive model generator 120 may generate the first predictive model 122 by using the AMI power usage data D4 among the data preprocessed by the data preprocessor 110 . Here, the first prediction model 122 may predict the maximum power demand of the next day based on the AMI power usage data D4, and may output the predicted maximum power demand of the next day as an output value.

Also, the predictive model generator 120 may generate the first predictive model 122 by using the power generation stage load data D5 among the data preprocessed by the data preprocessor 110 . Here, the first prediction model 122 may predict the maximum power demand of the next day based on the power generation stage load data D5, and may output the predicted maximum power demand of the next day as an output value.

Also, the predictive model generator 120 may generate the first predictive model 122 by using the social statistics index data D6 among the data preprocessed by the data preprocessor 110 . Here, the first prediction model 122 may predict the maximum power demand of the next day based on the social statistics index data D6, and may output the predicted maximum power demand of the next day as an output value.

Also, the prediction model generator 120 may generate the second prediction model 124 by merging the plurality of first prediction models 122 . The second prediction model 124 is a model for predicting the maximum power demand for each time zone of the next day, and may be configured as a dense deep neural network.

The prediction model generator 120 may predict the maximum power demand for each time period from 12:00 pm on the same day to 24 pm the next day using the second prediction model 124, and output the predicted maximum power demand for each time period as an output value. have.

5 is a diagram illustrating a deep neural network-based power demand prediction method according to an embodiment of the present invention.

Referring to FIG. 5 , the data pre-processing unit 110 may pre-process data for predicting the amount of power demand ( S100 ). Data for predicting power demand may include weather observation, weather forecast, transmission load, AMI power usage, power generation end load, and social statistical indicators.

The weather observation data may be 1-minute measurement data provided by the Korea Meteorological Administration AWS, and the 1-minute measurement data may include temperature, humidity, wind speed, hourly cumulative precipitation, daily cumulative precipitation, and daily maximum wind speed.

The weather forecast data may be neighborhood weather forecast data in units of three hours.

The transmission load data (D3) may include the load data for each 15 regional headquarters collected by the SCADA system, and the load data for each regional headquarters in 5-minute units provided by load analysis and SOMAS.

The AMI power usage data may be power usage data in 15-minute units of high-voltage customers nationwide, and the power generation end load data may be current load data in units of 5 minutes provided by the Power Exchange. Social statistical indicators may include perceived temperature, discomfort index, air pollution information, and consumer psychology statistics.

The prediction model generator 120 may generate a first prediction model for each data for prediction using the pre-processed data ( S200 ). The first prediction model is a model for predicting the maximum power demand for the next day, and may be composed of a deep neural network of Long-Short Term Memory.

Also, the predictive model generator 120 may generate a second-side model by merging the first predictive models ( S300 ). That is, the prediction model generator 120 may generate a second prediction model by merging the plurality of first prediction models generated for each data for prediction. The second prediction model is a model for predicting the maximum power demand for each time period of the next day, and may be composed of a dense deep neural network.

The power demand prediction unit 130 may predict the power demand by using the first prediction model and the second prediction model ( S400 ). Here, the power demand prediction unit 130 reflects the result of predicting the maximum power demand for the next day using the first prediction model to the result of predicting the maximum power demand for each time period using the second prediction model, the value of the maximum power demand can be corrected.

The prediction accuracy analyzer 140 may determine the validity of the first prediction model and the second prediction model ( S500 ). The prediction accuracy analysis unit 140 performs the first prediction model and the second prediction based on the prediction accuracy according to the result of preprocessing the data by the data preprocessing unit 110 and the result of predicting the power demand by the power demand prediction unit 130 . The validity of the model can be evaluated.

As a result of evaluating the validity of the first prediction model and the second prediction model by the prediction accuracy analysis unit 140, when the validity of the first prediction model and the second prediction model do not meet the criteria, the prediction model generation unit 120 returns the second prediction model. The first predictive model and the second predictive model may be regenerated.

6 is a diagram illustrating a method of determining the validity of a first prediction model and a second prediction model according to an embodiment of the present invention.

Referring to FIG. 6 , the prediction accuracy analyzer 140 may determine whether the time when the first prediction model and the second prediction model were last generated exceeds 7 days ( S510 ).

When it is determined that the time at which the first prediction model and the second prediction model were last generated exceeds 7 days, the prediction accuracy analysis unit 140 uses the prediction model generation unit 120 to generate the first prediction model and the second prediction model. A predictive model may be newly generated (S550).

On the other hand, when it is determined that the time when the first prediction model and the second prediction model are last generated does not exceed 7 days, the prediction accuracy analyzer 140 determines whether the rate of change of the weather observation and weather forecast data exceeds a threshold value. It can be determined (S520).

When it is determined that the rate of change of the weather observation and weather forecast data exceeds a threshold, the prediction accuracy analysis unit 140 generates a first prediction model and a second prediction model newly through the prediction model generation unit 120 . can be (S550).

On the other hand, when it is determined that the rate of change of the weather observation and weather forecast data does not exceed the threshold, the prediction accuracy analysis unit 140 determines whether the rate of change of the weather observation and weather forecast data is changed from negative to positive or from positive to negative. It can be (S530).

When it is determined that the rate of change of weather observation and weather forecast data is changed from negative to positive or from positive to negative, the prediction accuracy analysis unit 140 generates the first prediction model and the second prediction model through the prediction model generation unit 120 . It can be newly created (S550).

On the other hand, when it is determined that the rate of change of the weather observation and weather forecast data does not change from negative to positive or from positive to negative, the prediction accuracy analysis unit 140 determines whether the number of days for which the average of the prediction accuracy continuously falls exceeds a specific number of days. It can be determined (S540).

When it is determined that the number of days for which the average of the prediction accuracy has continuously decreased exceeds a specific number of days, the prediction accuracy analysis unit 140 newly generates the first prediction model and the second prediction model through the prediction model generation unit 120 . It can be done (S550).

Meanwhile, the prediction accuracy analysis unit 140 may determine that the validity of the first model and the second model are appropriate when it is determined that the number of days for which the average of the prediction accuracy continuously decreases does not exceed a specific number of days.

As described above, according to an embodiment of the present invention, the power demand is predicted by preprocessing the missing data, configuring an independent prediction model for each data for prediction, and configuring a prediction model merging the independent prediction models. It is possible to realize a deep neural network-based power demand forecasting apparatus and method for

Those skilled in the art to which the present invention pertains should understand that the present invention may 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. .

100: power demand forecasting device
200: power system
300: acceptor
110: data preprocessor
120: predictive model generation unit
130: power demand forecasting unit
140: predictive accuracy analysis unit
122: first prediction model
124: second prediction model

Claims (18)

a data pre-processing unit for pre-processing data for predicting power demand;
a predictive model generator for generating a plurality of independent first predictive models for each data type using the data preprocessed by the data preprocessor, and merging the plurality of first predictive models to generate a second predictive model;
a power demand prediction unit for predicting an amount of power demand by using the plurality of first prediction models and the second prediction model generated by the prediction model generation unit;
Prediction accuracy analysis for evaluating the validity of the plurality of first prediction models and the second prediction models based on the prediction accuracy according to the result of preprocessing the data by the data preprocessing unit and the result of predicting the power demand by the power demand prediction unit including wealth,
The prediction accuracy analyzer may change the recent rate of change of weather observation and weather forecast data pre-processed by the data preprocessing unit to a threshold value, or change from negative to positive or positive to negative, or change the rate of change of weather observation and weather forecast data from negative to negative, or At least one of a case in which the average of the prediction accuracy of the first predictive model and the second predictive model decreases continuously for a specific day, or when the plurality of first predictive models and the second predictive model are generated exceeds one week In the case where the validity of the plurality of first prediction models and the second prediction model is determined, and as a result, when the validity of the plurality of first prediction models and the second prediction model does not meet the criteria, in the prediction model generation unit A deep neural network-based power demand prediction device that transmits a signal to regenerate a plurality of first and second prediction models.
According to claim 1,
The data for predicting the power demand is a deep neural network-based power demand prediction device including weather observation data, weather forecast data, transmission load data, AMI (Advanced Metering Infrastructure) power usage data, power generation end load data, and social statistical indicators.
According to claim 1,
The plurality of first prediction models are models for predicting the maximum power demand for the next day, and a deep neural network-based power demand prediction device comprising a long-short term memory deep neural network.
According to claim 1,
The second prediction model is a model for predicting the maximum amount of power demand for each time period of the next day, and a deep neural network-based power demand prediction device comprising a dense deep neural network.
According to claim 1,
The prediction model generator normalizes the data for predicting the power demand and converts it into a value having a distribution from 0 to 1 to generate the plurality of first prediction models independent for each data type. A deep neural network-based power demand prediction device .
According to claim 1,
The predictive model generator learns the plurality of first predictive models and the second predictive models by using other data except for data of the last month among the data preprocessed by the data preprocessor, and is preprocessed by the data preprocessor. A deep neural network-based power demand prediction apparatus for verifying the plurality of first prediction models and the second prediction models using data of the last month among data.
According to claim 1,
The power demand prediction unit reflects the result of predicting the maximum power demand for the next day using the plurality of first prediction models to the result of predicting the maximum power demand for each time period using the second prediction model, the value of the maximum power demand A deep neural network-based power demand forecasting device that calibrates.
delete delete In the deep neural network-based power demand prediction method by the deep neural network-based power demand prediction device,
pre-processing data for predicting power demand;
generating a plurality of independent first prediction models for each data type using the preprocessed data, and generating a second prediction model by merging the plurality of first prediction models;
predicting an amount of power demand by using the plurality of generated first prediction models and the second prediction models;
Evaluating the validity of the plurality of first prediction models and the second prediction model based on the prediction accuracy according to the result of pre-processing the data and the prediction of the power demand,
The step of evaluating the effectiveness,
The recent rate of change of the preprocessed weather observation and weather forecast data exceeds a threshold, or the recent rate of change of the weather observation and weather forecast data changes from negative to positive or positive to negative, or the plurality of first prediction models and the second prediction When it includes at least one of a case in which the average of the prediction accuracy of the model decreases for a specific day continuously, or when the time at which the plurality of first prediction models and the second prediction model are generated exceeds one week, the plurality of first prediction models and determining the validity of the second prediction model. and
generating the second predictive model when the validity of the plurality of first predictive models and the second predictive model is evaluated and the validity of the plurality of first predictive models and the second predictive model do not meet a criterion; A method for predicting power demand based on a deep neural network, comprising: transmitting a signal to regenerate the plurality of first predictive models and second predictive models.
11. The method of claim 10,
The data for predicting the power demand is a deep neural network-based power demand prediction method including weather observation data, weather forecast data, transmission load data, AMI (Advanced Metering Infrastructure) power usage data, power generation end load data, and social statistical indicators.
11. The method of claim 10,
The plurality of first prediction models are models for predicting the maximum power demand for the next day, and a deep neural network-based power demand prediction method comprising a long-short term memory deep neural network.
11. The method of claim 10,
The second prediction model is a model for predicting the maximum power demand for each time period of the next day, and a deep neural network-based power demand prediction method comprising a dense deep neural network.
The method of claim 10, wherein the generating of the second predictive model comprises:
A deep neural network-based power demand prediction method for generating the plurality of independent first prediction models for each data type by normalizing the data for predicting the power demand and converting it into a value having a distribution from 0 to 1.
The method of claim 10, wherein the generating of the second predictive model comprises:
The plurality of first prediction models and the second prediction models are learned using data other than the data of the most recent month among the preprocessed data, and the plurality of first prediction models are learned using data of the most recent month among the preprocessed data. A deep neural network-based power demand prediction method for verifying a prediction model and the second prediction model.
The method of claim 10, wherein predicting the power demand comprises:
Deep neural network-based power to correct the value of the maximum power demand by reflecting the result of predicting the maximum power demand for the next day using the plurality of first prediction models to the result of predicting the maximum power demand for each time period using the second prediction model How to forecast demand.
delete delete

Images