KR20220107794A - Power information prediction method and system based on transfer learning - Google Patents

Abstract

A transfer learning-based power information prediction system of the present invention comprises: a data collection unit that collects past power data and relevant environmental information obtainable by region or group; a data preprocessing unit that preprocesses the data collected by the data collection unit; a correlation analysis unit that selects a regional dataset with the highest Pearson correlation coefficient based on a target dataset for a specific region and other datasets for different unspecified regions; a model pretraining unit that trains a learning model in advance using the regional dataset selected by the correlation analysis unit; a model fine-tuning unit that fine-tunes the learning model using the target dataset; and a prediction unit that predicts items of the target dataset in the future using the learning model. The present invention aims to provide a transfer learning-based power demand prediction method and system capable of performing effective learning based on transfer learning when there is not sufficient leading data for learning.

Description

POWER INFORMATION PREDICTION METHOD AND SYSTEM BASED ON TRANSFER LEARNING

The present invention relates to a power demand prediction method for performing learning based on transfer learning using precedent data of other regions and/or groups of the same type in a state where precedent data for learning is insufficient It relates to a transfer learning-based power demand prediction method and system that can be implemented in the form of a learning-based monthly power demand prediction program.

Electric Load Forecasting is the basis for securing necessary power in advance for the efficient operation of the power system, and it ensures the reliability of power system equipment and maintains an active power reserve to prepare for losses due to power outages and overloads. used for

With the development of deep learning techniques, it is possible to create a predictive model with high accuracy, but there is a disadvantage that accuracy cannot be guaranteed if the model is constructed using a small amount of data.

For example, monthly electricity demand forecast for each autonomous district is essential for annual electricity scheduling. However, the monthly electricity demand data is insufficient because only one point is generated per month.

On the other hand, with the development of IOT technology and the information providing technology that supports it, various types of information can be collected from various places in the electric power field, and various application services can be developed and performed by each local government using electric power field data. However, whenever each application service is developed, the aforementioned lack of preceding data will arise.

Republic of Korea Publication No. 10-2019-0062741

An object of the present invention is to provide a transfer learning-based power demand prediction method and system that can perform effective learning based on transfer learning in a state where there is not enough preceding data for learning.

More specifically, the present invention uses Pearson correlation coefficient analysis to solve the problem of data shortage, finds a dataset divided into homogeneous regions and/or groups related to the prediction target dataset, pre-trains the model, and selects the target dataset. We present a method to improve the accuracy of the prediction model by fine-tuning the DNN model using

A transfer learning-based power information prediction system according to an aspect of the present invention includes: a data collecting unit for collecting past power data and related environment information obtainable by region or group; a data pre-processing unit for pre-processing the data collected by the data collecting unit; a correlation analysis unit that selects a regional dataset having the highest Pearson correlation coefficient based on a target dataset for a specified region and other unspecified regional datasets; a model pre-training unit for pre-learning a learning model using the regional dataset selected by the correlation analysis unit; a model fine-tuning unit for fine-tuning the learning model using the target dataset; and a predictor for predicting items of the target dataset in the future by using the learning model.

Here, the data collection unit may further include a storage unit for storing the collected data and information, and for storing the learning model.

Here, the correlation coefficient may be a Pearson correlation coefficient, and the learning model may be a DNN model.

Here, the correlation analysis unit, if the average value of the correlation coefficient is 0.5 or less, the top 5 high correlation coefficient, if the average value 0.5 or more and less than 0.7, the top 10, if the average value 0.7 or more and less than 0.8, the top 20, the average value 0.8 or more In this case, the top 30 regional datasets can be selected.

Here, the correlation analysis unit may perform correlation analysis according to the following equation.

및

는 각각 X 및 Y의 평균값임)(where R _XY is the PCC between X and Y, n is the number of observations, X _i and Y _i are the values of time i,

and

is the average value of X and Y, respectively)

Here, the data collection unit may include a data communication module for accessing an external agency server having source information of the data to be collected and accessing and receiving the source information.

According to another aspect of the present invention, a method for predicting power information based on transfer learning includes dividing power data by region or group; Collecting obtainable past power data and related environmental information; pre-processing the collected past power data and generating data sets for each region or group according to the division by region or group; designating a target dataset among the datasets for each region or group, and selecting a data set having a high correlation coefficient with the target dataset from among other unselected datasets for each region or group; pre-training a learning model using the selected target data set and a data set for each region or group having a high correlation coefficient; fine-tuning the learning model using the target dataset; and predicting items of the target dataset in the future with the learning model.

Here, in the step of predicting the items of the target dataset in the future, it is possible to collect and use prediction values of other organizations for future related environmental information.

Here, for all regional datasets, the target dataset may be designated, and the step of selecting or predicting the one having a high correlation coefficient may be performed.

A method for predicting power information based on transfer learning according to another aspect of the present invention includes: collecting historical power data and related environment information obtainable in a specific region or group; generating a target data set by pre-processing the collected past power data; selecting one having a high correlation coefficient with the target dataset from among previously acquired regional datasets of other regions; pre-training a learning model using the selected target dataset and regional datasets having a high correlation coefficient; fine tuning the learning model using the target dataset; and predicting items of the target dataset in the future with the learning model.

When the method and system for predicting power demand based on transfer learning according to the spirit of the present invention having the above configuration are implemented, there is an advantage in that effective learning can be performed based on transfer learning in a state where the preceding data for learning is insufficient.

1 is a block diagram illustrating a power demand prediction system according to the spirit of the present invention.
2 is a data configuration diagram showing the monthly electricity data collection structure of Seoul as the amount of electricity used by each autonomous district in Seoul.
3 is a conceptual diagram illustrating an example of calculating a Pearson correlation coefficient with respect to the monthly power data collection structure of Seoul of FIG. 2 .
4A is a conceptual diagram showing the overall structure of transfer learning using weights generated from a neural network trained and trained using a large-scale dataset similar to the target domain dataset.
Figure 4b is a conceptual diagram showing a general DNN structure of MTLF as a general deep neural network structure.
4C is a conceptual diagram illustrating an example of configuring a transfer learning-based DNN model.
Figure 4d is a conceptual diagram showing the overall data configuration and processes for monthly power demand forecasting.
5 is a flowchart illustrating an embodiment of a method for predicting power information according to the spirit of the present invention.
6 is a flowchart illustrating another embodiment of a method for predicting power information according to the spirit of the present invention.
7A and 7B are boxed graphs for MAPE and NRMSE of predictive models.
8 is a graph showing the relationship of mean absolute percentage error (MAPE) reduction versus Pearson's correlation coefficient.

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 terms used herein are used only to describe specific embodiments, and are 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 possibility of the presence 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.

Transfer learning is a task of pre-training a model using external data and then applying it to prediction, which has a great advantage in predicting monthly electricity demand with insufficient data. In the present invention, a Pearson correlation coefficient is used to select external data for prior learning.

1 is a block diagram illustrating a power demand prediction system according to the spirit of the present invention.

The illustrated power demand prediction system includes: a data collection unit 110 for collecting historical power data and related environmental information obtainable for each region or group; a data pre-processing unit 120 for pre-processing the data collected by the data collecting unit 110; a correlation analysis unit 140 for selecting a data set for each region or group with a high correlation coefficient based on the target dataset for a specified region or group and other unspecified region or group-specific datasets; a model pre-training unit 150 for pre-learning a learning model using the data set for each region or group selected by the correlation analysis unit 140; a model fine-tuning unit 160 for fine-tuning the learning model using the target dataset; and a prediction unit 170 for predicting items of the target dataset in the specified region in the future by using the learning model.

As shown, the data and information collected by the data collection unit 110 are stored, and the learning model used by the model pre-training unit 150 , the model fine-tuning unit 160 , and the prediction unit 170 is used. It can be defined as further comprising a storage unit 180 for storing.

The idea of the present invention proposes to secure data for transfer learning from other classification groups of the same kind when the preceding data for learning is insufficient.

The classification group may be determined by various criteria, and even if the classification criteria are changed, there is no difficulty in applying the spirit of the present invention as it is, which also falls within the scope of the present invention. However, for the convenience of explanation and understanding, if the classification group is divided according to the region (administrative region), which is the most widely used classification group, it is called region, and in the case of other classification groups, it is called group. In the following embodiment, the classification to which the autonomous district and industry category are applied is referred to as a region, and will be concretely illustrated as a case of being classified (divided) by region.

The learning model is exemplified by specifying the DNN model, and the advantages of selecting the DNN model will be described later.

In addition, the division by region will exemplify the case of classifying Seoul into autonomous districts, and the correlation coefficient will be exemplified by the Pearson correlation coefficient, which is easy to apply and has excellent correlation derivation.

The illustrated data collection unit 110 aims to collect data on weather, population, and power usage collected by region and date. As a source of data to be collected, a weather data open portal is an example of weather information. To this end, the data collection unit 110 may include a data communication module that accesses and receives the source information transmitted by accessing an external institution server having source information of the data to be collected.

The illustrated data collection unit 110 may collect monthly weather information as related environment information and collect past power usage data as power data. Among the collected data, weather, population, and electricity usage data a week ago can be used as input variables.

The illustrated data preprocessing unit 120 performs min-max scaling of the data collected by the data collection unit 110 and makes it into a form that can be used in the model (target data set form), The purpose is to send the scaled data to the correlation analysis unit.

The illustrated correlation analysis unit 140 may analyze elements having a similar monthly power usage pattern in terms of time, population, weather, and power one week ago, and as a result, select a regional dataset having a high correlation coefficient. In this case, the similarity of related environmental information may also be analyzed according to implementation.

The illustrated correlation analysis unit 140 calculates a Pearson correlation coefficient between the target dataset and a dataset of another autonomous region that may correspond thereto, and calculates the Pearson correlation coefficient of 10, 20, 30 (T10) having the highest Pearson correlation coefficient. , T20, and T30) for selecting regional datasets.

FIG. 2 is an example of electric power consumption by autonomous district of Seoul, and it can be seen that each district is divided into five categories for each use of electricity, such as "Household" and "Public", and collected/managed.

3 shows an example of calculating the Pearson correlation coefficient. In FIG. 3, T is the Total of FIG. 2, H is Household, P is Public, S is Service, and I is Industrial.

A high correlation coefficient means, for example, that the value of each cell in FIG. 3 is close to 1. Since high and low correlation coefficients are inevitably relative, the actual implementation of selecting a high correlation coefficient may be performed by selecting a predetermined number of higher values close to 1.

The regional datasets are selected in a predetermined number and used for pre-learning following the predictive learning model. If performed, the predictive learning model may rather become heterogeneous from the target domain in the pre-learning process.

According to the experiment described below, the following optimal relationship was derived with respect to the degree of correlation coefficient of all regional datasets with the target dataset and the number of selected regional datasets.

That is, if the average value of the correlation coefficient is 0.5 or less, the top 5 regions with high correlation coefficients are selected, if the average value is 0.5 or more and less than 0.7, the top 10 regions, if the average value is 0.7 or more and less than 0.8, the top 20 regions, and if the average value is 0.8 or more, the top 30 regional data Selecting sets and applying them to pre-learning results in optimal performance.

The illustrated model pre-training unit 150 pre-trains the DNN model as a learning model using the dataset of the autonomous region selected by the correlation analysis unit 140 and provides the pre-trained model to the model fine-tuning unit 160 . aim to do

When limited data is available in the target domain for learning-training, adaptive techniques such as transfer learning can be used to improve the performance of predictive models.

Figure 4a shows the general structure of transfer learning using weights generated in a neural network trained-trained using a large-scale dataset similar to the target domain dataset. The illustrated process is called pre-learning.

A deep neural network according to the DNN model will be described.

An artificial neural network is a machine learning algorithm that mimics the human brain and consists of three layers: an input layer, one or more hidden layers, and an output layer, and each layer consists of several nodes called perceptrons.

Each node takes a value from a node in the previous layer and determines the activation of the next node through a sigmoid function, a hyperbolic tangent function, a rectified linear unit (ReLU), an exponential linear unit (ELU), and a scaled node activation function. The process is repeated with exponential linear units (SELU), etc. and nodes are activated in the final layer (the output layer) to produce the desired output. If the number of hidden layers of an ANN is two or more, the network is called a DNN.

Figure 4b shows a typical DNN structure of MTLF as a general deep neural network structure for monthly electrical load prediction.

The more layers hidden in the DNN, the more complex the network, and increasing network complexity can improve network performance. However, if the structure of the DNN is more complex than necessary, the model may perform poorly on unseen data, which is called overfitting. Therefore, it is essential to select an appropriate number of hidden layers.

In order to construct a DNN-based MTLF model, various hyperparameters such as the number of hidden layers, the number of nodes in the hidden layer, and the activation function must be considered.

The illustrated model fine-tuning unit 160 aims to retrain the model using the target dataset of the DNN model pre-trained in the model pre-training unit 150 . 4C is a schematic diagram of a model pre-training unit and a model fine-tuning unit, and is a conceptual diagram illustrating an example of configuring a transfer learning-based DNN model. 4D is a conceptual diagram illustrating overall data configuration and processes for monthly power demand forecasting.

After performing the pre-training, the pre-trained learning model network is trained again using a dataset corresponding to a smaller target domain, a process called fine tuning.

If the data of the target domain and the source domain are similar, the resulting model can show satisfactory performance. Therefore, in order to find similar source domain data, a correlation analysis method such as Pearson Correlation Coefficient (PCC) analysis can be used to determine the similarity between two domains.

및

는 각각 X 및 Y의 평균값이다.The PCC analysis is defined by Equation 1 below, where R _XY represents the PCC between X and Y, n is the number of observations, and X _i and Y _i are the values of time i. In addition,

and

are the average values of X and Y, respectively.

[Equation 1]

PCC is used to measure the linear correlation between two variables X and Y (values range from +1 to 1). A value of +1 (a value of -1) indicates a total positive (negative) linear correlation between them, whereas a value of 0 indicates no linear correlation.

4c previously looked at shows an example of configuring a transfer learning-based DNN model. where X _i,S and X _i,T are the inputs of the source and destination domains at time i, respectively, and Y _i,S and Y _i,T are the outputs of the source and destination domains at time i, respectively.

When selecting a source domain, it is advantageous to select a domain with an absolute value of PCC close to 1 in order to consider all possible positive and negative trends.

In general, a dataset corresponding to a target domain for fine tuning becomes the target dataset.

The illustrated power prediction unit 170 aims to predict the actual power demand by using the retrained model in the model fine tuning unit 160 and derive the actual predicted power value through reverse normalization.

5 is a flowchart illustrating an embodiment of a method for predicting power information according to the spirit of the present invention.

The illustrated power information prediction method includes the steps of dividing power data into regions (S10); Collecting obtainable past power data and related environmental information (S110); pre-processing the collected past power data and generating regional datasets according to regional classification (S120); designating a target dataset among the regional datasets (S130), and selecting a target dataset with a high correlation coefficient from among other unselected regional datasets (S140); pre-training a learning model using the selected target dataset and regional datasets having a high correlation coefficient (S150); fine tuning the learning model using the target dataset (S160); and predicting items of the target dataset in the future with the learning model (S170).

The power information prediction method according to the illustrated flowchart is embodied as a case of division by region, and power information prediction is performed for all of the entire regions divided by region or a plurality of designated target regions.

For example, a specific electric power information prediction program is introduced for the entire autonomous district of Seoul, and it can be viewed as a case of predicting the electric power demand of the entire autonomous district of Seoul by performing learning using only electric power data of the entire autonomous district of Seoul for the past one month. That is, it responds to the problem of an insufficient period of precedent data by performing transfer learning using precedent data of other regions with an insufficient period as well.

For example, when all regions are designated as target regions, all regional datasets are designated as the target dataset (S130 and S184), and the step (S140) to predicting (S140) to predicting (S130, S184) S170) is repeated or performed in parallel.

In the illustrated flowchart, after performing the step (S170) of predicting items of the target dataset in the future with respect to the target dataset of a specific region, check whether there are remaining regions without performing prediction among the target regions. (S180), the target data set is changed for one of the remaining regions (S184), and the process returns to the step (S140) of selecting the one having a high correlation coefficient.

The regional division of the power data in step S10 may be performed in advance by another institution (such as Seoul), and when the method for predicting power information according to the spirit of the present invention is performed, the region classification performed by another organization in step S10 is performed in advance. information can be received.

In step S110, necessary information and data may be transmitted from a server of a related institution. For example, past power consumption data may be transmitted from a power company server, weather information may be transmitted from a weather service server, and population information may be transmitted from a local government server.

In step S120, min-max scaling is performed on the data/information collected in step S110, and it can be changed into a format usable for the learning model. For example, in the case of FIGS. 2 and 3 , a regional dataset may be generated for all autonomous districts of Seoul. The data set for each region consists of items of total power consumption, household power consumption, public power consumption, service power consumption, and industrial power consumption per unit period of each autonomous district. can be done

In the step S130, a regional data set of an autonomous district, which is a target for obtaining the above-described items of future power consumption, may be designated as a target data set.

As the related environmental information, the resident population of each autonomous district, weather, etc. may be applied, and in the step (S170) of predicting the items of the future target dataset, the prediction time of each region (ie, autonomous district) of the future It is advantageous to apply the resident population and weather information.

To this end, in the step of predicting the items of the target dataset in the future ( S170 ), prediction values of other organizations (eg, the Meteorological Administration) for future related environmental information may be collected and used.

6 is a flowchart illustrating another embodiment of a method for predicting power information according to the spirit of the present invention.

The illustrated method for predicting power information includes: collecting historical power data and related environmental information that can be acquired in a specific area (S210); generating a target data set by pre-processing the collected past power data (S220); selecting one having a high correlation coefficient with the target dataset from among previously acquired regional datasets of other regions (S240); pre-training a learning model using the selected target dataset and regional datasets having a high correlation coefficient (S250); fine tuning the learning model using the target dataset (S260); and predicting items of the target dataset in the future with the learning model (S270).

The power information prediction method according to the illustrated flowchart is embodied as a case of division by region, and shows a process of performing power information prediction of one specific region.

For example, a specific electric power information prediction program is introduced for a specific autonomous district of Seoul. Learning is performed using only the electric power data of the specific autonomous district of the past month, but the prediction is made using the past electric power data of other autonomous districts in Seoul. can take a look At this time, one or more of the other autonomous regions may have relatively long past power data of more than two months in the past.

Next, we will look at the effect of transfer learning using different regional datasets of the same type having high correlation according to the spirit of the present invention.

The proposed model for effect verification is a monthly load prediction model, and the dataset used to construct it is described. The dataset provided by the Seoul Metropolitan Government includes the monthly electrical loads of 25 districts in Seoul from January 2005 to December 2018. The data set can be divided into a total of five categories, such as household, public, industrial, and service, as shown in FIGS. 2 and 3 .

The residential dataset consists of monthly electricity consumption by pure residential customers, the public dataset includes monthly electricity consumption for public purposes such as infrastructure and government agencies, and the industrial dataset consists of the monthly electricity consumption for mining, manufacturing, construction, etc. It shows monthly electricity consumption, the service dataset includes monthly electricity consumption for water/wastewater treatment, high-speed transportation systems, commercial office and retail buildings, and finally the total dataset is the sum of these four categories of electricity consumption. indicates.

As a result, the dataset consists of 125 sequence data over 14 years covering 5 categories and 25 boroughs, and is used as a source domain for transfer learning. Here, a regional data set is configured for each autonomous district.

Looking at the input data (information) configuration, first, there is calendar data as related environment information. In general, power consumption can be determined by various factors. Of these, three main factors can be considered. The first major element is calendar data, such as the number of years, months, seasons, days, weekends, and holidays.

Months have periodic properties. For example, since December and January are temporally adjacent, they have similar characteristics in terms of temperature and season. In order to reflect this periodicity, the monthly data can be converted into continuous data reflecting a sine wave.

Also, power consumption is generally reduced during public holidays. Therefore, the number of days and holidays in a month are related to the electricity consumption in that month. In particular, it is advantageous to consider the number of weekday holidays for the input variable, as the number of weekday holidays can have a large impact on the model. In addition, if the holiday period is extended, there is a possibility that an unusual power demand pattern may occur, so the maximum length of holiday for the month is also taken into consideration. For example, the website (https://www.timeanddate.com) may collect holiday data from the time and date. As a result, 24 input variables can be selected from the calendar data.

Another environmental information considered for model building is local population data. Information on the population of Seoul can be found at the Seoul Open Data Plaza (http://data.seoul.go.kr/). Here, identifiable population data includes population, area of district, population density, in-migration, out-of-migration and net migration. The data is determined in March of each year based on ID card information, and the two types of population data are annual and monthly.

Annual population data includes population, global area and population density, and monthly population data includes during migration, outward migration and net migration. The population of a district (ie, a borough) is highly correlated with electricity consumption in that district. Monthly population data can be considered for further analysis.

As the last relevant environmental information, we apply weather data such as temperature, humidity and precipitation. Monthly weather data were collected from the Korea Meteorological Administration (KMA), which provides various types of forecasts, including three-day, medium-term, and long-term forecasts consisting of one-month/three-month forecasts. Monthly weather forecasts can be provided using MTLF's one-month forecasts. One-month projections include average temperature, average maximum temperature, average minimum temperature, and total precipitation in one month.

The target learning model aims to provide a power usage prediction model for a specific autonomous district. Accordingly, the target domain, that is, the target dataset, is expressed as the amount of electricity for each autonomous district, and the amount of electricity for each autonomous district of another autonomous region becomes the source domain, that is, the data set for each region. As mentioned earlier, there are a total of 125 unions in 25 boroughs and 5 categories.

First, a monthly electrical load prediction model is constructed using similar electrical load data from the remaining 124 electrical load data as the source domain. To this end, we first divided each source data and target data into training and test sets. To find similar load data in the source domain, we computed the PCC between the training sets of the source and target data, and then constructed a DNN-based predictive learning model using the selected source data.

Finally, the predictive learning model was fine-tuned using the training set of the target data (target dataset) and the model was evaluated using the test set of the target data. The overall steps of constructing the predictive learning model are as shown in FIGS. 4B to 4D above.

A very extensive experiment was performed to evaluate the effectiveness of the predictive learning model. For the experiment, five items of monthly electrical load data were collected from January 2005 to December 2018 in 25 districts of Seoul. Among them, data from January 2005 to December 2016 were used as the training set, and data from January 2017 to December 2018 were used as the test set. In our experiments, we considered all combinations of districts and categories as target domains and other data as source domains. All experiments were performed in Python 3.7.6 and the model was constructed using TensorFlow 1.13.1.

A part of the table including the PCC values calculated for 125 domains is as shown in FIG. 3 above. Here, each cell is marked with a unique color according to the PCC value.

3 schematically shows data characteristics for each autonomous district and each category. For example, the total of Jongno is very similar to that of most other autonomous districts because the PCC value is high. On the other hand, the industry in Yongsan is very different from most of the other regions because the PCC value is low. This can be interpreted that the entire data of Jongno is suitable for transfer learning, whereas the industrial data of Yongsan is not. As a result, if similarity between data is well considered, electrical data from various areas can be effectively used for transfer learning.

In addition, three DNN models were constructed by selecting the top 10 (DNN_T10), 20 (DNN_T20), and 30 (DNN_T30) domains that were most similar in terms of PCC to observe the effect of the number of data for training the predictive model.

Table 1 below lists the averages of the highest 10, 20, and 30 PCC correlation coefficient values in the five categories for each borough. Here, the mean absolute percentage error (MAPE) and the normalized root mean square error (NRMSE) were considered for performance comparison.

[Table 1]

MAPE and NRMSE may be defined according to Equations 1 and 2 below.

[Equation 1]

[Equation 2]

및

는 각각 시간 i에서의 실제 값과 예측 값,

는 실제 값의 평균이다.In the above equations, n is the number of observations,

and

are the actual and predicted values at time i, respectively,

is the average of the actual values.

In addition to the above three models, the following four predictive models were considered for comparison - multiple linear regression (MLR), RF, extreme gradient boost (XGB) and baseline DNN. MAPE and NRMSE values for all classifications were calculated and then the average value for each school district was calculated.

Tables 2 and 3 below show the average MAPE and NRMSE of the five datasets for each autonomous district.

[Table 2]

(DNN: deep neural network; DNN_T10: DNN pre-trained using top 10 similar data; DNN_T20: DNN pre-trained using top 20 similar data; DNN_T30: DNN pre-trained using top 30 similar data ;MLR: Multiple Linear Regression, RF: Random Forest, XGB: Extreme Gradient Boosting)

[Table 3]

(DNN: deep neural network; DNN_T10: DNN pre-trained using top 10 similar data; DNN_T20: DNN pre-trained using top 20 similar data; DNN_T30: DNN pre-trained using top 30 similar data ;MLR: Multiple Linear Regression, RF: Random Forest, XGB: Extreme Gradient Boosting.)

The most commonly used method for tuning hyperparameters is a grid search that tries all possible combinations of hyperparameters of interest. Therefore, we used grid search with cross-validation to select optimal hyperparameters for RF and XGB models.

By the way, the hyperparameter tuning method that divides the data into training, validation, and test sets can cause the problem of overfitting the validation set. Cross-validation was performed.

7A and 7B are box graphs for MAPE and NRMSE of predictive models.

It can be seen that the three transfer learning-based MTLF models have better predictive performance than the baseline DNN and other machine learning-based models.

8 is a graph showing the relationship between the mean absolute percentage error (MAPE) reduction rate versus the Pearson correlation coefficient.

The graph of FIG. 8 shows the trend intuitively by presenting the MAPE reduction rate for the PCC values for DNN_T10, 20, and 30. The x-axis represents the PCC, the y-axis represents the MAPE reduction rate, and each dot represents the MAPE reduction rate for each PCC in the proposed model. The central red line represents the trend line of MAPE values. Since the graph shown shows a positive trend line slope, it can be seen that an increase in the PCC value corresponding to transfer learning is closely related to the performance improvement.

In summary, in a specific configuration for examining the effects of the present invention, elements with similar monthly power usage patterns in terms of time, population, weather, and power one week ago are analyzed and modeled using Pearson correlation coefficient analysis and transfer learning. use for learning By pre-training the model with the associated data, the accuracy of the model can be increased.

In particular, FIG. 7A is a graph comparing prediction performance of several prediction models and the present models (DNN_T10, DNN_T20, DNN_T30). Even if various information such as time, population, weather, and electricity a week ago, as well as gender and age group are added, it can be analyzed in the same process. A heat map allows you to visualize elements and the degree of association between elements.

It can be seen that the annual power scheduling can be applied to the actual usage through more accurate monthly power demand prediction than when general data is used, thereby reducing power loss.

The above description is merely illustrative of the technical spirit of the present invention, and various modifications and variations will be possible without departing from the essential characteristics of the present invention by those skilled in the art to which the present invention pertains. Therefore, the embodiments disclosed in the present invention are not intended to limit the technical spirit of the present invention, but to illustrate, and the scope of the technical spirit of the present invention is not limited by these embodiments. The protection scope of the present invention should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

110: data collection unit
120: data preprocessor
140: correlation analysis unit
150: model pre-learning unit
160: model fine tuning unit
170: prediction unit
180: storage

Claims (10)

a data collection unit for collecting historical power data and related environmental information obtainable by region or group;
a data pre-processing unit for pre-processing the data collected by the data collecting unit;
a correlation analysis unit for selecting a regional dataset having the highest Pearson correlation coefficient based on a target dataset for a specified region and other unspecified regional datasets;
a model pre-training unit for pre-learning a learning model using the regional dataset selected by the correlation analysis unit;
a model fine-tuning unit for fine-tuning the learning model using the target dataset; and
A prediction unit that predicts items of the target dataset in the future by using the learning model
Power information prediction system comprising a.
According to claim 1,
A storage unit that stores the data and information collected by the data collection unit, and stores the learning model
Power information prediction system further comprising a.
According to claim 1,
The correlation coefficient is a Pearson correlation coefficient,
The learning model is a power information prediction system that is a DNN model.
4. The method of claim 3,
The correlation analysis unit,
If the average value of the correlation coefficient is 0.5 or less, the top 5 high correlation coefficients are selected, if the average value is 0.5 to less than 0.7, the top 10 are selected, if the average value is 0.7 to less than 0.8, the top 20 data sets, and if the average value is 0.8 or more, the top 30 regional datasets are selected. A power information prediction system that selects.
4. The method of claim 3,
The correlation analysis unit,
Power information prediction system for performing correlation analysis according to the following equation.

(where R _XY is the PCC between X and Y, n is the number of observations, X _i and Y _i are the values of time i,

and

is the average value of X and Y, respectively)
According to claim 1,
The data collection unit,
A data communication module that connects to an external agency server that has the source information of the data to be collected and accesses and receives the source information
Power information prediction system comprising a.
classifying the power data by region or group;
Collecting obtainable past power data and related environmental information;
preprocessing the collected past power data, and generating regional or group-specific datasets according to region or group classification;
designating a target dataset from among the datasets for each region or group, and selecting a data set having a high correlation coefficient with the target dataset from among other unselected datasets for each region or group;
pre-training a learning model using the selected target data set and a data set for each region or group having a high correlation coefficient;
fine-tuning the learning model using the target dataset; and
predicting items of the target dataset in the future with the learning model
Power information prediction method comprising a.
8. The method of claim 7,
In the step of predicting the items of the target dataset in the future, the power information prediction method for collecting and using prediction values of other organizations for the future related environmental information.
8. The method of claim 7,
Power information prediction method, characterized in that for all regional datasets, specifying the target dataset, and selecting or predicting the one having a high correlation coefficient.
Collecting obtainable past power data and related environmental information of a specific region or group;
generating a target data set by pre-processing the collected past power data;
selecting one having a high correlation coefficient with the target dataset from among previously acquired regional datasets of other regions;
pre-training a learning model using the selected target dataset and regional datasets having a high correlation coefficient;
fine-tuning the learning model using the target dataset; and
predicting items of the target dataset in the future with the learning model
Power information prediction method comprising a.

Images