KR102593144B1 - Anomaly detiection and repair based electrical load forecasting device and method - Google Patents

Abstract

The present invention relates to an apparatus and method for predicting strategic demand, and more specifically, to an apparatus and method for accurately predicting future strategic demand through machine learning. The power demand prediction device according to an embodiment of the present application includes a data collection unit that collects weather data, power data, and time data; a data pre-processing unit that performs pre-processing operations on the weather data, the power data, and the time data; An outlier detection unit implemented through a variational autoencoder and classifying the preprocessed power data into abnormal data and normal data; An outlier restoration unit implemented through a random forest model and restoring the abnormal data to generate restored data; and a power demand prediction unit that is implemented through a sliding window-based LightGBM model and learns a prediction model based on the normal data and the restored data and predicts power demand. The power demand prediction device according to an embodiment of the present application detects outliers, restores them, and uses them as learning data. By restoring outliers and using them as learning data in situations where the number of data is insufficient, the strategic demand prediction device according to the embodiment of the present application can provide improved forecast performance without overfitting.

Description

Anomaly detection and restoration based power demand prediction device and method {ANOMALY DETIECTION AND REPAIR BASED ELECTRICAL LOAD FORECASTING DEVICE AND METHOD}

The present invention relates to an apparatus and method for predicting strategic demand, and more specifically, to an apparatus and method for accurately predicting future strategic demand through machine learning.

Smart grid is attracting a lot of attention as a feasible solution to environmental problems and resource depletion problems occurring around the world, and various systems such as energy storage systems, energy management systems, and renewable energy systems that make up the smart grid are efficiently used. To do this, highly accurate power demand forecasting is required.

As computer technology has recently developed, research on machine learning and deep learning-based prediction models is actively underway and is showing good performance. These machine learning and deep learning-based prediction models are greatly influenced by the quantity and quality of data.

If there are many outliers or missing values in the data used to train the prediction model, it may interfere with model learning and lower prediction accuracy. When a sufficient amount of data has been collected for learning, there is no problem in learning the model after removing outliers and missing values. However, if the number of outliers and missing values is removed when the number of data is insufficient, the model learning itself may be hindered due to overfitting. It can get difficult.

The purpose of the present invention is to provide a strategic demand forecasting device that can improve the performance of a prediction model without overfitting through outlier detection and restoration in situations where the number of data is insufficient.

The power demand prediction device according to an embodiment of the present application includes a data collection unit that collects weather data, power data, and time data; a data pre-processing unit that performs pre-processing operations on the weather data, the power data, and the time data; An outlier detection unit implemented through a variational autoencoder and classifying the preprocessed power data into abnormal data and normal data; An outlier restoration unit implemented through a random forest model and restoring the abnormal data to generate restored data; and a power demand prediction unit that is implemented through a sliding window-based LightGBM model and learns a prediction model based on the normal data and the restored data and predicts power demand.

In an embodiment, the data preprocessor may include a weather data preprocessor that performs a normalization operation on the weather data; A power data preprocessor that performs a normalization operation on the power data; and a time data preprocessor that converts the time data into two-dimensional time data.

In an embodiment, the weather data preprocessor generates perceived temperature data and discomfort index data based on temperature, humidity, and wind speed data among the weather data, and performs a normalization operation on the generated perceived temperature data and discomfort index data. .

In an embodiment, the time data preprocessor converts the time data into two different two-dimensional time data using a periodic function.

In an embodiment, the weather data collected by the data collection unit is either weather forecast data or measured weather data.

In an embodiment, the power demand prediction unit has a window size of 7.

The power demand prediction method according to an embodiment of the present application includes collecting weather data, time data, and power data in a data collection unit; In a data preprocessing unit, performing a preprocessing operation on the weather data, the time data, and the power data; In an outlier detection unit, detecting outliers among the preprocessed power data; In an outlier restoration unit, restoring the outliers to generate restored data; and, in a power demand prediction unit, learning a prediction model using the restored data as learning data, and predicting power demand based on the learned prediction model.

In an embodiment, performing the preprocessing operation may include normalizing the weather data and the power data; and converting the time data into two different time data using a periodic function.

In an embodiment, normalizing the weather data includes generating perceived temperature data and discomfort index data based on temperature, humidity, and wind speed data among the weather data; and performing a normalization operation on the generated perceived temperature data and the generated discomfort index data.

In an embodiment, the weather data is either weather forecast data or measured weather data.

In an embodiment, the outlier detection unit is implemented through a variational autoencoder.

In an embodiment, the outlier restoration unit is implemented through a random forest model.

In an embodiment, the power demand prediction unit is implemented through sliding window-based LightGBM.

In an embodiment, the power demand prediction unit has a window size of 7.

The power demand prediction device according to an embodiment of the present application detects outliers, restores them, and uses them as learning data. By restoring outliers and using them as learning data in situations where the number of data is insufficient, the strategic demand prediction device according to the embodiment of the present application can provide improved forecast performance without overfitting.

Figure 1 is a block diagram showing a strategic demand prediction device 10 according to an embodiment of the present application.
FIG. 2 is a block diagram showing the data preprocessing unit 200 of FIG. 1 in more detail.
FIG. 3A is a diagram showing an example of categorical time data (Data_T).
FIG. 3B is a diagram showing an example of two-dimensional time data (Data_T1 and Data_T2).
FIG. 4 is a diagram showing the outlier detection unit 300 of FIG. 1 in more detail.
Figure 5 is a graph showing the results of an outlier detection experiment by the outlier detection unit 300.
FIG. 6 is a diagram showing the outlier restoration unit 400 of FIG. 1 in more detail.
Figure 7 is a graph showing the results of a restoration experiment of the outlier restoration unit 400.
FIG. 8 is a diagram showing the power demand prediction unit 500 of FIG. 1 in more detail.
Figure 9 is a graph showing the results of an experiment for determining the window size of the power demand prediction unit 500.
Figure 10 is a graph comparing the prediction results of the power demand prediction device 10 according to an embodiment of the present application with other models.
FIG. 11 is a flowchart showing the operation of the power demand prediction device 10 of FIG. 1.

Hereinafter, in order to explain in detail enough to enable a person skilled in the art of the present application to easily implement the technical idea of the present application, embodiments of the present application will be described in more detail with reference to the accompanying drawings. will be.

Figure 1 is a block diagram showing a strategic demand prediction device 10 according to an embodiment of the present application.

Referring to FIG. 1, the strategic demand prediction device 10 includes a data collection unit 100, a data preprocessing unit 200, an outlier detection unit 300, an outlier restoration unit 400, and a power demand prediction unit 500. Includes.

The data collection unit 100 collects various data necessary for predicting power demand. For example, the data collection unit 100 may collect weather data, power data, and time data.

For example, the data collection unit 100 may collect data about various types of weather from the outside. The meteorological data collected by the data collection unit 100 may be, for example, data on daily maximum temperature, daily minimum temperature, temperature, humidity, wind speed, total cloud cover, and precipitation. The data collection unit 100 may collect meteorological data through, for example, the Korea Meteorological Administration's meteorological data open portal.

For example, the data collection unit 100 may collect power data about the amount of power consumed from at least one cluster. In this case, clusters may be buildings for different purposes to prevent data bias. For example, the data collection unit 100 may collect power data from cluster A comprised of educational buildings, cluster B comprised of dormitories, cluster C comprised of engineering college labs, and cluster D comprised of science college labs. However, this is an example, and the number and type of clusters can be set in various ways.

For example, when collecting various weather and power data, the data collection unit 100 may also collect corresponding time data. For example, when collecting weather data, the data collection unit 100 may also collect time data for month, day, hour, and minute. Additionally, when collecting power data, the data collection unit 100 may also collect time data for month, day, hour, and minute.

In one embodiment of the present application, the data collection unit 100 may collect not only actually measured weather data but also weather forecast data predicted a predetermined period of time from the past. For example, when collecting power consumption data and weather data on May 5, 2018, the weather data corresponding to May 5, 2018 is the minimum daily temperature forecast on May 4, 2018, the day before. , it may be weather forecast data for daily maximum temperature, daily average temperature, temperature, humidity, wind speed, total cloud cover, precipitation, etc. In this way, by collecting not only the actual measured weather data but also the weather forecast data and providing this as learning data for implementing a learning model, the power demand prediction device 10 according to the embodiment of the present application uses the weather forecast data and the actual weather forecast data. Learning operations can be performed by considering errors between weather data.

However, this is an example, and the data collection unit 100 may collect only actually measured weather data. In this case, the learning operation may be performed using only the actually measured weather data. As another example, the data collection unit 100 may collect only weather forecast data, and in this case, the learning operation may be performed using only weather forecast data.

The data pre-processing unit 200 receives weather data, time data, and power data from the data collection unit 100. The data preprocessing unit 200 performs a preprocessing operation on the received weather data, power data, and time data so that they can be used in the anomaly detection unit 300.

For example, the data pre-processing unit 200 may receive weather data and power data from the data collection unit 100 and perform a normalization operation on the received weather data and power data. Additionally, the data pre-processing unit 200 may receive one-dimensional time data from the data collection unit 100 and convert the received one-dimensional time data into two-dimensional time data. The configuration and operation of the data pre-processing unit 200 will be described in more detail in FIGS. 2 and 3 below.

The outlier detection unit 300 receives preprocessed power data from the data preprocessor 200. The outlier detection unit 300 may detect outliers in preprocessed power data.

In one embodiment of the present application, the outlier detection unit 300 may detect outliers through a variational autoencoder (VAE). Since the variational autoencoder generates output values by learning the distribution of input values, the outlier detection unit 300 can better detect outliers that deviate from the general power demand distribution. The outlier detection unit 300 will be described in more detail in FIGS. 4 and 5 below.

The outlier restoration unit 400 receives abnormal data from the outlier detection unit 300. The outlier restoration unit 400 restores the abnormal data based on other input variables at the time corresponding to the abnormal data and generates restored data.

In one embodiment of the present application, the outlier restoration unit 400 may restore restored data from outlier data through a random forest (RF) model. The outlier restoration unit 400 will be described in more detail in FIGS. 6 and 7 below.

The power demand prediction unit 500 receives input data including restored data from the outlier restoration unit 500. Here, the input data includes data preprocessed in the data preprocessing unit 200 and data restored in the outlier restoration unit 400. The power demand prediction unit 500 may learn a model for predicting power demand using input data.

In one embodiment of the present application, the power demand prediction unit 500 may be implemented through a sliding window-based Light GBM model. The power demand prediction unit 500 will be described in more detail in FIGS. 8 and 9 below.

As described above, the power demand prediction device 10 according to an embodiment of the present application detects outliers through a variational autoencoder and restores the outliers through a random forest model. Therefore, stable training data can be provided even in situations where the number of data is insufficient, and accordingly, the power demand prediction device 10 can provide improved prediction performance without overfitting. In addition, the power demand prediction device 10 implements a prediction model through a sliding window-based Light GBM model, and thus can properly reflect the latest data patterns close to the prediction time, thereby further improving prediction performance.

FIG. 2 is a block diagram showing the data preprocessing unit 200 of FIG. 1 in more detail.

Referring to FIG. 2, the data pre-processing unit 200 includes a meteorological data pre-processing unit 210, a power data pre-processing unit 220, and a time data pre-processing unit 230.

The weather data pre-processing unit 210 receives weather data (Data_W) from the data collection unit 100. The weather data preprocessor 210 performs a preprocessing operation on the received weather data (Data_W) to generate first weather data (Data_W1).

For example, the weather data pre-processing unit 210 may receive data such as daily maximum temperature, daily minimum temperature, temperature, humidity, wind speed, total cloud cover, and precipitation amount from the data collection unit 100. The weather data preprocessor 210 may generate first weather data (Data_W1) by performing a normalization operation on the received weather data.

Additionally, the weather data preprocessor 210 may generate perceived temperature data and discomfort index data based on the received temperature, humidity, and wind speed data. Thereafter, the weather data preprocessor 210 may generate second weather data (Data_W2) by normalizing the perceived temperature data and discomfort index data.

The power data preprocessor 220 may receive power data about the amount of power consumed in at least one cluster from the data collection unit 100. The power data preprocessor 220 may perform a preprocessing operation on the received power data to generate preprocessed power data (Data_Pp).

The time data pre-processing unit 230 receives time data (Data_T) from the data collection unit 100. Here, time data may be one-dimensional data that is good for reflecting categorical data. The time data preprocessor 230 outputs one-dimensional time data (Data_T) as is, or converts one-dimensional time data (Data_T) into two-dimensional time data ((Data_T1, Data_T2) that can reflect periodicity information. can do.

To explain in more detail, one-dimensional time data includes categorical information representing months such as January and February, categorical information representing days such as the 1st and 2nd, and 1 o'clock and 2 o'clock. Likewise, categorical information representing hours is reflected well. However, one-dimensional time data has a problem in that it does not reflect periodicity information well. For example, although 23:00 and 0:00 are continuous times, a difference of 23 occurs in one-dimensional data.

Therefore, so that the periodicity information of time can be well reflected, the time data preprocessor 230 can convert one-dimensional time data (Data_T) into two-dimensional time data ((Data_T1, Data_T2). At this time, time The data pre-processing unit 230 may perform a conversion operation through periodic functions such as a sine function and a cosine function (cos). For example, the time data pre-processing unit 230 uses the following equation: 1 One-dimensional time data (Data_T) can be converted into two-dimensional time data (Data_T1, Data_T2).

Here, “cycle” represents the cycle of time data. For example, if "time" is month data, "cycle" may be "12", and if "time" is day data, "cycle" may be the number of days in the month. the Month), and if “time” is hour data, “cycle” may be “24”. Also, here: " "and " " may correspond to Data_T1 and Data_T2 in FIG. 2, respectively.

Meanwhile, the reason why the time data preprocessor 230 expresses it in two dimensions through two trigonometric function values is that, for example, when expressing through one trigonometric function value with a period of 12, the same y for two x values This is because the value is determined, and in this case, it is difficult to specify the time only with the y value. Therefore, in order to solve this problem by using two trigonometric functions with different y values even if the x value is the same, the time data preprocessor 230 uses two trigonometric functions to generate two-dimensional data (Data_T1, Data_T2). Create.

As described above, the data preprocessor 200 may perform preprocessing operations on weather data (Data_W), power data (Data_P), and time data (Data_T). In particular, the data preprocessor 200 can output one-dimensional time data (Data_T) that well reflects categorical information and two-dimensional time data (Data_T1, Data_T2) that well reflect periodic information.

FIG. 3 is a diagram showing an example of time data output by the time data preprocessor 230 of FIG. 2. Specifically, FIG. 3A shows an example of categorical temporal data (Data_T), and FIG. 3B shows an example of two-dimensional temporal data (Data_T1 and Data_T2).

As shown in FIG. 3, the time data preprocessor 230 may perform a preprocessing operation on time data Data_T to generate two-dimensional time data Data_T1 and Data_T2.

FIG. 4 is a diagram showing the outlier detection unit 300 of FIG. 1 in more detail, and FIG. 5 is a graph showing the results of an outlier detection experiment of the outlier detection unit 300.

Referring to FIG. 4, the outlier detection unit 300 receives preprocessed power data (Data_Pp) from the data preprocessor 200. The outlier detection unit 300 can detect outliers in the preprocessed power data (Data_Pp) and distinguish them into normal data and abnormal data.

In an embodiment of the present application, the outlier detection unit 300 may be implemented through a variational autoencoder. That is, the outlier detection unit 300 can learn a variational encoder using preprocessed power data and detect an outlier using the difference between the output value and the input value reconstructed through the learned variational autoencoder.

If the outlier detection unit is implemented through a general autoencoder (AE), the autoencoder-based anomaly detection unit reconstructs all input variables and then adds all the reconstruction errors of each input variable. Afterwards, the autoencoder-based anomaly detection unit determines that it is an outlier when the sum of reconstruction errors exceeds a certain value. However, in this case, even when an outlier does not occur in actual power consumption, there is a risk that the autoencoder-based outlier detection unit may determine it to be an outlier.

For example, let's assume that it suddenly rained heavily at a certain point due to the rainy season during the summer. In this case, because the autoencoder-based anomaly detection unit considers the reconstruction error of all input variables, even though the reconstructed input variables, including power usage, are normal, a large difference occurs in the reconstructed rainfall-related variables, resulting in an outlier. You may judge incorrectly.

To prevent this risk of error, the outlier detection unit 300 according to an embodiment of the present application can be implemented through a variational autoencoder. In this case, rather than determining the outlier by adding the reconstruction errors of all input variables, the outlier is determined using only the reconstruction error of the preprocessed power data (Data_Pp), so the outlier detection unit ( 300) may have more improved outlier detection performance.

Referring to FIG. 5, it can be seen that the outlier detection unit 300 based on a variational autoencoder according to an embodiment of the present application has improved performance compared to other models. In Figure 5, IQR, IForest, and LOF mean an outlier detection model using the interquartile range (IQR), an Isolation Forest outlier detection model, and a Local Outlier Factor outlier detection model, respectively.

FIG. 6 is a diagram showing the outlier restoration unit 400 of FIG. 1 in more detail, and FIG. 7 is a graph showing the results of a restoration experiment of the outlier restoration unit 400.

Referring to FIG. 6, the outlier restoration unit 400 receives normal data and abnormal data from the outlier detection unit 300. The outlier restoration unit 400 may restore abnormal data into repair data based on other input variables at the time the abnormal data was detected.

In one embodiment of the present application, the outlier restoration unit 400 may be implemented using a random forest (RF) model. That is, the outlier restoration unit 400 can learn a random forest model using normal data and generate restored data with values derived when input variables at the time an outlier occurs are input into the learned random forest model. Random forest models use partial data sets to learn each decision tree. Accordingly, the outlier restoration unit 400 not only does not cause overfitting, but also can have improved performance despite the number of input variables.

Referring to FIG. 7, it can be seen that the random forest-based outlier restoration unit 400 according to an embodiment of the present application has improved performance compared to other models. In Figure 7, Zero, Linear, and RF indicate when an outlier is restored using zero interpolation, when an outlier is restored using linear interpolation, and when an outlier is restored using Random Forest, respectively.

FIG. 8 is a diagram showing the power demand prediction unit 500 of FIG. 1 in more detail, and FIG. 9 is a graph showing the results of an experiment for determining the window size of the power demand prediction unit 500.

Referring to FIG. 8, the power demand prediction unit 500 receives normal data and restoration data on power consumption from the outlier restoration unit 400, and receives first and second weather data from the data preprocessor 200. Data_W1, Data_W2), categorical time data (Data_T), and periodic time data (Data_T1, Data_T2) can be received. The power demand prediction unit 500 may learn a power demand prediction model based on the received input data and predict future power demand through the learned power demand prediction model.

In one embodiment of the present application, the power demand prediction unit 500 may be implemented through a sliding window-based Light GBM model. By applying the sliding window technique, the power demand forecasting unit 500 can appropriately reflect the latest trends and patterns.

For example, in FIG. 8, each dot may represent one day's worth of input data. In this case, since the power demand forecasting unit 500 uses input data from the previous week from the point in time to be predicted, it can reflect the latest data and have good performance.

In one embodiment of the present application, the power demand prediction unit 500 may set '7' as the window size. Referring to Figure 9, the performance improves as the window size increases from '1' to '7', but when the window size exceeds '7', the performance improvement is minimal, while only the learning time increases. there is. This is because data on electricity demand tends to repeat in a certain periodic pattern. In other words, if the point in time you want to predict is Monday, entering last Monday's data as an input variable can reflect periodic patterns and improve performance. The power demand prediction unit 500 according to an embodiment of the present application sets the window size to '7' to reflect this periodic pattern of power demand.

On the other hand, when applying the sliding window technique, there is a disadvantage that a new model must be constructed each time a prediction is desired. In order to compensate for these shortcomings, the power demand prediction unit 500 according to an embodiment of the present application can be implemented using the Light GBM model, which has a fast model construction speed and excellent prediction performance. In this case, the power demand forecasting unit 500 uses the GOSS (Gradient-based One Side Sampling) technique, which obtains information using only the portion of the data with a large gradient, and the EFB (Exclusive Feature Bundling) technique, which bundles and processes mutually exclusive variables. This can speed up model construction.

Figure 10 is a graph comparing the prediction results of the power demand prediction device 10 according to an embodiment of the present application with other models.

As shown in Figure 10, as a result of predicting the power demand for four different clusters, the MAPE (Mean Absolute Percentage Error) of the power demand prediction device 10 is 4.545%, 3.755%, 2.7%, and 2.144%, respectively. was recorded, and it can be confirmed that it has excellent prediction performance compared to other models.

FIG. 11 is a flowchart showing the operation of the power demand prediction device 10 of FIG. 1.

In step S110, the data collection unit 100 may collect weather data, time data, and power data.

In step S120, the data preprocessing unit 200 may perform a preprocessing operation on the collected weather data, time data, and power data. For example, the data preprocessor 200 may perform a normalization operation on weather data and power data. For example, the data preprocessor 200 may convert one-dimensional time data into two-dimensional time data.

In step S130, the outlier detection unit 300 may perform an operation to detect an outlier among the preprocessed power data. For example, the outlier detection unit 300 may classify power data into abnormal data and normal data, and the outlier detection unit 300 may be implemented through a variational autoencoder (VAE).

In step S140, the outlier restoration unit 400 may restore the abnormal data. For example, the outlier restoration unit 400 may restore abnormal data using a random forest model to generate restored data.

In step S150, the power demand prediction unit 500 may learn a prediction model based on input data including restored data. For example, the power demand prediction unit 500 may be implemented through sliding window-based LightGBM. Through the learned prediction model, the power demand prediction unit 500 can accurately predict power demand at a future point in time.

In the above, preferred embodiments according to the present invention are shown and described. However, the present invention is not limited to the above-described embodiments, and various modifications and modifications may be made by anyone skilled in the art without departing from the gist of the present invention attached to the scope of the patent claims. .

100: Data collection unit
200: Data preprocessing unit
300: Outlier detection unit
400: Outlier restoration unit
500: Power demand prediction unit

Claims (14)

A data collection unit that collects weather data, power data, and time data;
a data pre-processing unit that performs pre-processing operations on the weather data, the power data, and the time data;
An outlier detection unit implemented through a variational autoencoder that detects outliers using only reconstruction errors of the preprocessed power data, and classifies the preprocessed power data into abnormal data and normal data;
An outlier restoration unit implemented through a random forest model and restoring the abnormal data to generate restored data; and
It is implemented through a sliding window-based LightGBM model, and includes a power demand prediction unit that learns a prediction model and predicts power demand based on the normal data and the restored data,
The LightGBM model is constructed through a GOSS (gradient-based one side sampling) technique that obtains information based on the gradient of the normal data and the restored data and an EFB (exclusive feature bundling) technique that processes mutually exclusive variables by grouping them together. , power demand prediction device. According to claim 1,
The data preprocessing unit
a weather data preprocessor that performs a normalization operation on the weather data;
A power data preprocessor that performs a normalization operation on the power data; and
A power demand prediction device comprising a time data preprocessor that converts the time data into two-dimensional time data. According to clause 2,
The weather data preprocessor generates perceived temperature data and discomfort index data based on temperature, humidity, and wind speed data among the weather data, and performs a normalization operation on the generated perceived temperature data and discomfort index data. A power demand prediction device . According to clause 2,
The time data preprocessor converts the time data into two different two-dimensional time data using a periodic function. According to claim 1,
The weather data collected by the data collection unit is either weather forecast data or measured weather data. According to claim 1,
The power demand prediction unit has a window size of 7. In the data collection unit, collecting weather data, time data, and power data;
In a data preprocessing unit, performing a preprocessing operation on the weather data, the time data, and the power data;
In an outlier detection unit implemented through a variational autoencoder that detects outliers using only reconstruction errors of the preprocessed power data, classifying the preprocessed power data into abnormal data and normal data;
In an outlier restoration unit, restoring the abnormal data to generate restored data; and
In a power demand prediction unit implemented through a sliding window-based LightGBM model, learning a prediction model using the restored data as learning data, and predicting power demand based on the learned prediction model,
The LightGBM model is constructed through a GOSS (gradient-based one side sampling) technique that obtains information based on the gradient of the normal data and the restored data and an EFB (exclusive feature bundling) technique that processes mutually exclusive variables by grouping them together. , electricity demand prediction method. According to clause 7,
The step of performing the preprocessing operation is
normalizing the weather data and the power data; and
A method for predicting power demand, comprising converting the time data into two different time data using a periodic function. According to clause 8,
The step of normalizing the weather data is
Generating perceived temperature data and discomfort index data based on temperature, humidity, and wind speed data among the weather data; and
A power demand prediction method comprising performing a normalization operation on the generated perceived temperature data and the generated discomfort index data. According to clause 9,
The weather data is either weather forecast data or measured weather data. A method of predicting power demand. delete According to clause 7,
A method for predicting power demand, wherein the outlier restoration unit is implemented through a random forest model. delete According to clause 7,
A method for predicting power demand, wherein the power demand prediction unit has a window size of 7.