KR20180060317A - Apparatus and method for predicting electricity demand based on deep neural networks - Google Patents

Abstract

The present invention provides a device for predicting deep neural network-based customer power demand and a method thereof, which can accurately predict power demand of a customer in a dynamically changing grid environment. To this end, the device for predicting deep neural network-based customer power demand collects weather data, historical daily power usage data and date data for the historical daily power usage data for predicting the power demand of the customer, generates a learning data set based on the collected data, generates a basic deep neural network based on the learning data set, applies a set learning algorithm to the basic deep neural network to generate a deep neural network-based customer power demand prediction model, and predicts the power demand of the customer based on the deep neural network-based customer power demand prediction model.

Description

[0001] APPARATUS AND METHOD FOR PREDICTING ELECTRICITY DEMAND BASED ON DEEP NEURAL NETWORKS [0002]

The present invention relates to an apparatus and method for estimating a demand for a consumer electric power, and more particularly to an apparatus and a method for estimating an electric power demand of a consumer using a depth network.

The Smart Grid is a new type of power network that combines the existing power grid and information and communication technology (ICT), enabling bi-directional information exchange between suppliers and consumers. In addition, the boundary between suppliers and consumers has been broken down through ESS and renewable energy technologies. As a result, a wide range of demand changes will occur as the choice of the consumer becomes wider from generation, consumption, and charge discharge in a single option of power consumption, and the proportion of the consumer in the smart grid will also increase significantly.

Here, power demand forecasting technology is essential for optimized network operation and design. In particular, demand for short - term demand forecasts is expected to increase in order to efficiently utilize new services available in the smart grid, such as demand response and real - time rate plan.

In this regard, there are various demand forecasting methods in the past, but it can be classified into forecasting based on time series analysis and forecasting based on artificial neural network. Among these methods, the time series analysis technique can reflect the characteristics of the repeated power pattern at each cycle, but there is a disadvantage that it is relatively difficult to analyze and apply the external factors influencing the past power use history.

In addition, demand prediction methods based on artificial neural networks are mainly used for fuzzy neural networks or multilayer perceptron, which can easily reflect external factors. However, given the change in the role of the consumer in the smart grid, the need for a demand prediction model that can abstract the dynamics between power consumption and various factors at a higher level than the existing three-layer perceptron model of input layer, hidden layer and output layer Is emerging. However, in the case of the existing artificial neural network, when the number of hidden layers responsible for the abstraction is increased, the phenomenon that the predictive model is not learned properly due to vanishing gradient problem or the like occurs.

Korean Registered Patent No. 1025788 (Name: Power Usage Prediction Apparatus and Method Using Predictive Accuracy Feedback)

It is an object of the present invention to provide an apparatus and method for predicting a demand for a customer based on a neural network that can accurately predict the demand for power of a customer in a dynamically changing grid environment.

In order to solve the above-mentioned problems, the inventive deep-network neural network-based demand power demand forecasting apparatus includes data for collecting date data on weather data, past daily power use data and past daily power use data Collecting section; A data processing unit for generating a learning data set based on weather data, past daily power use data and date data; A model learning unit for generating a depth neural network based customer power demand prediction model by generating a basic depth neural network based on the learning data set and applying the set learning algorithm to the underlying neural network; And a demand forecasting unit for predicting power demand of a customer based on a depth-of-neural network-based consumer power demand prediction model.

In addition, the depth-of-neural network-based consumer demand forecasting model can be generated either through prior learning through the Restricted Boltzmann Machine (RBM) or by using ReLU (Rectified Linear Unit) as an active function.

Further, the model learning unit can select the learning parameters based on at least a part of the learning data sets, and can generate the underlying depth-based neural network based on the selected learning parameters.

The model learning unit also determines whether there is a K-fold cross validation for the training data set, generates a verification data set comprising at least a portion of the training data sets when performing the K-fold cross validation, and the learning parameters are based on the verification data set Can be selected.

Further, the data processor may determine a learning window used for determining a period of past data used for learning among past data.

Further, the learning data set is divided into a label and learning data. The label represents a load profile corresponding to a period to be predicted. The learning data includes meteorological data corresponding to the size of the learning window, daily power use data and date data .

In addition, the data processor includes a data preprocessing module for determining whether there is abnormal data by analyzing weather data, past daily power usage data, and date data, and the data preprocessing module, when abnormal data exists, The abnormal data can be corrected based on the data.

Further, the weather data may include at least one of an average temperature, an average humidity, an average wind speed, a singular value, and a circulation rate.

Further, the underlying neural network may be composed of an input layer, a hidden layer, and an output layer, and the input layer, the hidden layer, and the output layer may each be composed of an artificial nerve.

According to another aspect of the present invention, there is provided a method for predicting demand for a catering server based on neural network, comprising the steps of: estimating power demand of a customer based on meteorological data, past daily power usage data, Collecting date data; Generating a learning data set based on meteorological data, past daily power usage data and date data by a data processing unit; Generating a neural network based on the neural network based on the learning data set by the model learning unit and applying the set learning algorithm to the underlying neural network; And predicting, by the demand predicting unit, the power demand of the customer on the basis of the deep-network neural network-based consumer demand forecasting model.

In addition, a depth-of-neural network-based consumer demand forecasting model can be generated either through prior learning through RBM, or by using ReLU as an active function.

The generating of the depth network-based CAPR model may include: selecting a learning parameter based on at least a portion of the learning data set; And generating an underlying neural network based on the selected learning parameters.

The step of generating a depth-of-neural network-based CAPR model further includes the steps of: determining the presence or absence of a K-fold cross-validation of the learning data set; And generating a set of verification data comprising at least a portion of the set of learning data at the time of performing the K-fold cross-validation, wherein the learning parameters may be selected based on the verification data set.

In addition, generating the training data set may include determining a learning window that is used to determine a period of historical data used for learning among historical data.

Further, the learning data set is divided into a label and learning data. The label represents a load profile corresponding to a period to be predicted. The learning data includes meteorological data corresponding to the size of the learning window, daily power use data and date data .

Before the step of generating the learning data set, the step of determining whether the abnormal data exists by analyzing the meteorological data, the past daily power use data and the date data by the data processing unit; And correcting the abnormal data based on the data at the past time point more than the abnormal data when the abnormal data exists.

Further, the weather data may include at least one of an average temperature, an average humidity, an average wind speed, a singular value, and a circulation rate.

Further, the underlying neural network may be composed of an input layer, a hidden layer, and an output layer, and the input layer, the hidden layer, and the output layer may each be composed of an artificial nerve.

According to an apparatus and method for predicting a demand for a cryptographic power demand based on a neural network according to an embodiment of the present invention, it is possible to predict a power demand of a customer accurately in a dynamically changing grid environment, It is possible to reduce energy costs.

FIG. 1 is a flowchart illustrating a method of predicting a demand for a customer based on a neural network according to an embodiment of the present invention.
2 is a flowchart of a step of generating a learning data set according to an embodiment of the present invention.
FIG. 3 is a flowchart illustrating a step of generating a depth-of-field-based power demand prediction model through a learning algorithm according to an embodiment of the present invention.
4 is a structural diagram of a depth-of-field network according to an embodiment of the present invention.
5 is a structural view of an artificial nerve according to an embodiment of the present invention.
6 is a conceptual diagram of an RBM structure according to an embodiment of the present invention.
FIG. 7 is a block diagram of an apparatus for predicting a demand based on a neural network based on an embodiment of the present invention.

The present invention will now be described in detail with reference to the accompanying drawings. Hereinafter, a repeated description, a known function that may obscure the gist of the present invention, and a detailed description of the configuration will be omitted. Embodiments of the present invention are provided to more fully describe the present invention to those skilled in the art. Accordingly, the shapes and sizes of the elements in the drawings and the like can be exaggerated for clarity.

Hereinafter, a method and an apparatus for predicting a demand for a customer based on a neural network according to an embodiment of the present invention will be described.

FIG. 1 is a flowchart illustrating a method of predicting a demand for a customer based on a neural network according to an embodiment of the present invention.

Step S110 is a step of collecting date data for weather data, past daily power usage data, and past daily power usage data by the data collecting unit in order to predict the electric power demand of the customer. Specifically, the step S110 collects data necessary for prediction from the available external database or the AMI, and the data to be collected includes weather data including average temperature, average humidity, average wind speed, Usage data (load profile), and date data (season, month, day, workday) for power consumption data.

In addition, the step S110 may further store the collected data in the database. Further, the data stored in the database may be subjected to modification of the data through the data processing unit as described below. A database can also be implemented as a physical disk on which data can be stored, for example, to store data in the form of a csv file.

Step S120 is a step of generating a set of learning data used for predicting the demand for catering power, which will be described below, by processing the data stored in the database by the data processing unit. Here, the step S120 may be roughly divided into two functions: a function of extracting and modifying abnormal data from a database, and a function of generating a learning data set. To this end, step S120 may include the steps shown in FIG. In Fig. 3, steps S121 to S124 show steps corresponding to functions for extracting and modifying abnormal data, and steps S125 to S129 show steps for generating a learning data set.

Step S121 is a step of selecting parameters for abnormal data determination and correction processing.

In step S122, data validity determination is performed on the data stored in the database. Here, the abnormal data can be classified into a case where there is no data for the date or time, a case where the data is zero, and a case where the data has a specific value. As a result of the determination in step S122, if data detected as abnormal data is found, control is passed to step S123 to perform a correction process on the abnormal data.

For example, if abnormal data (for example, data out of the distribution of the same time zone data of the other day based on the parameter determined in step S121) is searched, the step S123 is performed based on the parameter determined in the step S121 Lt; RTI ID = 0.0 > data. ≪ / RTI > For example, it can be modified to data within the preset interval based on the same time data of different dates or the same time of different dates. When the processing for all the abnormal data stored in the database is completed, the control is transferred to the step S124 to update the database.

Steps S125 to S129 represent a series of transformation steps for generating a learning data set. That is, steps S125 to S129 are steps of generating a learning data set based on data stored in the database before the prediction date to be predicted. However, since the process for generating learning data differs depending on the type of data, the process of dividing the data type of the data stored in the database is performed through step S125 before the generation of the learning data set.

As described above, the database may store date data for meteorological data, past daily power usage data, and past daily power usage data. In step S125, the type of each data stored in the database is classified, so that the weather data and past daily power consumption data can be processed through step S126, and the date data can be processed through step S129.

Step S126 is a step of performing maximum value calculation and normalization. Specifically, the step S126 is a step of performing normalization so as to have a value between 0 and 1 by calculating the maximum value among the past data in the weather data and the past daily power use data.

Step S129 is a step of performing mapping so as to have a discrete value between 0 and 1 in the date data such as the season, month, day of the week, presence or absence of the working day, and the like.

In step S127, a learning window for determining a period of past data to be used for prediction is determined. In step S128, based on the converted data and the learning window through steps S125, S126, S127, and S129, .

Here, the learning data set can be largely divided into a label and learning data. Here, the label represents a normalized load profile corresponding to a period to be predicted. For example, when the power demand for 24 hours after the prediction time is predicted, the 24-hour power demand for the past day corresponds to the load profile. The learning data in the learning data set includes power usage data from the load profile corresponding date to the past n days and date data (n) according to the size (n) of the learning window determined in step S327 from the data measured before the date corresponding to each load profile And may include date information and weather data of the load profile corresponding date. For example, the training data set can be set as shown in Table 1 below.

Referring again to FIG. 1, in step S130, a model learning unit generates a depth NN based on a learning data set, and applies a set learning algorithm to a basic NN to generate a depth NN based consumer power demand prediction model . Specifically, step S130 can generate a depth-of-neural network-based consumer demand forecasting model by including the steps shown in FIG.

Step S131 is a step of determining whether or not a K-fold verification is performed for selecting a learning parameter. As a result of the determination in step S131, when the K-fold cross validation is performed, control is passed to step S133. Otherwise, control passes to step S132.

Step S133 is a step performed at the time of performing the K-fold cross-validation, and is a step of generating a verification data set composed of at least a part of the learning data sets. Based on the verification data set, the model with the smallest prediction model learning result error is used for demand forecasting.

Step S132 is a step of selecting a learning parameter. Specifically, if the K-fold cross validation is not performed in step S131, the learning parameter determined by the user is basically selected in step S132. If the K-fold cross validation is performed in the opposite manner, as described with reference to step S133 Based on the verification data set, the learning parameters are selected based on the model with the smallest prediction model learning result error.

Step S134 is a step of generating the structure of the underlying NN using the learning parameters determined through steps S131 to S133. As described above, demand prediction methods based on artificial neural networks in general are advantageous in that they are easily reflected on external factors in the form of fuzzy neural network or multilayer perceptron. However, when the number of hidden layers responsible for abstraction is increased, The problem is that the prediction model is not learned properly.

Accordingly, in order to overcome such a disappearing slope problem, the depth-based neural network-based demand power demand prediction method according to an embodiment of the present invention uses an RBM pre-learning method or ReLU (Rectified Linear Unit) A neural network model can be selected.

As shown in FIG. 4, the deep layer neural network is composed of an input layer, a hidden layer, and an output layer, and each layer is composed of the artificial nerve of FIG. The learning parameters such as the number of artificial nerves constituting each layer and the number of learning repetition times of each layer follow the values selected in step S132. Can be determined through the verification data set used in the determination. However, the number of nodes in the input layer and the output layer is set to the same value as the length of the data and label in the learning data set, respectively. The in-depth neural network is learned using a set of learning data. Learning refers to sequentially updating the weights of the entire network so as to have a value matching a specific purpose. Learning in depth neural networks aims at minimizing the error between prediction and labeling through input data. This learning can be done through an error backpropagation algorithm.

If the RBM model is used, steps S136 and S137 may be sequentially performed (i.e., the RBM pre-learning step and the fine tuning step may be performed).

Here, in step S136, the RBMs are sequentially stacked so as to have the same structure as the target network structure. Learning in the RBM aims to maximize the probability that it exists in the state of the RBM corresponding to the given learning data. At this time, the probability that the RBM exists in a specific state with respect to the learning data is defined as the energy of the RBM. Here, an example of the structure of the RMB is shown in Fig. In addition, the state probability of the RBM can be expressed by the following equation (1).

In Equation (1), v and h denote the on / off states of the visible layer and the hidden layer, respectively, and E (v, h) denotes the energy of the RBM depending on the states of the visible layer and the hidden layer. Also, because Z is a value for normalization, it is difficult to calculate the RBM effectively. Therefore, the RBM is learned by the Constructive Divergence algorithm in actual learning. In step S136, the weight of the network generated through the pre-learning process may be set as the initial weight of the neural network, and the learning can be performed through the error back propagation algorithm.

Thereafter, in step S138, a depth neural network-based customer demand prediction model can be generated.

As a result of the determination in step S135, if the ReLU model is used, the model can be learned by the error back propagation algorithm using ReLU as an active function through step S139. Herein, in the case of the depth neural network using ReLU, the activation function in the artificial nerve of FIG. 6 is applied as Equation 2 below.

)에 대한 가중합(weighted summation)을 나타낸다.In Equation (2), z is the input data vector of the artificial nerve

) Of the weighted summation.

Step S140 is a step of predicting the demand for the electric power of the customer based on the hypothetical network-based demand power demand prediction model generated by the demand forecasting unit in step S130. Specifically, in step S140, the depth-based neural network-based power demand prediction model generated in step S130 predicts the demand for power for the period after the prediction time using the data collected until the prediction time. Since the predicted power demand is normalized, it is multiplied by the maximum value of the historical data divided.

FIG. 7 is a block diagram of an apparatus for predicting a demand for a CNN-based demanded electric power according to an embodiment of the present invention. As described above, the neural network-based demand power demand forecasting apparatus 100 according to an embodiment of the present invention is characterized by accurately predicting a demand for power of a consumer in a dynamically changing grid environment. The apparatus for predicting the demand for a CA based on the neural network according to an embodiment of the present invention includes a data collecting unit 110, a data processing unit 130, a model learning unit 140, a demand predicting unit 150, As shown in FIG. Referring now to FIG. 7, a description will be given of each of the configurations included in the apparatus for predicting the demand for a ground based neural network based on a demand according to an embodiment of the present invention.

The data collecting unit 110 collects daily data on weather data, past daily electricity consumption data, and past daily electricity consumption data for predicting a demand for electric power of a customer. Here, the data collection may be from an available external database or AMI. The data collected through the data collecting unit 110 includes at least one of weather data including average temperature, average humidity, average wind speed, unity, electricity, etc., past daily power use data (load profile) Date data (season, month, day of week, working day), and the like.

In addition, the data collection unit 110 may store the collected data in the database 120. As described above, the database 120 may be implemented as a physical disk capable of storing data, and may store data in a csv file format, for example.

The data processing unit 130 functions to generate a learning data set based on weather data stored in the database 120, past daily power usage data, and date data. Specifically, the data processing unit 130 may perform a function of extracting and modifying abnormal data from a database and a function of generating a learning data set based on the data stored in the database 120. To this end, a data preprocessing module (131) and a data conversion module (132).

The data preprocessing module 131 determines whether or not abnormal data exists by analyzing the data stored in the database 120 (specifically, the weather data of the past time point, the electricity use amount data and the date data of the past days). Here, the condition for determining the presence or absence of the abnormal data (i.e., for judging the validity of the data) may include, for example, whether data having a specific value among the distribution of the same time zone data of different dates exists. To this end, the data preprocessing module 131 may select parameters for abnormal data determination and correction processing and may perform the above-described validity determination process based on the selected parameters.

In addition, when abnormal data is detected, the data preprocessing module 131 can correct the abnormal data with the data within the predetermined interval based on the synchronous time data of another date or the same time of the other dates.

The data preprocessing module 131 may further update the data stored in the database 120 by storing the correction result in the database 120 after the abnormal data detection and correction process described above.

The data conversion module 132 functions to generate a learning data set based on the data stored in the database 120, which is preprocessed through the data preprocessing module 131. Specifically, the data conversion module 132 performs a conversion process in consideration of the data type (weather data, past daily power usage data and date data) of the data stored in the database 120, As shown in FIG. To this end, the data conversion module 132 normalizes the meteorological data and the past daily power usage data so as to have a value between 0 and 1 by calculating the maximum value among the past data, A learning window to be used for prediction is determined, and a learning data set can be generated based on the conversion result and the learning window described above.

As described above, the learning data set can be largely divided into a label and learning data. Here, the label indicates a normalized load profile corresponding to a period to be predicted, and the learning data in the learning data set may include meteorological data corresponding to the size of the learning window, past daily power consumption data, and date data.

The model learning unit 140 generates a depth neural network based on the learning data set generated through the data processor 130 and applies a set learning algorithm to the underlying neural network to generate a depth neural network- .

For this, the model learning unit 140 can select the learning parameters based on at least a part of the learning data sets, and can generate the underlying depth-based neural network based on the selected learning parameters. In addition, the model learning unit 140 may determine whether there is a K-fold cross validation for the training data set, and generate a verification data set composed of at least a part of the learning data sets when performing the K-fold cross validation. The learning parameters described above can be selected based on the verification data set when K-fold cross-validation is performed, and the learning parameters determined from the user can be basically selected when K-fold cross-validation is not performed.

In addition, the model learning unit 140 generates an underlying neural network based on the selected learning parameters, and generates preliminary learning through RBM (Restricted Boltzmann Machine) or by using ReLU (Rectified Linear Unit) We can generate neural network based power demand forecasting model.

The demand predicting unit 150 predicts the power demand of the customer on the basis of the hypothetical neural network based power demand prediction model generated through the model learning unit 140. Specifically, the demand forecasting unit 150 is a depth-based neural network-based power demand forecasting model generated through the model learning unit 140. The demand forecasting unit 150 uses the data collected until the prediction time point, Predict power demand. Since the predicted power demand is normalized, it is multiplied by the maximum value of the historical data divided.

In addition, the apparatus and method for predicting the demand of a customer based on the neural network of the present invention can be applied to the prediction of renewable energy generation amount in addition to the above method. For example, in the data processing unit 130 described above, a method of predicting renewable energy generation amount such as solar heat, wind power generation, etc. by changing the configuration of the learning data set can be considered. In addition, prediction accuracy can be improved by adding a mechanism for automatically determining the network structure to the model learning unit 140 in the future.

As described above, an optimal embodiment has been disclosed in the drawings and specification. Although specific terms have been employed herein, they are used for purposes of illustration only and are not intended to limit the scope of the invention as defined in the claims or the claims. Therefore, those skilled in the art will appreciate that various modifications and equivalent embodiments are possible without departing from the scope of the present invention. Accordingly, the true scope of the present invention should be determined by the technical idea of the appended claims.

100: Demand forecasting system based on in-depth network
110: Data collecting unit 120: Database
130: Data processing unit 131: Data preprocessing module
132: data conversion module 140: model learning unit
150: Demand forecasting section

Claims (18)

A data collecting unit for collecting daily data on meteorological data, past daily electricity use data and past daily electricity use data for predicting a demand for electric power of a customer;
A data processing unit for generating a learning data set based on the weather data, the past daily power usage data, and the date data;
A model learning unit for generating a depth neural network based power demand prediction model by generating an underlying neural network based on the learning data set and applying a set learning algorithm to the underlying neural network; And
And a demand predicting unit for predicting the power demand of the customer on the basis of the hypothetical network-based demand power demand prediction model. The method according to claim 1,
The depth-based neural network-based power demand forecasting model is generated by prior learning through a Restricted Boltzmann Machine (RBM) or by using ReLU (Rectified Linear Unit) as an active function. Device. The method according to claim 1,
Wherein the model learning unit selects learning parameters based on at least a part of the learning data sets and generates a base NN based on the selected learning parameters. The method of claim 3,
Wherein the model learning unit determines whether there is a K-fold cross validation for the training data set, generates a verification data set comprising at least a part of the learning data sets when performing K-fold cross validation, Wherein the power demand prediction device is based on a neural network. The method according to claim 1,
Wherein the data processing unit determines a learning window used to determine a period of past data used in the learning among past data. 6. The method of claim 5,
Wherein the learning data set is divided into a label and learning data, the label indicates a load profile corresponding to a period to be predicted, and the learning data includes meteorological data corresponding to the size of the learning window, Data based on at least one of the at least two types of data. The method according to claim 1,
Wherein the data processing unit includes a data preprocessing module for determining whether there is abnormal data by analyzing the weather data, the past daily power usage data, and the date data,
Wherein the data preprocessing module corrects the abnormal data based on data of a past time point more than the abnormal data when there is abnormal data. The method according to claim 1,
Wherein the meteorological data includes at least one of an average temperature, an average humidity, an average wind speed, a singlet, and a circulation rate. The method according to claim 1,
Wherein the base layer NN comprises an input layer, a hidden layer, and an output layer, and wherein the input layer, the hidden layer, and the output layer each comprise artificial neurons. Collecting the date data for the electricity demand data of the past, the past electricity consumption data for the past and the electricity consumption data for the past by the data collection unit for the prediction of the electric power demand of the customer;
Generating a learning data set based on the weather data, the past daily power usage data, and the date data by a data processing unit;
Generating a neural network based on the learning data set by the model learning unit and applying the set learning algorithm to the basic neural network; And
And estimating power demand of the customer based on the depth network-based customer demand forecast model by the demand forecasting unit. 11. The method of claim 10,
Wherein the neural network-based demand power demand prediction model is generated by pre-learning through RBM or by using ReLU as an active function. 11. The method of claim 10,
The method of claim 1, wherein the generating the deep-
Selecting a learning parameter based on at least a portion of the learning data sets; And
And generating an underlying neural network based on the selected learning parameters. ≪ RTI ID = 0.0 > 11. < / RTI > 13. The method of claim 12,
The method of claim 1, wherein the generating the deep-
Determining whether there is a K-fold cross-validation of the learning data set; And
Generating a set of verification data comprising at least a portion of the learning data sets when performing K-fold cross validation, wherein the learning parameters are selected based on the verification data set. Prediction method. 11. The method of claim 10,
Wherein the step of generating the learning data set comprises determining a learning window used to determine a period of historical data used in the learning among past data. 15. The method of claim 14,
Wherein the learning data set is divided into a label and learning data, the label indicates a load profile corresponding to a period to be predicted, and the learning data includes meteorological data corresponding to the size of the learning window, Data based on the power consumption of the network. 11. The method of claim 10,
Determining whether the abnormal data exists by analyzing the weather data, the past daily power consumption data, and the date data before the step of generating the learning data set; And
And correcting the abnormal data based on data of a past time point more than the abnormal data when abnormal data exists. 11. The method of claim 10,
Wherein the meteorological data includes at least one of an average temperature, an average humidity, an average wind speed, a singular value, and a circulation rate. 11. The method of claim 10,
Wherein the base layer neural network is comprised of an input layer, a hidden layer and an output layer, and wherein the input layer, the hidden layer and the output layer each comprise artificial neurons.

Images