KR20190129675A - System and Method for Energy Forecasting using Directional Cost Function - Google Patents

Abstract

The present invention relates to an apparatus and method for energy demand prediction using a directional cost function to apply the directional cost function to a deep learning model to efficiently predict the energy demand of buildings, such as buildings or factories. According to the present invention, the apparatus comprises: a module generation unit for generating a structure of a deep learning based demand prediction model; an omni-directional calculation unit which performs omni-directional calculation while inputting data according to the structure of the model according to time; a cost calculation unit which calculates the difference by calculating the difference between a result value and an actual value of the omni-directional calculation unit by applying the directional cost function; a backward learning unit for performing backward learning using a cost value calculated in the cost calculation unit; and a demand prediction model unit which generates a forward demand model by repeating forward calculation, cost calculation, and backward learning until the cost value converges to a constant value.

Description

System and Method for Energy Forecasting using Directional Cost Function

The present invention relates to energy demand prediction, and more specifically, an apparatus for predicting energy demand using a directional cost function to efficiently predict the energy demand of a building or a building by applying a directional cost function to a deep learning model, and It is about a method.

As the problems of climate change, such as global warming, have become more severe worldwide, in December 2015, ministers from 195 countries met in Paris to sign the Paris Convention on Greenhouse Gas Emissions.

Accordingly, the demand for renewable energy or efficient use of energy, rather than the discovery of fossil fuels, is increasing, aiming to reduce 37% of the emission forecast (BAU) by 2030.

An energy management system is a representative means of energy efficient use, and existing vendors are trying to introduce artificial intelligence based prediction technology to increase the savings efficiency.

A typical artificial intelligence service that can be considered in an energy management system is energy demand forecasting.

In particular, electric energy is an indispensable public good in modern society.

Electricity is produced and consumed at the same time, so it is possible to operate the equipment smoothly only when it is possible to predict the amount of electricity used and to supply it.

In particular, the prediction of the energy demand of buildings is very important given that the energy used in buildings in modern society is a big part of the total energy consumption.

As the building energy consumption continues to increase, the importance of demand forecasting as well as building energy management is emerging.

Accurate forecasting of building energy demands can lead to an optimized plan for hourly or daily building supply and demand.

However, building energy demand prediction is difficult because it is affected by various factors such as weather conditions, building materials, and building air conditioning systems.

1A and 1B are diagrams illustrating a model for predicting general energy demand and a demand forecast result.

In the prior art models for predicting energy demand, as shown in FIGS. 1A and 1B, the prediction performance decreases as the error increases when one or more energy demand predictions occur.

In addition, when there are a large number of buildings to be managed, the daily demand forecasting puts a large load on the server, and thus, a technology development for a long-term power prediction model of one or more days is required.

And because it reacts sensitively to noise (power outage, overload) data, the development of technology to filter such noise is required.

2A and 2B are diagrams illustrating an error accumulation problem of a model for predicting energy demand in the prior art.

As shown in FIG. 2A, the model for predicting energy demand in the prior art uses the predicted value of power demand as an input for the next time when predicting energy demand, and thus an error accumulates over time.

In addition, since the loss function used to reduce the error is reflected only on the amount of error, it is not known whether the error occurs along the pattern or is opposite to the pattern.

Therefore, although the direction of the pattern is different as shown in FIG. 2B, since the amount of error is the same, there is a problem of recognizing the same error.

In order to solve these problems, it is required to develop new technology for predicting building energy demand that can increase the accuracy of prediction.

Republic of Korea Patent No. 10-1168153 Republic of Korea Patent Publication No. 10-2016-0130023 Republic of Korea Patent Publication No. 10-2015-0125171

The present invention is to solve the problem of the conventional energy demand prediction technology, by using the directional cost function to apply the directional cost function to the deep learning model to efficiently predict the energy demand of buildings, such as buildings or factories It is an object of the present invention to provide an apparatus and method for predicting energy demand.

The present invention adds sensitivity to the model using Mean Absolute Percentage Error (MAPE) and predicts it using the directional cost function using Mean Directional Accuracy (MDA) to reflect the direction of learning. It is an object of the present invention to provide an apparatus and method for predicting energy demand using a directional cost function to increase accuracy.

The objects of the present invention are not limited to the above-mentioned objects, and other objects that are not mentioned will be clearly understood by those skilled in the art from the following description.

An apparatus for predicting energy demand using a directional cost function according to the present invention for achieving the above object includes a model generator for generating a structure of a deep learning-based demand forecasting model; An omnidirectional computing unit for performing the omni-directional calculation; Cost calculation unit for calculating the error by calculating the difference between the result value and the actual value of the omnidirectional calculation unit by applying the directional cost function; Reverse using the cost value calculated in the cost calculation unit Reverse learning unit for performing the learning; Demand prediction model unit for generating a demand prediction model by repeating forward calculation, cost calculation, backward learning until the cost value converges to a constant value; characterized in that it comprises a; .

Where the directional cost function is

으로 정의되고, n은 전체 데이터의 개수, t는 시간, A_t는 실제 값, F_t는 예측 값인 것을 특징으로 한다.

N is the total number of data, t is the time, A _t is the actual value, F _t is a prediction value.

The actual value is the collected real data, and the predicted value is the value predicted by the demand forecasting model, and the predicted value is the error that is derived by using the directional cost function. It is characterized by the value that is derived after repeated until.

The omni-directional calculation in the omni-directional calculation unit is an operation for calculating the node of the hidden layer in the input layer of the demand prediction model for energy demand prediction and the calculation of the node of the output layer in the hidden layer. Learning is characterized in that the process of updating each error to the node of the hidden layer when the error is calculated through the omni-directional operation.

In order to generate the demand forecast model, the demand forecast model unit adds sensitivity to the demand forecast model using Mean Absolute Percentage Error (MAPE), and average directivity accuracy (MDA, Mean) to reflect the direction of learning. Directional Accuracy).

Mean Absolute Percentage Error (MAPE) is used, and the sensitivity and direction are given to the cost function by giving a penalty by checking whether the actual value and the predicted value have similar directions.

The method for energy demand prediction using the directional cost function according to the present invention for achieving another object comprises the steps of generating a structure of the deep learning-based demand prediction model in the model generator; Comprising the omni-directional operation while inputting the data; Computing the difference between the actual value and the result value of the omni-directional calculation unit by applying the directional cost function in the cost calculation unit; Cost calculated by the cost calculation unit Performing backward learning in the backward learning unit using the value; finally generating the demand forecasting model by repeating forward calculation, cost calculation, and backward learning until the cost value converges to a constant value in the demand prediction model unit. It characterized in that it comprises a step.

The apparatus and method for energy demand prediction using the directional cost function according to the present invention has the following effects.

First, the directional cost function is applied to the deep learning model to efficiently predict the energy demand of buildings such as buildings or factories.

Second, add sensitivity to the model using Mean Absolute Percentage Error (MAPE), and predictive accuracy using a directional cost function using Mean Directional Accuracy (MDA) to reflect the direction of learning. Can increase.

1A and 1B are diagrams illustrating a model for predicting general energy demand and a demand forecast result.
2a and 2b is a configuration diagram showing the error accumulation problem of the model for energy demand prediction in the prior art
3 is a block diagram of a model for energy demand prediction according to the present invention
4 is a detailed configuration diagram of an apparatus for predicting energy demand using a directional cost function according to the present invention;
5 is a flow chart illustrating a method for energy demand prediction using a directional cost function in accordance with the present invention.
6 and 7 are graphs for predicting the energy demand of the building for three days using the existing cost function and the cost function according to the present invention

Hereinafter, a preferred embodiment of the apparatus and method for energy demand prediction using the directional cost function according to the present invention will be described in detail.

Features and advantages of the apparatus and method for energy demand prediction using the directional cost function according to the present invention will become apparent from the following detailed description of each embodiment.

3 is a block diagram of a model for predicting energy demand according to the present invention, and FIG. 4 is a detailed block diagram of an apparatus for predicting energy demand using a directional cost function according to the present invention.

The model for energy demand prediction according to the present invention solves the problem of accumulating errors by using the seq2seq model structure which is mainly used for language translation, and correlates the correlation between the power consumption time of the encoder and the power demand prediction time of the decoder. It can include constructs that learn directly.

In particular, the present invention is to apply the directional cost function to the deep learning model to efficiently predict the energy demand of buildings, such as buildings or factories, the sensitivity to the model using the Mean Absolute Percentage Error (MAPE) In addition, the directional cost function using Mean Directional Accuracy (MDA) can be used to increase the prediction accuracy to reflect the direction of learning.

Deep learning-based energy prediction models are mostly based on cyclic neural networks capable of learning time series data.

Using this model, the performance can be improved to some extent compared to the existing machine learning based model, but the accuracy is low for the actual industry.

To improve performance, there are ways to learn a large amount of structured data accumulated over time and to improve the model.

In the first method, it takes a long time to accumulate energy usage data for the initial application of the building.

Therefore, the second method should be approached to improve the model, and in the present invention, the prediction function is improved by improving the cost function which has an important effect when the energy demand prediction model is trained.

Most existing energy demand prediction models use root mean square error (RMSE) as a cost function when learning.

Using this method, we can obtain a model with the pattern of the form that is required for the prediction, but since the actual energy use pattern is affected by various variables, the pattern of the form can only get the minimum accuracy according to the shape. have.

Therefore, we need a model that can respond sensitively to these shapes and a new type of cost function rather than the mean square root error method.

In the present invention, the directional cost function using Mean Directional Accuracy (MDA) to add sensitivity to the model using Mean Absolute Percentage Error (MAPE) rather than square root error and reflect the direction of learning. To solve the problem.

4 illustrates a configuration for generating an energy demand prediction model. The model generator 110 generates a structure of a deep learning-based demand prediction model, and performs omni-directional calculation while inputting data according to the structure of the model according to time. An omnidirectional calculator 120 and a cost calculator 130 that calculates an error by calculating a difference between the actual value and the resultant value of the omnidirectional operator 120 by applying a directional cost function, and a cost calculator ( The backward learning unit 140 performs backward learning using the cost value calculated in 130), and the forward prediction, cost calculation, and reverse learning are repeated until the cost value converges to a constant value. Demand prediction model unit 150 for generating a.

Here, the directional cost function is defined by Equation 1 below.

n is the total number of data, t is the time, A _t is the actual value, F _t is the prediction value.

Instead of using the mean square root error, you use a mean absolute percentage and see if the actual and predicted directions are similar. Thus, sensitivity and direction can be given to the cost function.

Here, the actual value is the actual data collected, and the predicted value is a value predicted by the model, and Equation 1 is used to calculate an error after the omnidirectional operation is completed.

The predicted value is a value that is obtained after updating the error derived through Equation 1 to the model through backward learning and repeating it until a certain value is reached.

In addition, the model generator 110 uses a model for learning time series data mainly used in deep learning, and typically, RNN-LSTM (Recurrent Neural Networks-Long Short Term Memory) may be used, but is not limited thereto.

In addition, the omni-directional calculation in the omni-directional calculation unit 120 is an example of a model for predicting energy demand. For example, the RNN-LSTM model calculates a node of a hidden layer in an input layer and a node of an output layer in a hidden layer. The operation to do.

In the basic model structure as shown in FIG. 2A, Layer1 is an input layer, Layer2 is a hidden layer, and the uppermost y-value is an output layer.

The y value means a value predicted by the model, and when the omnidirectional operation is performed, the predicted y value is obtained and the difference is calculated from the actual value to calculate the error, and the directional cost function according to the present invention is obtained. Is used.

In addition, 'reverse learning' is a process of updating each error to a hidden layer node when an error is calculated through an omnidirectional operation. This is called backpropagation.

In addition, the 'constant value' in the process of finally generating the demand forecast model by repeating forward calculation, cost calculation, and reverse learning until the cost value in the demand forecast model unit 150 converges to a constant value is represented by the model. Predictive error rate, which is empirically set by the user.

Input of the demand forecast model is an essential element, which can include power consumption and additional environmental information (temperature, humidity, etc.). When the input values enter the demand forecast model, it is possible to derive the predicted value through omni-directional calculation and use this value as the actual power. Compare the values to find the error, then update the model through backward learning.

As such, this process is repeated until the predicted value reaches a predetermined value determined by the user, and finally a demand forecasting model is generated.

The energy demand prediction of the device for energy demand prediction using the directional cost function according to the present invention having such a configuration proceeds as follows.

5 is a flow chart illustrating a method for energy demand prediction using a directional cost function in accordance with the present invention.

In the method for energy demand prediction using the directional cost function according to the present invention, the model generator 110 generates a structure of the deep learning based demand prediction model (S201).

Subsequently, the omni-directional calculation unit 120 performs omni-directional calculation while inputting data according to the time according to the structure of the model.

In addition, the cost calculator 130 calculates an error by calculating a difference between the actual value and the resultant value of the omnidirectional calculator 120 by applying the directional cost function (S203).

Subsequently, the backward learning unit 140 performs backward learning using the cost value calculated by the cost calculating unit 130 (S204).

Then, the demand forecasting model unit 150 generates a demand forecasting model by repeating forward calculation, cost calculation, and backward learning until the cost value converges to a constant value (S205).

6 and 7 are graphs for predicting the energy demand of the building for three days using the existing cost function and the cost function according to the present invention.

FIG. 6 is a result of prediction using a mean square root error in a cyclic neural network, and FIG. 7 is a result of applying a directional cost function to a cyclic neural network.

The prediction error rate results of the two graphs are shown in Table 1.

As can be seen from Table 1, the difference in the mean error rate for the three-day prediction value of the two models is 3.68, indicating a large difference.

In addition, when the directional cost function is used, the predicted error rate of day 1 is reduced by about 2 times, and the error rates of day 2 and day 3 remain similar, while the error rate increases with time when the mean square root error is used. You can see that it grows.

The apparatus and method for energy demand prediction using the directional cost function according to the present invention described above are to apply the directional cost function to a deep learning model to efficiently predict the energy demand of buildings, such as buildings or factories.

To do this, we add sensitivity to the model using Mean Absolute Percentage Error (MAPE), and predict it using the directional cost function using Mean Directional Accuracy (MDA) to reflect the direction of learning. This is to increase the accuracy.

It will be understood that the present invention is implemented in a modified form without departing from the essential features of the present invention as described above.

Therefore, the described embodiments should be considered in descriptive sense only and not for purposes of limitation, and the scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the equivalent scope are included in the present invention. It should be interpreted.

110. Model generator 120. Omnidirectional calculator
130. Cost calculator 140. Reverse learner
150. Demand prediction model

Claims (9)

A model generator for generating a structure of a deep learning based demand prediction model;
An omnidirectional calculator configured to perform omni-directional calculation while inputting data according to time according to the structure of the model;
A cost calculator for calculating an error by calculating a difference between a result value and an actual value of the omnidirectional calculator by applying a directional cost function;
A reverse learning unit performing backward learning using the cost value calculated by the cost calculating unit;
Energy demand forecasting using a directional cost function, comprising: a demand forecasting model unit that generates a demand forecasting model by repeating forward calculation, cost calculation, and backward learning until the cost value converges to a constant value. Device for. The method of claim 1, wherein the directivity cost function

Defined as
and n is the total number of data, t is the time, A _t is the actual value, and F _t is the predictive value. The method of claim 2, wherein the actual value is actual data collected, and the predicted value is a value predicted by the demand forecasting model.
The predicted value is an energy demand prediction using the directional cost function, which is a value obtained by updating the error derived using the directional cost function to the demand prediction model through backward learning and repeating it until a certain value is reached. Device for. The method of claim 1, wherein the omni-directional calculation in the omni-directional calculation unit is an operation of calculating a node of a hidden layer in an input layer of a demand prediction model for energy demand prediction and a node of an output layer in a hidden layer
Backward learning in the backward learning unit is an apparatus for predicting energy demand using the directional cost function, characterized in that each error is updated to the node of the hidden layer when the error is calculated through the forward operation. The method of claim 1, wherein in order to generate a demand forecast model in the demand forecast model unit,
Add sensitivity to demand forecast models using Mean Absolute Percentage Error (MAPE),
Apparatus for predicting energy demand using the directional cost function, characterized by using Mean Directional Accuracy (MDA) to reflect the direction of learning. 6. The directional cost according to claim 5, wherein a mean absolute percentage error (MAPE) is used to determine whether the actual value and the predicted value are similar to each other, and a penalty is given to the cost function to give sensitivity and direction. Device for forecasting energy demand using a function. Generating a structure of the deep learning based demand prediction model in the model generator;
Performing omni-directional calculation by inputting data according to time according to the structure of the model in the omni-directional calculation unit;
Calculating an error by calculating a difference between a result value and an actual value of the omnidirectional calculator by applying the directional cost function in the cost calculator;
Performing backward learning in the backward learner using the cost value calculated in the cost calculator;
Generating a demand forecast model by repeating forward calculation, cost calculation, backward learning until the cost value converges to a constant value in the demand forecast model unit; energy using a directional cost function Method for forecasting demand. 8. The method of claim 7, wherein the directivity cost function is

Defined as
n is the total number of data, t is the time, A _t is the actual value, F _t is the predicted value method for energy demand prediction using the directional cost function. 8. The method of claim 7, in order to generate a demand forecast model,
Add sensitivity to demand forecast models using Mean Absolute Percentage Error (MAPE),
Method for predicting energy demand using directional cost function, characterized by using Mean Directional Accuracy (MDA) to reflect the direction of learning.

Images