Prediction of Daily Electricity Use in Residential High-Rise Buildings

Using Artificial Neural Networks

Ahmad Rofii and Hanif Ibrahim

Universitas 17 Agustus 1945 Jakarta, Indonesia

Keywords: JST, Backpropagation Method, MATLAB, MSE, MAPE.

Abstract: The high demand for electrical energy from consumers requires producers to provide a reliable but economic

supply of electrical energy. Therefore, strategies and methods are needed to match the power generation and

demand. This can be achieved by planning a good and proper operation. One of the important steps in planning

the operation of the electric power system is predicting the need for electrical loads. One method of predicting

electrical loads is to use ANN (Artificial Neural Networks). ANN is an information processing system with

characteristics similar to biological neural networks. This method uses ANN with a backpropagation

algorithm, and the prediction results are obtained by adding electrical load data (KW) for the selected similar

days. ANN processing using MATLAB software. The artificial neural network architecture uses 15 input

layers, 15 output layers, and ten hidden layers, and the activation function used is logsig and purelin. Logsig

for hidden layers and purelin for output layers. The results of the electrical load prediction using an artificial

neural network with the backpropagation method, the Mean Square Error (MSE) value of network training is

0.1, and the MAPE value of data testing is 6.5%. The results of the prediction of electricity use in high-rise

residential buildings in February 2023 are predicted.

1 INTRODUCTION

One of the types of energy that propagates through

the cable network. electricity, has played an essential

role in the progress of human civilization in various

fields. The use of electricity will be affected by

increased human activities. Electricity providers must

provide enough electricity to meet high consumer

demand. Second type will cause obesity and lack of

physical activity. The electricity sector is considered

to be a field that requires long-term forecasts so that

the power plant infrastructure is ready to be supplied

with electricity. However, long predictions are

difficult to achieve. Timeframes and financing factors

are often the obstacles faced. Therefore, to anticipate

the events mentioned above, it is necessary to make

projections to estimate how much electrical energy

will be consumed.

Artificial intelligence software is now being

developed due to computational advances to create

alternative techniques for long-term electrical energy

forecasting. In addition to being easier to use,

computational innovations result in more accurate

findings. Experts are working to develop an artificial

intelligence system that can estimate future electrical

energy needs.

Artificial neural networks are one the most

effective intelligent systems in making predictions.

Artificial neural networks are used to predict the use

of electricity that will be used in buildings. in this

way we can predict the use of electricity in a building

2 LITERATURE REVIEW

Research conducted by Fathur Rohman et al. (2021)

with the title "Electrical Load Prediction Using

Artificial Neural Network Method Backpropagation".

This study used the method of predicting electrical

loads using JST (Artificial Neural Network)

Backpropagation. The study

has indicated electrical loads using artificial neural

networks backpropagation method of the largest

MAPE (Mean Absolute Percentage Error) value

obtained with a value of 4.32 %. And the smallest

MAPE value is obtained with a value of 2.71 %.

(Rohman, 2022).

Research conducted by Yuan Octavia et al. (2018)

with the title "Study of electrical load forecasting

Roﬁi, A. and Ibrahim, H.

Prediction of Daily Electricity Use in Residential High-Rise Buildings Using Artiﬁcial Neural Networks.

DOI: 10.5220/0011980100003582

In Proceedings of the 3rd International Seminar and Call for Paper (ISCP) UTA â

Z45 Jakarta (ISCP UTA’45 Jakarta 2022), pages 293-299

ISBN: 978-989-758-654-5; ISSN: 2828-853X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

293

using artificial neural network method."In a case

study of the distribution of electrical energy in the

Mojokerto Region, this study using the Artificial

Neural Network (JST) approach with the use of

backpropagation algorithms to predict long-term

electricity needs. In this study, there were eight

variables used, with variables bound is the amount of

electricity consumption. The free variable is the sum

of population, GRDP, number of household sector

customers, number of sector customers, industry,

number of business sector customers, number of

social sector customers, number of business sector

customers, and distribution losses. According to this

analysis, the electrical load of The Mojokerto region

is predicted to grow by 22.641 percent between 2018

and 2030, an average of 1.728% per annum. (Yuan et

al., 2018).

Research conducted by Diah Setyowati and Said

Sunardiyo (2020) under the title "Forecast of

Electrical Energy Needs with Artificial Neural

Networks (Artificial Neural Network)

Backpropagation Method 2020-2025". With using the

Artificial Neural Network Backpropagation

technique using MATLAB software, this research

predicts the electrical energy needs of PT PLN 7

(Persero) UP3 Semarang in 2020–2025. The study

resulted in a growth annual of a total percent (GOT%)

of 2.7% and the average percentage of errors absolute

(MAPE) of 0.4%. (Setyowati & Sunardiyo, 2020).

2.1 Theoretical Basis

2.1.1 Predictions

Forecasting is the practice of predicting events or

items in the future (Jay, 2009). Three prediction

categories depending on the time frame that can be

made: short-term, medium-term, and long-term

(Aryan Hamidie, 2009). Hourly, daily, and three-

month periods are included in the short-term

forecasts. Predictions for the medium term or medium

period often range from three months to two years.

Long-term predictions are generally for two-year

planning or more (Sugiarto & Harijono, 2000). To

foresee events that are not desirable and get ready to

take the necessary actions and predictions required

(Arifah et al., n.d.). Although it is difficult to foresee

the future, Forecasts can be used as a guide to reduce

errors. Achievement of long-term goals in the

installation of production control systems

(production) and allocation of power lines require

accurate projections, in particular for energy

producers.

2.1.2 Electrical Energy

Energy from natural resources is converted into

electrical power by generating electricity. According

to the resources used, Power plants are classified as

PLTU, PLTA, PLTN, PLTS, and others. Generators

in power plants convert mechanical energy into

electricity using the concept of conductors and

electric fields. The energy produced will be stored in

accumulators or energy storage devices.

Transmission lines will channel the electrical energy

that has been kept to the customer. Based on the

voltage value, shape, the type of conductor used, line

arrangement, and circuit array, transmission and

distribution lines Classified. However, the alternating

current voltage transmission line of the 3-phase and

1-phase is usually used to distribute electricity from

power plants. However, some transmission lines also

use direct current voltage.

2.1.3 Electric System Power

Electrical power components, such as electric power

production, transmission systems, and distribution

systems, form the electric power system. Third, these

components comprise most of the power grid that

transports electricity from the production station to

the load station. Transmission lines connect

production stations to distribution systems and,

through interconnections, to other power systems. All

individual loads are connected to the transmission

line through a distribution system at substations that

handle voltage conversion and switching operations.

Figure 2.1 below shows the power system circuit

electricity (Eirene & Sau, 2019).

The electrical installation system in high-rise

buildings is divided into utility lines, lighting lighting,

air conditioning channels and stop contact channels.

How to find out the use of electrical power in each

room of the building is to place a measuring

instrument on the main panel as well as in every room

that is considered a tenant or occupant. To identify the

use of electrical power in each room, a set of

measuring instruments is installed starting from

measuring power, current, voltage, and cos phi.

Recording is carried out in accordance with

operational standards determined by the building

management. the relationship between current,

voltage, power based on the following formula

sebagai berikut

P=√3 V_p I_p cos∅

Where Vp is the phase voltage, Ip is the phase

current and cos∅ is the power factor.

In the electrical system of residential buildings,

the parameters measured are single-phase parameters,

so to determine and predict the use of electrical power

ISCP UTA’45 Jakarta 2022 - International Seminar and Call for Paper Universitas 17 Agustus 1945 Jakarta

294

in residential buildings, it is necessary to carry out the

results of phase measurements into 3 phase

calculations.

2.1.4 Artificial Neural Network

The functioning system of neural networks can be

compared with the human brain. Many neurons are

found in the neural network, which is connected. The

information obtained through the outgoing

connection will be altered by these neurons and sent

to different contacts. Report (called input). It will

have a certain arrival weight when it is delivered to

neurons. By summing all the weight values in this

input, a particular propagation function will process

them. What will then compare the results with a

specific threshold value using the activation function

of each neuron? The neuron is not triggered if the

input drops below the threshold value, but if it is, it is

involved and transmits the output to all neurons

connected to it with its output weight (Kusumadewi,

2004)

Figure 1: Activation Function on Simple Neural Networks

In the figure, a neuron will process N inputs (x

x2, ..., x

) which each have weights w

, w

, ..., wN,

and bias weight b, with the formula :

𝑎 

∑





𝑥



𝑤



Then the activation function F will activate into

the output of the network y.

2.1.5 Backpropagation Algorithm

According to (Kusumadewi, 2004), Perceptrons with

multiple layers often use backpropagation-guided

learning techniques to modify the weights that

connect neurons in the hidden layer. The

backpropagation method adjusts the backward weight

value by using the error output. What must complete

the forward propagation stage first to get this error.

Neurons are triggered using the differential activation

function as propagation progress.

2.1.6 Backpropagation Architecture

Artificial neural network-based backpropagation or

propagation feedback consists of many units in one or

more hidden layers. The artificial neural network base

in the feedback propagation architecture drawing

contains n inputs (plus bias), hidden layers with p

units (plus bias), and m output units (Prasetyo &

Sahala, 2014).

3 RESEARCH METHODOLOGY

The research methods used in the preparation of this

final project report are:

a. Literature studies examine the necessary

theories from handbooks that support and

relate to the themes taken to be used as

theoretical foundations.

b. Discussion, namely conducting a question and

answer with supervisors and technology in the

field and friends (University of August 17,

1945 Jakarta).

c. Observation Method, plunge directly into the

field to study the selected object.

d. Perform calculations and analyzes.

3.1 Flowchart Research

This project research method explains the estimated

consumption of electrical energy. The preparation of

this final project can be seen in the flowchart.

Figure 2: Research Flowchart

∑

Prediction of Daily Electricity Use in Residential High-Rise Buildings Using Artiﬁcial Neural Networks

295

4 RESULTS AND DISCUSSION

What can analyze the estimated electrical energy

consumption for the next month based on connected

power (VA) and the amount of energy (kW). What

will discuss this regarding the results of estimating

electrical energy consumption using the artificial

neural network method of backpropagation.

If any, should be placed before the references

section without numbering.

Table 1: Three-phase data

The neural network training process for artificial

backpropagation requires several parameters,

including the number of hidden layers, epoch, error

goals, and learning rate. Parameter changes made

during training are the number of hidden layers 10,

starting from the number of epochs 4000, 6000, and

8000, the number of learning rates starting from

0.001, 0.01,0.1, and for error goals 0.001. This is done

to obtain good training results. The training data used

for the training are connected power data (VA) and

amount of energy (kW) from January 1, 2021 to

January 30, 2021 and training target data, namely

connected power data (VA) and energy amount

(kWh) from January 16 to January 31. Before the data

is entered into the artificial neural network, the data

must be normalized in the range [0 to 1] because the

input data of the artificial neural network uses the

logsig activation function (binary sigmoid). To

convert the original data into normalization data using

a formula.

Min – max normalization:

𝑋





𝑋𝑚𝑖𝑛 𝑋

𝑟𝑎𝑛𝑔𝑒𝑋



𝑋𝑚𝑖𝑛 𝑋



𝑋



𝑚𝑖𝑛 𝑋

𝑋





𝑋𝑚𝑒𝑎𝑛 𝑋

𝑆𝐷𝑋

Table 2: normalization data.

Backpropagation Artificial Neural Network

Program for electric power load forecasting using 15

input and 15 target data. It consists of 15 input units

and ten layers on the hidden layer and 15 units on the

output layer. The activation function of the input unit

to the hidden layer is a binary sigmoid (logsig) and

purelin.

Figure 3: Network creation toolbox

ISCP UTA’45 Jakarta 2022 - International Seminar and Call for Paper Universitas 17 Agustus 1945 Jakarta

296

Figure 4: Best training and gradient with learning rate 0.1

and purelin activation function

Figures 3 and 4 show the best training and

gradients using ten hidden layers, learning rate 0.1

and binary sigmoid activation functions (logsig) and

purelin. In training with the best iteration with

conditions without isolation using the activation

function of binary sigmoids (logsig) and purelins,

binary sigmoids with a range of 0 to 1 while purelin

is a function of linearizing each input into its linear

value. The linear process will produce an input value

equal to the output value (y = x). These two activation

functions are commonly used for artificial neural

networks trained using the backpropagation method.

The identity activation function returns an MSE value

close to reaching the target. This is affected because

the input of the activation function is the same as the

output.

Figure 5: Regression of daily electricity use of high-rise

residential buildings

Figure 5 is a Regression of training results using

Hidden Layer 10, epoch 8000, and Learning Rate 0.1.

With a regression value of 0.91, the degree of

proximity between the training target and the training

result is very close.

Figure 6: Training Parameters

Figure 6 is the parameter used to perform JST

testing with the backpropagation method. The

parameters used are epochs 8000, learning rate 0.1,

min_grad 10-9, and max_fail 7000. Comparison table

and errors that what can obtain in the prediction of

daily electricity use in high-rise buildings: Based on

table 3, we can see that the results of predicting the

use of electricity in high-rise buildings calculate the

accuracy of the forecasting/prediction that has been

carried out, namely using calculations and analysis of

the Mean Absolute Percentage error (MAPE). Mean

Absolute Percentage error (MAPE) is the percentage

of the average error absolutely (absolute). The

definition of Mean Absolute Percentage Error is a

statistical measurement of the accuracy of forecasts

(predictions) on forecasting methods. The wider

community can use measurement using Mean

Absolute Percentage Error (MAPE) because MAPE

is easy to understand and apply in predicting

forecasting accuracy. The Mean Absolute Percentage

Error (MAPE) method provides information on how

much the forecasting error is compared to the actual

value of the series. The smaller the percentage error

value on MAPE, the more accurate the forecasting

results will be.

The Mean Absolute Percentage Error (MAPE) value

is analyzed:

Prediction of Daily Electricity Use in Residential High-Rise Buildings Using Artiﬁcial Neural Networks

297

Table 3: Comparison of electric power load forecasting in

high-rise residential buildings

Table 4: Range Mean Absolute Percentage Error (MAPE)

From the table above, we can understand the

range of values that show the meaning of the error

percentage value on MAPE, where the MAPE value

can still be used if it does not exceed 50%. If the

MAPE value is above 50%, then the forecasting

model cannot be used.

The calculation of the MAPE method is as follows:

𝑀𝐴𝑃𝐸 

∑

1







𝐴



𝐹



𝐴



100

𝑛

Information

= Actual request to t

= forecasting result to t

N = magnitude of forecasting data

Where there is an absolute symbol in the MAPE

formula indicating that the negative value of the

calculation result will remain positive.

Table 4: Range Mean Absolute Percentage Error (MAPE)

𝑀𝐴𝑃𝐸 

0.972491563

 100  6.483 %

Based on the calculation results, it can be seen that

the MAPE value is 6.5%. This shows that the

prediction of electricity use in residential buildings

using the JST Backpropagation method has excellent

forecasting capabilities.

Table 5: Results of Predicted Electricity Consumption of

Residential High-Rise Building

From the results of the predicted data, it can be

concluded that the network can study the dispersion

of data based on 30 data used and 560 data patterns

used. The network can also find the optimal solution

to minimize cost functioning. The result of prediction

or forecasting for 15 days in the coming month is a

solution in this case, which is useful for

understanding the peak load in the upcoming month

so that you can determine how much energy will be

used.

5 CONCLUSIONS

Based on the training results using several

parameters, the best results for the testing and

prediction process were obtained, namely using

Hidden Layer 10, epoch 8000 and Learning Rate 0.1.

The results of calculating the MAPE value of 6.5%

show that the prediction of electricity use in

residential buildings using the JST Backpropagation

method has excellent forecasting accuracy

capabilities.

ISCP UTA’45 Jakarta 2022 - International Seminar and Call for Paper Universitas 17 Agustus 1945 Jakarta

298

REFERENCES

Arifah, N., Murnomo, A., & Suryanto, D. A. (n.d.).

Implementing Neural Network on Matlab for Forecast

of Electricity Load Consumption in Ponorogo Regency,

East Java.

Arya Hamidie, K. (2009). Energy coefficient method for

short-term load forecasting on the Balinese Madura

champion network. 1–10.

Eirene, H., & Sau, M. (2019). Textbook of Energy And

Electrical Power Operation ETAP Applications. CV.

Budi Utama.

Fayeldi, T., Murniasih, T. R., & Yunus, A. (2016). Basics

of Computer Programming Using MATLAB.

Jay, H. (2009). Operations Management (C. Sungkono

(ed.); 9th ed.). Salemba Four.

Khair, A. (2011). Term Electric Load Forecasting. 138.

Kuncoro, A. H., & R, D. (2005). Application of Artificial

Neural Networks for Long-Term Electric Power Load

Forecasting in Electrical Systems in Indonesia. Year

XIX, 3, 211–217.

Kusumadewi, S. (2004). Building artificial neural networks

(using MATLAB and Excel links) (F. W. Nurwiyati

(ed.); First Edition).

Marsudi, D., Generator, P., & Tobing, B. (2016). Gear

Tension crowbar. Generation of Electrical Energy, 7(1),

4–31.

Panjang, J., & Setiabudi, D. (2015). Electrical Load

Forecasting Information System. 1(1), 1–5.

Prasetyo, E., & Sahala, A. (2014). Data mining: processing

data into information using MATLAB. ANDI

OFFSET.

Rahman, A., Abdullah, A. G., & Hakim, D. L. (2012).

Long-Term Peak Load Forecast in Indonesia's

Electrical System Using adaptive neuro-fuzzy inference

system algorithm. Electrans, 11(2), 18–26.

Rohman, F. (2022). Electrical Load Prediction Using

Artificial Neural Networks Backpropagation Method.

JOURNAL OF SOLAR ENERGY, 5(2).

https://doi.org/10.32502/jse.v5i2.3092

RUPTL, P. P. 2021-2030. (2021). Electricity Supply

Business Plan (RUPTL) of PT PLN (Persero) 2021-

2030. Electricity Supply Business Plan 2021-2030,

2019-2028.

S. Haykin, S. (2009). Neural Networks and Learning

Machines. In Encyclopedia of Bioinformatics and

Computational Biology: ABC of Bioinformatics (3d

Edition, Vols. 1–3). Pearson Education, Inc.

https://doi.org/10.1016/B978-0-12-809633-8.20339-7

Setyowati, D., & Sunardiyo, S. (2020). Forecast of

Electrical Energy Needs with Artificial Neural Network

Backpropagation Method 2020-2025. In EECCIS

Journal (Vol. 14, Issue 1). https://jurnaleeccis.ub.ac.id/

Siang, J. J. (2009). Artificial Neural Network &

Programming Using Matlab (Kedu Edition). ANDI

OFFSET.

Stevenson, J. W. D. (1984). Elements of Power System

Analysis 4th Edition. In Power System Analysis (p.

365).

Sugiarto, & Harijono. (2000). Business Forecasting.

Gramedia Main Library.

Sulasno. (1993). Analysis of electric power systems (Cet.

1). Satya Discourse.

Sunardiyo, S. (2009). Load Flow Analysis Study of Electric

Power System Implementation on Electrical Network at

UNNES.

Yuan, O. D. P., Afandi, A. N., & Putranto, H. (2018).

TEKNO Journal of Electrical and Vocational

Technology The study of electrical load forecasting

using artificial neural network methods. In Department

of Electrical Engineering (Vol. 28).

http://journal2.um.ac.id/index.php/tekno

Zebua, F. Y., Mulyani, S. H., & H., M. E. (2012). Modelling

of Liver Cirrhosis Disease Detection using Artificial

Neural Networks. Sisfotenika, 2(2).

https://doi.org/10.30700/jst.v2i2.71

Prediction of Daily Electricity Use in Residential High-Rise Buildings Using Artiﬁcial Neural Networks

299