International Journal of Artificial Intelligence Trends (IJAIT) 1

Vol. 2, Issue 4; No. 43, April, 2023, pp. 453-460

DEVELOPMENT OF SHORT-TERM LOAD FORECASTING SYSTEM FOR DEMAND SITE

MANAGEMENT OF TRANS-EKULU 33/11KVA DISTRIBUTIVE TRANSFORMER USING ARTIFICIAL
INTELLIGENCE
1
Elom Hyginus, 2Ibekwe Basil
Department of Electrical and Electronic Engineering, Faculty of Engineering,
1

Enugu State University of Science and Technology (ESUT)

Article Info ABSTRACT

Received: 11/03/ 2023 This paper presents the development of an artificial intelligence-based short-term
Revised: 03/04/2023 load forecasting system for demand site management of the Trans-Ekulu
Accepted: 11/04/2023 33/11KV, 15MVA distributive transformer. Empirical data collection and
Published: 19/04/2023 analysis were performed using Newton Raphson load flow analysis to study
feeder performance, and the amount of generated power on a short-term basis (24
Corresponding Author’s hours) was estimated using a unit cost of 5.22 naira per kW. The collected data
Email: was used to train an artificial neural network-based load forecasting system,
[email protected] which was then implemented using Simulink. The system was tested through
simulation, and the results showed that the algorithm was able to accurately
Corresponding Author’s estimate the behavior of the transformer in the next 24 hours, with a regression of
Tel: 0.9989. The load forecasting system was integrated into the EEDC center for
+234 703 322 5267 managing the Trans-Ekulu feeder, and the results showed that the behavior of the
feeder was correctly estimated by the forecasting system

Keywords: Power: load Forecasting; Transformer; Artificial Intelligence; Newton Raphson; Algorithm; Feeder

1. INTRODUCTION The importance of load forecasting cannot be

Recognizing the overall amount of power overstated. It has received tremendous research focus,
consumption patterns in the future is a vital part of aiming to improve accuracy, as the merit derived
exploring and planning electrical power systems. The from the forecasting system depends on the
quality control and effectiveness of these power prediction's success. Load forecasting can be
system operations heavily depend on accurately classified into three main categories: short-term load
forecasting future demand for effective site forecasting, medium-term load forecasting, and long-
management. Thus, forecasting the overall power term load forecasting. Short-term load demand
usage per given future time accurately is a mandatory forecasting predicts for one hour or more but less
prerequisite for managing power system generation, than 24 hours, medium-term load demand forecasting
transmission, and distribution operations. Load predicts for 2 days to around 2 weeks or a month,
forecasting has been an integral part of power system while long-term load forecasting lasts up to 6 months
planning since the 19th century and has gained to a year (Ashigwuike et al., 2020).
increased research attention worldwide. Accurate Traditionally, load forecasting decisions were based
load forecasting will help power generation and on analyzing reports for a case study area. However,
distribution companies make precise unit this approach is prone to human error and is
commitment decisions regarding voltage control, unreliable. To improve performance, various models
power supply quality, load switching, network such as fuzzy logic, bank Jones algorithm, Kalman
reconfiguration, and infrastructural development. filters, expert systems, wavelet technique, and neural
Furthermore, it will ensure low power generation network were developed. These approaches have
costs and minimize unit costs for consumers, their advantages and disadvantages. However, the use
guaranteeing power system reliability. of neural networks offers better success than the
counterparts (Zina et al., 2018).
 2  +234 703 322 5267

Trans-Ekulu, a place in Enugu, Nigeria, has Maew et al. (2017) presented a study on demand site
experienced an influx of people from different parts management using artificial neural networks in a
of the state, resulting in industrial, infrastructural, and smart grid environment. The study performed power
economic growth. This population increase has system planning on the area using ANN based load
proportionately increased the area's load demand forecasting technique using artificial neural networks.
beyond the conventional supply margin. The study did not consider cost of the load estimated
Consequently, the area has experienced poor power as a complete analysis and planning. Haider et al.
supply quality, power outages, fluctuations, and other (2016) presented a review on residential demand
challenges unsatisfactory to consumers. Developing a response of smart grid. The study reviewed various
load forecasting system to predict the amount of technique for demand site management with the use
power used by the area can enable better power of descriptive analysis. They recommended the use of
system planning and improve demand site artificial intelligence technique for future works.
management. O’connel et al. (2016) presented a study on benefits
Over the years, various techniques, including and challenges of electrical demand response. The
heuristic approaches, machine learning, demand study overviewed various benefits and challenges of
response, among others, have been proposed to demand response using a qualitative analysis
develop load forecasting systems for power planning technique. Artificial intelligence technique was
(Maew et al., 2017; Nauru, 2016; Zah et al., 2015; recommended as the best approach in future.
Hana, 2018). However, machine learning has 3. MTHODOLOGY
provided better results in terms of forecasting The methodology employed for this paper involved
accuracy when compared to the other techniques. experimental investigation of the Trans-Ekulu
Machine learning is a branch of artificial intelligence 15MVA. 33/11KV feeder using load flow analysis to
that can learn by training a dataset and make accurate collect the amount of power consumed by the testbed
decisions. Studies by Zah et al. (2015) revealed that on short term bases (24Hr). The equivalent cost of the
neural network-based machine learning algorithms load was also studied. The data collected was used to
are the most effective solution for power system develop a load forecasting program using artificial
planning and load forecasting. Thus, we will adopt neural network and then implement with neural
and use this algorithm to develop a load forecasting network toolbox in MATLAB.
model for the EEDC for better demand site 3.1 Engineering Materials Used
management.
2. LITERATURE REVIEW The engineering materials used for the study are
laptop which was installed with ETAP load flow
Nauru (2016) presented a new mathematical analyzer software to read the power consumed by the
approach and heuristic methods for forecasting in feeder transformer on daily basis considering active
smart grid. The study used mathematical models to and reactive power parameters. The 15MVA, 11KV
forecast load estimate in grid using heuristic method. feeder transformer was the testbed under study. The
The result can be improved with artificial intelligence price of the load consumed on daily based by the
technique. Zah et al. (2015) presented a study on the transformer was collected from the EEDC logbook;
application of Artificial neural network for load other instruments used were power meter to measure
forecasting in smart grid. The study performed long the overall power flow in the transformer within
term load forecasting on the distribution system using 24Hr.
artificial neural network technique. Hana (2018) 3.2 Method of Data Collection
researched on short term electric load forecasting
using demand site management technique. The study The primary source of data collection is the Enugu
performs load forecasting using demand site State Electricity Distribution Company (EEDC)
management technique for short term power planning which provided the data of the testbed. The data
in the area using demand response technique. The collection considers load flow attributes such as
estimated load usage accuracy can be improved with active power consumed by the feeder and also the
ANN.

Vol. 2, Issue 4; No. 43, April, 2023, pp. 453-460

International Journal of Artificial Intelligence Trends (IJAIT) 3
Vol. 2, Issue 4; No. 43, April, 2023, pp. 453-460

equivalent cost which was collected from the account

department of the EEDC.
3.3 Investigation of the Trans-Ekulu Feeder 3.4 Development of a load forecasting system for
Using Newton Raphson Load Flow Analysis the Area
The Newton Raphson load flow analysis in Onah et The method used for the development of the load
al. (2019) which consider the bus admittance matrix forecasting scheme is the structural method. This
of the feeder and then used the parameters to method was used due to its ability to model a system
compute the Jacobian matrix of the system and in a simple but clearly define the relationships
formulate the active and reactive power flow of the between all the variables put together to achieve the
transfer for the day was used to collect data of the new system. To develop the system, the artificial
total load on the 19th to 25th March, 2022 for 24Hr of neural network algorithm was adopted from Arjun et
each day. al. (2014). In the paper, the feed forward neural
The total price of the transformer load collected was network was modeled using the relationship between
calculated at the rate of 1Kw unit of energy per hour the weights, bias and activation function of a single
at 51.22 naira (class B (ie the class the feeder neuron to develop feedforward neural network
belongs)) in customer traffic plan. The results of the architecture. The model of the simple feedforward
data collection are all reported in chapter four. The neural network is presented in figure 1 as;
data analysis for the overall load consumed and the
price are all reported in the chapter four.

weight and bis functions incite the neurons as the

Figure 1: Model of a simple feedforward neural input. The summation of the input is activated for
network training to generate the desired load forecasting
The figure 1 presented a simple feedforward neural system. The neural network model in figure 1 was
network architecture configured with the weight, used to develop the neural network architecture used
bias, activation function, summation function and the in training the feeder data. The developed model was
activated output. The model shows how the internal presented in figure2;
structure of a neural network was modeled with the
 4  +234 703 322 5267

Figure 2: Model of the neural network

The figure 2 presented the model of the neural
network developed to train the data collection. The
model was developed using considering the number
of classes in the data collection which are load and
price. These were used to configure the input number
to the network. The hidden layers were configured
using the training parameter in table 1;
Training Parameters Values
Maximum number of epochs to train 100
Epoch between display 10
Maximum time to train in sec Infinity
Validation check 6
Initial step size 0.01
Minimum performance gradient 1e-6
Cost horizon 7
Control horizon 2
Number of bias functions 20
Training Iterations 56
Number of hidden layers 20
Number of weights 20
Number of input class 2

Algorithm
1. Start
2. Load training set
3. Configure neural network
4. Initialize epoch
5. Initialize Mean Square Error (MSE)
6. Adjust the neural network weights
7. Check for least MSE
8. If
9. Error is true
10. Then
Figure 3: The system flow chart
11. Stop training
12. Else
4 RESULT AND PERFORMANCE
13. Next epoch
ANALYSIS
14. Feedback output to input
This section presented the performance of the load
15. Adjust weights
forecasting algorithm developed. The result discussed
16. Check MSE
the training performance of the data collection with
17. Is least MSE = true
the neural network algorithm to develop the load
18. Stop training
forecasting system using MSE and Regression. The
19. Reference mode result of the algorithm was validated and analyzed.
20. Stop The MSE analysis was presented in figure 4;

Vol. 2, Issue 4; No. 43, April, 2023, pp. 453-460

International Journal of Artificial Intelligence Trends (IJAIT) 5
Vol. 2, Issue 4; No. 43, April, 2023, pp. 453-460

10 at epoch 415. The implication of this result

Figure 4: MSE result of the algorithm showed that the error which occurred during the
The result of the MSE in figure 4 was used to training process is minimal and hence indicated that
measure the error which occurred in the neural the learning process was good and the algorithm was
network during the training process. The idea was to able to predict the transformer behavior on a day time
achieve error of zero or approximately zero. From the series. The next result presented the regression
result it was observed that the MSE result is 1.1597E- performance as in figure 5;

5 1.2297E-10 0.9879
Figure 5: Regression Performance 6 1.1387E-10 0.9309
TABLE 2: VALIDATION RESULT OF THE 7 1.1597E-10 1.0000
ALGORITHM 8 1.1997E-10 0.9879
S/N MSE Regression 9 1.1447E-10 0.9934
1 1.1597E-10 0.9999 10 1.1397E-10 0.9952
2 1.2597E-10 0.9679 Average 1.1537E-10 0.98469
3 1.1897E-10 0.9939 The table2 presented the result of the validation
4 1.1297E-10 0.9899 performance and it was observed that the MSE value
 6  +234 703 322 5267

is 1.1537-E which is approximately zero thus

indicating good training performance. The result of 13 15.28
the regression was also validated and the result of 14 13.58
0.98469 was achieved, which is also very close to the 15 16.13
ideal regression value of 1. The result of the system 16 13.78
integration presented the impact of the algorithm on 17 12.46
the EEDC system for the load forecasting of the 18 13.57
feeder performance in the next 24hrs. The result is 19 14.38
presented in the table 3; 20 16.97
TALE 3: FORECASTED DATA OF THE 21 14.29
FEEDER IN THE NEXT 24HRS.
22 15.78
Time (Hr) Average daily load (Kw/h)
23 15.78
1 14.98
24 13.51
2 10.56
Total 340.52
3 16.29
Cost kW/Hr 17,441.4344
4 16.11
Cost Kw/day 418,594.426
5 14.39
The result of the table 3 presented the performance of
6 14.39 the algorithm when used to forecast the load
7 14.32 consumption of the transformer in the next 24hrs.
8 13.14 The estimated amount generated per hour and total
9 13.14 amount expected from the feeder in the next 24Hr
10 11.22 was presented which is 418,594.426. The result
11 13.62 showed that the transformer will consume 340.52kW
12 15.29 of power.

Figure 6: Estimated behavior of the feeder with ANN 4.1 COMPARATIVE ANALYSIS
The figure 6 presented the load forecasted behavior In this section, the data acquired after the
of the transformer showing estimated load implementation of the algorithm was compared with
consumption performance of the system in the next the actual data collected on the date and the to
24Hr. The result showed that the estimated peak determine the system reliability and the results are
period is 8pm (20:00) while the off-peak period is presented in table 4;
2:00 am.

Vol. 2, Issue 4; No. 43, April, 2023, pp. 453-460

International Journal of Artificial Intelligence Trends (IJAIT) 7
Vol. 2, Issue 4; No. 43, April, 2023, pp. 453-460

TABLE 4: COMPARATIVE RESULT OF THE TRANSFORMER DAILY DATA

Time (Hr) Estimated Load Actual Load 21 14.29 14.44
1 14.68 15.00 22 15.78 15.90
2 10.46 10.67 23 15.78 15.90
3 16.19 16.40 24 13.51 14.00
4 16.11 16.30 Total 340.52 344.46
5 14.29 14.50 Cost
6 14.39 14.50 kW/Hr 17,441.4344 17,643.2412
7 14.12 14.44 Cost
8 13.14 13.24 Kw/day 418,594.426 423,437.789
9 13.14 13.24
10 11.22 11.34 The result in the table 4 presented the performance of
11 13.52 13.74 the comparative performance of the transformer using
12 15.29 15.40 the actual and estimated data. The result showed that
13 15.28 15.40 the estimated total load consumed by the transformer
14 13.58 14.07 is 340.53kW as against 344.46kW. The implication
15 16.13 16.24 of the result showed that the load forecasting
16 13.78 13.90 algorithm was able to closely estimate the
transformer load. The graphical analysis was used to
17 12.46 13.00
read other behavior of the transformer as shown in
18 13.57 13.70
figure 7;
19 14.28 14.50
20 16.67 16.83

was able to estimate that the actual peak and off-peak

Figure 7: Comparative transformer behavior period and other load behavior. The result also
The figure 7 presented the comparative result of the estimated the amount of load consumed by the
load forecasting algorithm as against the actual data. transformer and the actual load consumed as
From the result, it was observed that the algorithm presented in the figure 8;
 8  +234 703 322 5267

Figure 8: Comparative cost analysis

The figure 8 presented the comparative feeder Nigerian Journal of Technology (NIJOTECH)
performance in term of cost. The result estimated the Vol. 39, No. 3, July 2020, pp. 860 – 870; Print
amount of load consumed by the transformer on ISSN: 0331-8443, Electronic ISSN: 2467-8821
hourly basis and also on daily basis. The result
Haider HT, See O., Elmenreich W. (2016). A review
showed that the load flow analysis system was able to
correctly estimate the amount of power consumed of residential demand response of smart grid.
from the load. The idea is to know which is expected Renew Sustain Energy Rev 2016;59:166–78.
from the feeder at the end of the day and then https://doi.org/10.1016/J.RSER.2016.01.016
compare with the actual cost realized from the feeder. Hana. T., (2018). Short term electric load forecasting.
This will help in better demand site management. Ph.D. dissertation, North Carolina State
4. CONCLUSION University, 2018.
Maew. Macedo, J. Galo, L. de Almeida, and A. de C.
From the experimental investigation conducted on the
load flow of the 11kV feeder, it was uncovered that Lima, (2018). Demand site management using
the behavior of the transformer varies on daily basis arti_cial neural networks in a smart grid
and as a result there is need for a smart strategy for environment. in Renew. Sustain. Energy Rev. 41.
demand site management. This solution was Nauru Cetinkaya, (2016). A new mathematical
developed in this research using artificial neural approach and heuristic methods for load
network to develop an algorithm for load forecasting. forecasting in smart grid. 2016 12th International
The algorithm developed was tested, validated and
Conference on Natural Computation, Fuzzy
deployed at the EEDC for the management of the
Trans-Ekulu feeder. The result showed that the Systems and Knowledge Discovery (ICNC-
system developed was able to correctly estimate the FSKD), Aug 2016, pp. 1103-1107.
performance of the load for seven consecutive days O’Connell N, Pinson P, Madsen H, O’Malley M.
and also determine the total amount expected to be (2016). Benefits and challenges of electrical
generated from the feeder. demand response: a critical review. Renew
5. REFERENCES Sustain Energy Rev 2016;39: 686–99.
https://doi.org/10.1016/J.RSER.2014.07.098.
Arjun B., Kumar G., Sudhansu K. M., (2014). A
Zah. H. T, X. F. Y, and Z. L., (2015). Artificial
Review of Short-Term Load Forecasting using
neural network for load forecasting in smart grid,
Artificial Neural Network Models. International
2015, pp. 128-133.
Conference on Intelligent Computing,
Zina, B., C., Ahmed R. and Haritza C. (2018). A
Communication & Convergence (ICCC-2014),
Nonlinear Autoregressive Exogenous (NARX)
Ashigwuike E.C, A. R. A. Aluya, J. E. C. Emechebe
Neural Network Model for the Prediction of the
and S. A. Benson (2020). Medium Term Electrical
Daily Direct Solar Radiation. Energies Vol 620.
Load Forecast Of Abuja Municipal Area Council
Using Artificial Neural Network Method “

Vol. 2, Issue 4; No. 43, April, 2023, pp. 453-460