You are on page 1of 6

Available online at www.sciencedirect.

Available online at
Energy Procedia
Energy 00 (2011)14
Procedia 000–000
(2012) 794 – 799

2011 2nd International Conference on Advances in Energy Engineering (ICAEE 2011)

December 27-28, 2011, Bangkok, Thailand

Long-term load forecasting based on adaptive neural fuzzy

inference system using real energy data
Bayram Akdemira, Nurettin Çetinkayab, a*
Selcuk University, Department of Electrical and Electronics Engineering, Konya, 42075, Turkey


Energy production and distributing have critical importance for all countries especially developing countries. Studies
about energy consumption, distributing and planning have much importance at the present day. In order to manage
any power plant or take precautions about energy subject, many kinds of observations are used for short, mid and
long term forecasting. Especially long term forecasting is in need to plan and carry on future energy demand and
investment such as size of energy plant and location. Long term forecasting often includes power consumption data
for past years, national incoming per year, rates of civilization, increasing population rates and moreover economical
parameters. Long term forecasting data vary from one month to several years. Some of the forecasting models use
mathematical formulas and statistical models such as correlation and regression models. In this study, artificial
intelligence is used to forecast long term energy demand. Artificial intelligences are widely used for engineering
problems to solve and obtain valid solutions. Adaptive neural fuzzy inference system is one of the most famous
artificial intelligence methods and has been widely used in literature. In addition to numerical inputs, Adaptive neural
fuzzy inference system has linguistics inputs such as good, bad and ugly. Adaptive neural fuzzy inference system is
used to obtain long term forecasting results and the results are compared to mathematical methods to show validity
and error levels. In order to show error levels, mean absolute error and mean absolute error percentage are used.
Mean absolute error and mean absolute error percentages are very common and practical methods in literature. The
obtained error results, from 2003 to 2025, mean absolute error and mean absolute percentage error are 1.504313 and
0.82439, respectively. Success of Adaptive neural fuzzy inference system for energy demand forecasting is 99.17%.
© 2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of the organizing committee
© 2nd
of 2011International
Published by Elsevier Ltd.
Conference Selectioninand/or
on Advances Energypeer-review
Engineeringunder responsibility
(ICAEE). of ICAEE 2011
Open access under CC BY-NC-ND license.

Keywords: Adaptive neural fuzzy inference system; long term forecasting; mean absolute error; mean absolute error percentage; real
data set.

* Corresponding author. Tel.: +90-332-223-2042; fax: +90-332-241-0635.

E-mail address:

1876-6102 © 2011 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of the organizing committee of 2nd International
Conference on Advances in Energy Engineering (ICAEE). Open access under CC BY-NC-ND license.
2 Bayram Akdemir
Bayram andand
Akdemir Nurettin \’c7etinkaya\
Nurettin Çetinkaya // Energy
Energy Procedia
Procedia 14
00 (2012)
(2011) 794 – 799
000–000 795

1. Introduction

Energy transitions are defined as; investments in possibly cleaner technologies to replace and expand
the depreciating capital stock to meet growing energy demand. When considered a longer time horizon,
significant changes in energy technologies and consumption could be observed [1]. Electricity energy is a
vital input for technical, social and economic development of a country. Therefore, the identification and
analysis of energy issues and development of energy policy options are of prime importance [2]. Power
load forecasting can be divided into load forecasting and electrical consumption forecasting according to
forecasting matter. Based on different forecast time, it can be divided into long-term, mid-term and short-
term forecasting. Short-term, mid-term and long-term load forecasts range from an hour to one week, one
month to one year and one year to decades, respectively.
From short-term load forecasting used in the daily unit allocations, to mid-term load forecasting used
for fuel budgeting, and to the long-term load forecasting used for resource planning and utility expansion,
a wide variety of techniques have been proposed in the last 30 years [3].
The studies on electrical load forecasting of Turkey began at 1960s. In Turkey, energy consumption
projections are made by Ministry of Energy and Natural Resources of Turkey (MENR). Since 1984,
MENR prepares energy demand forecasts by using Model for Analysis of Energy Demand (MAED)
simulation technique. MAED requires several types of data related to social, economical and
demographical structure of country. However, the governments may tend to use target values of data that
mentioned former statement instead of estimated values. For this reason, it is stated in a number of studies
that forecast values founded by MAED technique are higher than real values. In order to achieve lower
forecast error, utilizing different techniques based on direct observation values, instead of scenario-based
techniques like MAED, will be useful [4].
Artificial intelligence is widely used for nonlinear problems in many engineering such as computing,
structure, medical electronics [5-7]. Artificial intelligence often generates simple solution for difficult
problems based on experience without using equations, formulas and derivatives.
The use of neural networks is nowadays an alternative method that is becoming an efficient technique
to solve the electric load forecasting problem. There are researches that treat this problem using the back-
propagation algorithm. This algorithm is considered on the specialized literature a benchmark in precision.
However, the convergence is slow, although there are some adaptations to improve the performance.
Therefore, the idea is to use a neural network that combines good results with a faster processing [8].
Neuro-fuzzy modeling refers to the way of applying various learning techniques developed in the
neural network literature to fuzzy modeling or a fuzzy inference system (FIS). Neuro-fuzzy system, which
combines neural networks and fuzzy logic have recently gained a lot of interest in research and
application. A specific approach in neuro-fuzzy development is the adaptive neuro-fuzzy inference system
(ANFIS), which has shown significant results in modeling nonlinear functions [9].
In this study, ANFIS is used to solve long term forecasting problem for energy demand from 2003 to

2. Material and Methods

Long term forecasting is one of the famous subjects in energy demands problems. Many kinds of
observations and background data can be used for long term data structure. Carrying on energy demands
in long term, energy demands and power plant location size of plant are very important and must be
decided taking care of many variables. In this study, ANFIS is used to forecast long term energy demands
up to 2025. Used data includes population, reached maximum power demand in an hour and total energy
level given in giga-watt hours (GWh). The estimations for energy demands for future years covering all
country were implemented using the projections of official distributing energy service called Turkey
Electric Transmission Company (TEIAS) cover all country which is naming [10].
796 Bayram
Bayram Akdemir
Akdemir andand Nurettin
Nurettin \’c7etinkaya\
Çetinkaya / Energy
/ Energy Procedia
Procedia 14 (2012)
00 (2011) 794 – 799
000–000 3

2.1. Data Structure and Data Arranging

Raw data set has four features and 27 attributes. Features of data set are population, peak energy level
in an hour per year and total observed or guessed official energy demand per year as GWh. Raw data
were re-generated on time axis as six years attributes. Windowed data structure includes past four years,
present year and future year. After arranging data set new structure turned into 17x23 attributes. Table 1
shows raw data structure. From the table, past year energy demand was written as observed data and
future year energy demand written and used as official government guessed energy demand through the

Table 1. Raw data structure and data features

Years Peak load Population Incoming Energy demand

1999 18.938 63.366 1.071 118.485
2000 19.39 64.259 1.127 128.276
… ... ... ... ...
2009 30.982 71.897 1.498 200.137
2010 31.373 72.698 1.501 202.368
… ... ... ... ...
2025 68.547 83.566 2.905 335.225

After arranging the data, although data structure gets larger, estimation capability can be improved by
converting data for ANFIS [11]. Converted data structure includes six years packages as five years
backwards data and next year energy demand as target. Target is next year of the energy demand as GWh.

2.2. Adaptive Neural Fuzzy Inference System

ANFIS is a very common and widely used artificial intelligence technique in literature. This method
covers advantages of Fuzzy Logic and Artificial Neural Networks in same structure. Although Neural
Network (NN) has powerful learning ability, Fuzzy structure has strong inference system and has no
learning ability. In contrary, ANFIS merges these two desired features in the same topology. ANFIS was
presented by R.Yang in 1993 [12]. ANFIS is widely used in engineering applications for solving many
kinds of nonlinear problems. [13-15]. For Sugeno type, ANFIS structure be explained in five stages.
Stage 1: This layer can be called as fuzzification layer. Used parameters in this stage is called premise
parameters and re-arranged according to output error in every loop. These parameters are membership
grades of a fuzzy set and input parameters in this layer.
Stage 2: A fixed node labeled ∏ whose output is the product of all the incoming signals can be
computed. Every output of the stage 2 affects the triggering level of the rule in the next stage. Trigger
level is called firing strength and ∏ norm operator is called AND operator in fuzzy system.
Stage 3: This layer is called as normalization layer. For this layer, all firing strengths are re-arranged
again by considering own weights.
Stage 4: Defuzzication, this layer is a preliminary calculation of the output for real world. This layer
has adaptive nodes and it is expressed as functions and if ANFIS model is Sugeno type then is valid
calculation styles turn to linear approach. This type is called first order Sugeno type [16].
Stage 5: Summation neuron; this layer is a fixed node, which computes the overall output as the
summation of all incoming signals. ANFIS Learning Ability; ANFIS has two times error correction
ability in one loop. This correction is processed backwards and forward. For backward correction, the
4 Bayram Akdemir
Bayram andand
Akdemir Nurettin \’c7etinkaya\
Nurettin Çetinkaya // Energy
Energy Procedia
Procedia 14
00 (2012)
(2011) 794 – 799
000–000 797

antecedent parameters are tuned while the consequent parameters are kept fixed. Least square estimator
arranges the parameters to minimize the squared error.
In this study, data set is not normalized to obtain real response from ANFIS structure. Natural
condition of data set has been kept and low ANFIS rule structure was employed to obtain fast response.
ANFIS setting are configured as range of influence 0.9, squash factor 5, reject ratio 0.25, and accept ratio
is 0.6. Range of influence and squash factor are increased to obtain low level ANFIS rule structure and to
decrease training cycles.

2.3. Performance Measuring and Testing

Performance measuring is a way to compare and present for any implementation. Every measurement
method has its own characteristic and implementation method. Mean absolute error and mean absolute
error percentage are among the most famous methods to measure success of any implementation [17-19].
• Mean absolute Error: Mean absolute error is one of the simplest ways to evaluate any success and
depends on mean of difference among observations and real values [20-21]. Mean absolute error is
calculated by Equation 1.

1 n xim − xip
MAE = ∑ x
n i =1

where xim is the ith measured value and xip is the forecasted value.
• Mean absolute percentage error: Some conditions require small numbers and small quantities. In this
case mean absolute error and its derivatives may lead to misunderstanding or may not explain correctly.
Mean absolute percentage error is used to support mean absolute error results. Mean absolute error
percentage, given in Equation 2, has no unit and very common in energy forecasting applications [22-

1 n xim − xip
MAPE = 100 × ∑ x
n i =1

where xim is the ith measured value and xip is the forecasted value.

3. Proposed Method Implementation and Results

Long term forecasting data structure and ANFIS are used to forecast energy demand and tested up to
2025. Data set is windowed in six parts. Created new data set via windowing is performed by ANFIS as
training and testing. The obtained mean absolute error and mean absolute error percentage for the testing
values for forecasting and forecasting results are graphed in 3-D in Figure 1a and Figure 1b. From Figure
1a, forecasted and real energy demand values are close to each other and have same slope through the
years. Error levels for mean absolute error and mean absolute error percentage are shown in Figure 1b.
Especially for the future years, error levels are decreasing for MAE and MAPE too.
798 Bayram
Bayram Akdemir
Akdemir andand Nurettin
Nurettin \’c7etinkaya\
Çetinkaya / Energy
/ Energy Procedia
Procedia 14 (2012)
00 (2011) 794 – 799
000–000 5

(a) (b)

Fig. 1. (a) Forecasted and real energy demands through the years; (b) Test and train errors as MAPE

Characteristics of errors are listed in Table 2. Maximum error is obtained as 6.8174 GWh for year of
2009 and minimum error is obtained as 0.00000347 GWh for year of 2007. For future forecasting, ANFIS
has very accurate forecasting results. For future, ANFIS has just negligible error 0.81163829 GWh as
MAE and it appears 0.373457789% as MAPE from 2010 to 2025.

Table 2. Obtained results from ANFIS as MAE and MAPE

For Past Years For Future Years

Max 6.8174 (2009) 3.752 (2003) 1.4492 (2011) 3.0086 (2011)
Min 0.4652 (2007) 0.2442 (2007) 3.47E-06 (2007) 1.03E-05 (2007)
Average 3.0875 1.8550 0.81163829 0.373457789
Total Average MAPE: 0.82439 MAE: 1.504313

4. Conclusion

Long term forecasting is one of the most famous energy problems in literature and has no linear
correlation among the input variables. In this study, ANFIS is used to evaluate real energy demand
through the years as long term forecasting. Proposed method was implemented based on ANFIS and
backwards data were created and extended. Data structure was windowed as six parts from last four years,
present year and future year. After data turned into new time series structure, ANFIS is used for
evaluation. ANFIS has low error level especially for the future years forecasting. Average forecasting
error is 1.504313 GWh, although energy demands vary from 118.485 to 335.225 GWh. Considering the
results obtained in this study, we can say that it is possible to use the proposed forecasting method for
long term energy demand.


[1] Ünler A. Improvement of energy demand forecasts using swarm intelligence: The case of Turkey with projections to 2025.
Energy Policy 2008; 36: 1937-1944.
6 Bayram Akdemir
Bayram and
Akdemir Nurettin
and \’c7etinkaya\
Nurettin Çetinkaya // Energy
Energy Procedia
Procedia 00
14 (2011)
(2012) 000–000
794 – 799 799

[2] Toksari MD. Estimating the net electricity energy generation and demand using the ant colony optimization approach: Case
of Turkey. Energy Policy 2009; 37: 1181-1187.
[3] Kermanshahi B, Iwamiya H. Up to year 2020 load forecasting using neural nets. Electrical Power and Energy systems 2002;
24: 798-797.
[4] Hamzaçebi C. Forecasting of Turkey’s net electricity energy consumption on sectoral bases. Energy Policy 2007; 35: 2009-
[5] Özkan AO, Kara S, Sall A, Sakarya ME, Güneş S. Medical diagnosis of rheumatoid arthritis disease from right and left hand
Ulnar artery Doppler signals using adaptive network based fuzzy inference system (ANFIS) and MUSIC method. Advances in
Engineering Software(AES) 2010; 37, 8, 2010: 5720-5727.
[6] Yao X, Evolving Artificial Neural Networks. In Proc. of the IEEE,87(9), 1999: 1423-1447.
[7] Polat K, Güneş S, A novel hybrid intelligent method based on C4.5 decision tree classifier and one-against-all approach for
multi-class classification problems, Expert Systems With Applications, 36(2-1), 2009: 1587-1592.
[8] Lu´cia ML, Minussi CR, Diva AP. Electric load forecasting using a fuzzy ART&ARTMAP neural network. Applied Soft
Computing 2005; 5: 235-244.
[9] Nadimi V, Azadeh A, Pazhoheshfar P, Saberi M. An Adaptive-Network-Based Fuzzy Inference System for Long-Term
Electric Consumption Forecasting (2008-2015): A Case Study of the Group of Seven (G7) Industrialized Nations: U.S.A, Canada,
Germany, United Kingdom, Japan, France and Italy. UKSim Fourth European Modelling Symposium on Computer Modelling and
Simulation IEEE 2010; 301-305.
[11] Akdemir B, Yu L. Elliot Waves Predicting for Stock Marketing Using Euclidean Based Normalization Method Merged
with Artificial Neural Network. 2009 Fourth International Conference on Computer Sciences and Convergence Information
Technology, Seoul, Korea, November 2009: 24-26.
[12] Jang J-SR. ANFIS: Adaptive-network-based fuzzy inference system. IEEE Transactions on System, Man, and Cybernetics,
No. 5, Vol. 23, May 1993; 665-685.
[13] Roger JS, Sun C T. Neuro-Fuzzy Modeling and Control, The Proceedings of the IEEE, No. 3, 83, March 1995: 378-406.
[14] Yanmei L, Zhen C, Real-Time Controlling of Inverted Pendulum by Fuzzy Logic, International Conference on Automation
and Logistics, doi:978-1-4244-4795-4/ 09/2009: IEEE, Aug.2009: 1180-1183.
[15] Akdemir B, Oran B, Güneş S, Karaaslan S. Prediction of cardiac end-systolic and end-diastolic diameters in m-mode values
using adaptive neural fuzzy inference system, Expert Systems with Applications, 2010: 37(8), 5720-5727.
[16] Narukawa Y, Murofushi T, Sugeno M. Regular fuzzy measure and representation of comonotonically additive functional.
Fuzzy Sets and Systems, 112(2), June 2000; 177-186.
[17] Başar ME, Akdemir B. Adaptive neural fuzzy inference system (ANFIS) to determine possible relation between bricks
porosity and weight losing, Scientific Research and Essays Vol. 6(12), 18 June, 2011: 2480–2486.
[18] Mokhtarpour H, Teh CBS, Saleh G, Selamat AB, Asadi ME, Kamkar B. Non-destructive estimation of maize leaf area,
fresh weight, and dry weight using leaf length and leaf width, International Journal of the Faculty of Agriculture and Biology,
Warsaw University of Life Sciences, Poland, Communications in Biometry and Crop Science, 5, 1, 2010: 19–26.
[19] Loce RP, Dougherty ER. Mean-Absolute-Error representation and optimization of computational-morphological filters,
graphical models and image processing, doi:10.1006/gmip.1995.1004, 57, 1, January 1995: 27-37.
[20] Armstrong J S, Evaluating forecasting methods, Chapter 14 in Principles of forecasting: a handbook for researchers and
practitioners, ed., Kluwer Academic, 2001, Publishers: Norwell, MA.
[21] Hyndman RJ, Koehler AB. Another look at measures of forecast accuracy, monash econometrics and business statistics
working papers 13/05, Monash University, Department of Econometrics and Business Statistics Published at: International Journal
of Forecasting, Elsevier, 22, 4, 2005: 679-688.
[22] Akdemir B, Çetinkaya N. Importance of Holidays for Short Term Load Forecasting Using Adaptive Neural Fuzzy Inference
System, International Conference on Power and Energy Engineering, ICPEE-2011.
[23] Tayman J, Swanson DA. On the validity of MAPE as a measure of population forecast accuracy, Population Research and
Policy Review, 18, 4, 1999: 299-322.