Short term load forecasting system based on support vector kernel methods

International Journal of Computer Science & Information Technology (IJCSIT) Vol 6, No 3, June 2014
DOI:10.5121/ijcsit.2014.6308 93
SHORT TERM LOAD FORECASTING SYSTEM BASED
ON SUPPORT VECTOR KERNEL METHODS
Lal Hussain1
, M. Sajjad Nadeem2
, Syed Ahsen Ali Shah3
1,2,3
University of Azad Jammu and Kashmir, City Campus – 13100, Muzaffarabad,
Pakistan
ABSTRACT
Load Forecasting is powerful tool to make important decisions such as to purchase and generate the
electric power, load switching, development plans and energy supply according to the demand. The
important factors for forecasting involve short, medium and long term forecasting. Factors in short term
forecasting comprises of whether data, customer classes, working, non-working days and special event
data, while long term forecasting involves historical data, population growth, economic development and
different categories of customers.In this paper we have analyzed the load forecasting data collected from
one grid that contain the load demands for day and night, special events, working and non-working days
and different hours in day. We have analyzed the results using Machine Learning techniques, 10 fold cross
validation and stratified CV. The Machines Learning techniques used are LDA, QDA, SVM Polynomial,
Gaussian, HRBF, MQ kernels as well as LDA and QDA. The errors methods employed against the
techniques are RSE, MSE, RE and MAPE as presented in the table 2 below. The result calculated using the
SVM kernel shows that SVM MQ gives the highest performance of 99.53 %.
KEYWORDS
Forecasting, Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), QuadraticDiscriminant
Analysis (QDA), Mean Square Error (MSE), Relative Error (RE) and Mean Absolute Percentage Error
(MAPE), Cross Validation (CV)
1. INTRODUCTION
Accurate models predict electricity load power necessary for the operation and planning for the
utility company. Load forecasting electric tool helps to make important decisions, including
decisions relating to the purchase and electric power generation, load switching, and
infrastructure development. Expectations of pregnancy are extremely important for energy
suppliers, solidarity and financial situations, and other participants in the electric power
generation, transmission and distribution, and markets. And expectations of pregnancy can be
divided into three categories: short-term outlook, which is usually from one hour to one week,
and forecasts medium which is usually from one week to one year, and long-term forecasts that
are longer than a year. Expectations for different time horizons important for different processes
within the company benefit.

94
2. IMPORTANT FACTORS FOR FORECASTS
For short-term load forecasting should consider several factors, such as time factors, weather
data, and potential customer classes. Medium-term outlook and long-term take into account the
load historical data and weather, and a number of clients in different categories, Devices in the
region and their characteristics, including age, economic and demographic data and expectations,
and appliance sales data, and other factors. Time factors include time of year, and day of the week
and hours of the day. There are significant differences in pregnancy between weekdays and
weekends. The load on different days of the week also can behave differently. For example, it
may be adjacent Mondays and Fridays for the weekend, they have structurally different amounts
from Tuesday to Thursday. Is this is particularly true during the summer. The holidays are more
difficult to predict because of the holidays and a relatively rare occurrence. Weather conditions
affect the pregnancy. In fact, the Weather forecast parameters are the most important factors in
pregnancy expectations in the short term. Can be considered as different weather variables to
predict pregnancy. Temperature and humidity are the most commonly used predictors of
pregnancy.
3. LITERATURE REVIEW
To accurately forecasting the electric load we can save the electricity as well as can distribute the
efficient and limited energy sources. When the load forecasting of the electricity is not accurate it
may increase the cost for operating the forecasting. For instance, when the electric load is over
estimated it results unnecessary reservation of the spinning, inefficient distribution of the energy,
the limited energy sources are wasted moreover, the excess supply owning by the international
network is not acceptable. In comparison, when the electric load forecasting is under estimating it
results in failure to provide sufficient reservation and produce high costs for the peaking units due
to which the industrial and economic developments are discouraged. In order to achieve the
future electric load demands, it require efficient electricity load planning, secure operation field in
both the regional and national systems [5].
The demand for the electricity is usually modeled in terms of the weather variables(Hor et al.,
2005; Cancelo et al., 2008). However, for the short term forecasting, the univariate methods are
considered as sufficient as the weather variable smoothly changes in short time frames. Likewise,
to ensure the robustness, the weather-based online systems require default procedures(Bunn,
1982). According to the recent empirical study(Taylor, 2008a), a univariate method can better
perform than the multivariate method up to four hours ahead, however, by combining the
forecasting of these two methods can achieve the best performance up to a day ahead. This
implies that univariate can play a vital role for short term load forecasting (Soares and Medeiros,
2008) [7].
In order to best utilize the electric energy demands and resolve the conflict between supply and
demand, load forecasting is extremely dire to plan the operational electric power systems and
relevant research purposes as well.Short term load forecasting (STLF) is an essential part of load
forecasting for its controllable, random, real-time,non-linear and dynamic factors influenced on it.
As per characteristics of STLF, the neural networks are gaining an increased attention for
dynamic and non-linear behavior of the system having strong robustness, fault tolerance, wavelet

95
transform, grey system and support vector machine to achieve and face the demands and needs
for higher accurate STLF [3].
Presently, artificial intelligence techniques have been implied in order to increase the electricity
load forecasting demands and performance. Moreover, Rahman and Bhatnagar have presented
knowledge based expert system (KBES) approach to achieve better load forecasting for the
electricity, which can construct new rules based on the gathered information like daily
temperature, previous day load, day type etc. Using this approach attained through training rules,
the information received then transformed to the mathematical equations.Park et al. has
established a three layer back propagation neural network to solve the problems of daily
forecasting problems. The features extracted from the daily load are computed from temperature
indices: peak, average and lowest load, while the output is the peak load.This model performs
well for time series model and regression computed over the mean absolute percent error
(MAPE). Initially the SVMs have been applied for pattern recognition problems; however, with
the development of intensive loss function by Vapnik, SVMs are increasingly used in non-linear
regression problems as well as times series forecasting (Cae et. al.) [1].
SVM configuration affects the SVM performance such as type of kernel function, kernel
parameter and size of training sample. Thus the previous studies using SVM on wind forecasting
are limited in which particular kernel function and particular parameter are picked for these
studies. Kernel function to be chosen is very important for any particular application such as
linear, polynomial and Gaussian kernel in order to investigate the accuracy of the load forecasting
[8].
Support vector machine (SVM) algorithms have the capability to control the capacity of decision
function and choose the kernel function accordingly according to the sparsity of the solution.
SMS is basically structured on risk minimization factor that eventually yields high generalization
performance. Besides, SVM have the ability to solve both linear and quadratic programming
problems having always unique solutions and globally optimal unlike neural network which
require nonlinear optimization that may get stuck at local minima [6].
Short-term load forecast (STLF) always remains a demanding issue for power operation systems.
Various operating decisions are comprised on STLF including dispatch scheduling, reliability
analysis, security maintenance and assessment plans for generators. In addition, load forecasting
play a vital role in energy transaction in the competitive electricity market. The market profit,
share and share holder value are greatly affected by the forecast error. The electricity load
forecasting to the variability and nonstationarity of load series became difficult to forecast,
especially in electricity market [4].
Load forecasting approaches based on the demands have been generally classified into the time
series, kalman filtering technology, state space, artificial intelligence, regression models, fuzzy
logic methods. Time series approach such as Box–Jenkins ARIMA model, predict the electricity
load forecasting based on historical load data. On the other hand, Kalman filtering approach use
random process for load forecasting using 3-10 historical data in order to establish the periodic
load component of the depending variable such as load or temperature of power system. The
regression model establishes cause and effect relationship between the independent variable such
as climate factor, seasonal factor and social activities with the electricity load [2].

96
SVM developed by Vapnik 1998 is used in statistical machine learning and regression
problems(Akay, 2009; Cherkassky& Ma, 2004; He, Wang, & Jiang, 2008; Lazaro, Santamarı´a,
Perez-Cruz, &Artes-Rodrı´guez, 2005; Wu, Chau, & Li, 2008) as well as wind speed
prediction(Mohandes, Halawani, Rehman, & Hussain, 2004; Salcedo-Sanz et al., 2009).
Moreover, SVMr training are used for sequential minimal optimization
algorithms(Smola&Scholkopf, 1998) [9].
In the last decade due to the rapid increase in the wind generation of electricity, the capacity of
the wind power have been increased almost fourfold 24.3 GW to 203.5 GW. Short-term load
forecasting are used in the power system planning for unit commitment and dispatch (1 hr to 72
hr), while Medium-term forecasting ranging from 3 days to 7 days used to plan maintenance of
the wind farms, outage of thermal generation, unit commitment, storage operation and to schedule
grid maintenance of the energy. The wind power predictions have been described both in theories
and practices of meteorology and climatology specifically to wind power generation [10].
4. DATA ACQUISITION
Al-Khwarizmi Institute of Technology, The Punjab University Lahore, provide electricity hourly
load demands of the complete year 2008 and July to December 2009 of one grid. The data was
taken from a feeder in order to fulfill the maximum load requirementsfor this study. The data was
accordingly separated for morning, day, evening and night hours. Likewise, load data season wise
was separated. However, there was no special event data in the present study.
5. RESULT AND DISCUSSIONS
Electricity consumption on the basis of current and past, and is used to predict the electrical load
to estimate the load electric power at a later time. For one day predict pregnancy and a one-week
prediction pregnancy is to arrange a plan today schedule and plan the schedule for a week on the
unit, which includes the decision unit GE to get out, and coordination between hydropower and
electricity, heat and energy exchange in the connection, distribution of economic power load, and
Inspection Equipment, etc. ultra-short-term prediction of pregnancy within 1 hour and is used to
guide the distribution of economic output set startup and shutdown a range of meals, they are also
used to monitor the safety, preventive and control critical to deal with the situation. At present,
power plants, especially those in China, are in dire need of a very short pregnancy prediction, any
expectation within one minute or one hour. Load forecasting instruction judge dispatchers
download the very short term, both parts of the revolution enough at the right time in the near
future, rotary parts is an important factor in determining the integrity of the system category. In
addition, it helps dispatchers to adjust the distribution of power group, on and off from a move in
the peak and trough packing, and it is also a must for automatic generation control (AGC) and the
achievement of economic dispatch dynamically. Long-term, very short power load forecasting is
the information that the energy market demands.
Smart Grid will in the future to achieve reliable operation of the power system and the economic
benefit from its resources. This requires predict pregnancy in a wide range of yarns time, from
minutes to several days. Carried predict loads integrated every hour outside for a day 1-1A week
ago and is usually referred to predict short-term pregnancy (STLF). These projections are used
each hour through energy management systems (EMS) for the development of operational plans

97
for power plants and their generation units, and energy planning transactions in the market. It is
also necessary for theSecurity Studies energy system, including the analysis and management of
emergency loads.
In addition, the load-frequency control and dispatch functions for EMS economic require
pregnancy forecasts in within a shorter lead time, from one minute to several tens of minutes.
And they are referred to and expectations of pregnancy for a very short time (VSTLF). These
projections, integrated with information about transactions Wheeling decision, and the
availability of transmission, and the generation cost, the spot market pricing of energy, spinning
reserve requirements imposed by the independent system operator, and is used to determine the
best strategy for the benefit of resources. Very short-term Load forecasting has become much
greater importance in today's liberalized energy industry.
Figure 1: Load Demand Winter Season January 01, 2008
In the above figure 1, load demand is plotted for January 01, 2008, morning hours, before noon,
after noon, evening and night hours. The load data collected from s grid station shows that
electricity demands are different during different hours of the day. Thus it is pertinent to supply
the load demands accordingly during different hours of the day. This comes in Very Short term
Load Forecasting (VSTLF) and Short Term Load Forecasting (STLF) that is load demands for
every next hour and next few minutes of the day. This will help us to know the minimum
requirements of load demands of electricity during different hours of the day. The figure 1 shows
that load demand during morning and night hours is more in winter seasons. Thus supply shall
accordingly be made as required which will smooth the system and avoid to load-shedding.
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x 10
19
Hours: hours
Demand
Load Demand for Winter Season(January 01 2008(Hourly Basis)
Morning
Beforenoon
Afternoon
Evening
Night

98
Table 1: Performance Using 10 folds Cross Validation
The Table 1 computes the performance using 10 fold cross validation for Nov. Dec. data 2008 as
training and 2009 as testing and vice versa as well as data of July to Dec.2008 and 2009
accordingly. The accuracy and error computed in each case is shown in the table 1.
Real time forecasting system performance is measured on similar load situations as well as on
different load situations in working days, weekends and seasonal demands.
Various techniques are appliedinvolving Support Vector Machine (SVM) kernels, Linear
Discriminant Analysis (LDA), QuadraticDiscriminant Analysis (QDA), the re substitution error is
calculated in each case. It is noticed that SVM Radial Base Function Kernel using 10 fold cross
validation gives the best performance of 99 %, however, other techniques also have almost
similar results as shown in Table 1 and Table 2.
In order to analyze the performance indexes, the actual and predicted values are computed using
Mean Square Error (MSE), Relative Error (RE) and Mean Absolute Percentage Error (MAPE).
The performance of forecasting effect using RE is computed using following formula.
( ) =
( ) − ( )
( )
× 100 % (1)
Where, ( ) is actual value and ( ) is the forecasting value.
The most commonly used aggregative indicator in power systems is MAPE, used to evaluate the
forecasting performance of whole predicting process in a more comprehensive way.
Mathematically,
=
1
|
( ) − )
( )
| (2)
Where, N is the sample number.
Support Vector Machine (SVM), kernels applied on load forecasting data are polynomial,
Gaussian, Highly Radial Base Function and Multi Quadratic. Beside it, Linear Discriminant
Analysis (LDA) and Quadratic Discriminant Analysis (QDA) techniques are applied. The MAPE

99
Calculated in each case is shown in table 2. The results show that SVM MQ gives 99.53 %
performance of actual and forecasting data for Nov. Dec. 2008 and 2009 than other kernels.
Table 2: Performance using Re Substitution Error
Data for the
month
Technique Error Value Error Type
Nov.Dec.2008
and 2009
10 fold CV 0.0014 RSE
Stratified 10 fold CV 0.0027 -do-
LDA 0.01 -do-
QDA 0.01 -do-
10 fold CV 1.7203 MSE
SVM poly 0.7967 RE
SVM Polynomial 0.0088 MAPE
SVM Gaussian 0.0148 -do-
SVM HRBF 0.0216 -do-
SVM MQ 0.0047 -do-
SVM: Support Vector Machine, CV: Cross Validation, RSE: Re-substitution Error, MSE: Mean
Square Error, RE: Relative Error, MAPE: Mean Absolute Percentage Error, HRBF: Highly
Trained Radial Base Function, MQ: Multi Quadratic Kernel
The forecasting performance measured using RE and MAPE as shown in table 2. The figure 2
shows the corresponding plot of actual and forecasting values during November and December
2008 and 2009 where 2008 values are used as actual and 2009 as forecasting values. The
corresponding RE and MAPE are computed which are 0.7967 and 0.0047 least in case MQ kernel
respectively. The RE and MAPE are computed using SVM kernel which gives above 99 % of
performance accuracy. The similar results are to be obtained for other seasons, working and non-
working days.
Table 3: Electricity Load Demands of Working and Non-Working Days.

100
The Table 3 above shows the electricity demands of working and non-working days during the
winter season November& December 2008 and 2009. The maximum,minimum and average
demand is computed. The results in the table shows that the demand during both 2008 and 2009
of working days is more than that of non-working days in Nov. Dec. As maximum demand of
electricity of working days during 2008 is 2865 MW while in 2009 it is 2695 MW. However, it is
2432 MW and 2311 MW during non-working days accordingly. The same prediction can be
achieved for other seasons of working and non-working days.
Figure 2: Actual and predicted value plot of November and December 2008 and 2009
The table 3 shows the maximum, minimum and average demands of load. The figure 3 below
shows the load demand in graphical form for first week working days of January 2008. The
varying amplitude shows the different demands during morning, before noon, afternoon, evening
and night as well as load demands during working days in a week in January.

101
Figure 3: Load Demand for January 1st
working days 2008
6. CONCLUSION
Support Vector Machine (SVM) kernels are used for Short Term Load Forecasting (STLF) on
data. The results are computed and performance is measured using 10 fold cross validation, re-
substitution error, Mean Square Error, Relative Error and MAPE. The results obtained from SVM
are compared with LDA and QDA. SVM kernel gives higher performance of 99%. The SVM
kernel approach is suggested for STLF in seasonal demands and can have the adaptability on
different load situations such as working days and non-working days. Moreover, STLF can help
in generation dispatch, capacity switching, load management program, feeder reconfiguration,
voltage control and automatic generation control (AGC). In addition, it will help to establish
operational plans for power stations and their generation unit and to plan market energy
transaction and load management.
7. FUTURE WORK
For Medium Term and Long Term Forecasting data required is meteorological information such
as temperature, wind speed, illumination, special events and data from different locations.
Medium term forecasting
Data required description of appliances used by customers, customer behavior, size of house, age
of equipment, technology change and population dynamics used in end use application. In
addition, other factors required are economic factors such as per capita income, employment level
and electricity pricing.
Long term Forecasting
Forecasting on population growth, economic development, industrial construction and technology
development

102
REFERENCES
[1] Pai, Ping-Feng, and Wei-Chiang Hong. "Support vector machines with simulated annealing
algorithms in electricity load forecasting." Energy Conversion and Management 46.17 (2005): 2669-
2688.
[2] Pai, Ping-Feng, and Wei-Chiang Hong. "Forecasting regional electricity load based on recurrent
support vector machines with genetic algorithms." Electric Power Systems Research 74.3 (2005):
417-425.
[3] Amjady, N., and F_ Keynia. "Short-term load forecasting of power systems by combination of
wavelet transform and neuro-evolutionary algorithm." Energy 34.1 (2009): 46-57.
[4] Hong, Wei-Chiang. "Chaotic particle swarm optimization algorithm in a support vector regression
electric load forecasting model." Energy Conversion and Management 50.1 (2009): 105-117.
[5] Huang, Wei, YoshiteruNakamori, and Shou-Yang Wang. "Forecasting stock market movement
direction with support vector machine." Computers & Operations Research 32.10 (2005): 2513-2522.
[6] Gao, Ciwei, et al. "Price forecast in the competitive electricity market by support vector
machine." Physica A: Statistical Mechanics and its Applications382.1 (2007): 98-113.
[7] Taylor, James W. "Triple seasonal methods for short-term electricity demand forecasting." European
Journal of Operational Research 204.1 (2010): 139-152.
[8] Zhou, Junyi, Jing Shi, and Gong Li. "Fine tuning support vector machines for short-term wind speed
forecasting." Energy Conversion and Management 52.4 (2011): 1990-1998.
[9] Salcedo-Sanz, Sancho, et al. "Short term wind speed prediction based on evolutionary support vector
regression algorithms." Expert Systems with Applications 38.4 (2011): 4052-4057.
[10] Foley, Aoife M., et al. "Current methods and advances in forecasting of wind power
generation." Renewable Energy 37.1 (2012): 1-8.

Short term load forecasting system based on support vector kernel methods

More Related Content

What's hot (15)

Viewers also liked (20)

Similar to Short term load forecasting system based on support vector kernel methods (20)

Recently uploaded (20)

Short term load forecasting system based on support vector kernel methods