See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224309206

Short term load forecasting using Multiple Linear Regression

Conference Paper · October 2007

DOI: 10.1109/UPEC.2007.4469121 · Source: IEEE Xplore

CITATIONS READS

170 6,522

3 authors, including:

Cuneyt Suheyl Ozveren David J. King

71 PUBLICATIONS 710 CITATIONS

Abertay University
37 PUBLICATIONS 568 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Cuneyt Suheyl Ozveren on 01 July 2015.

The user has requested enhancement of the downloaded file.

 Short Term Load Forecasting using Multiple Linear Regression

N. Amral, C.S. Özveren, D King

University of Abertay Dundee, UK

Abstract

In this paper we present an investigation for the short term (up 24 hours) load forecasting of the
demand for the South Sulewesi’s (Sulewesi Island – Indonesia) Power System, using a Multiple
Linear Regression (MLR) method. After a brief analytical discussion of the technique, the usage
of polynomial terms and the steps to compose the MLR model will be explained. Report on
TM
implementation of MLR algorithm using commercially available tool such as Microsoft EXCEL
will also be discussed. As a case study, historical data consisting of hourly load demand and
temperatures of South Sulawesi electrical system will be used, to forecast the short term load.
The results will be presented and analysed potential for improvement using alternative methods is
also discussed.

Key words: Multiple Linear Regression, polynomial terms.

1). Peak load models; in these models, the

daily or weekly peak load is modelled, usually
1 INTRODUCTION as a function of weather. Time does not play a
role in such models. Peak load models can be
Due to its major role in an effective and economic divided into two parts: a base load that
operation of power utilities, electric load independent to the weather and variable load
which is weather dependent, this component
forecasting has received increasing attention over
is then added to the base load [2], [3]. The
the years by academic researchers as well as the
power systems engineers in the industry. parameters of the model are estimated
through linear or nonlinear regression. These
models do not define the time which the peak
An accurate short term load forecasts (STLF) with occurs and contain information about the
forecasting horizons up to 24 hours are necessary shape of the load curve. Correlation across
for scheduling functions such as hydro-thermal the period cannot be forecast as the models
power generation coordination (unit commitment) are essentially static.
in order to establish the hourly schedules for
generation resources that will minimize the system
operating cost. Such forecasts are also used for 2). Load shape models are observed in
particular time intervals, which are usually
economic dispatch, predictive frequency control,
sampled every one hour or one-half hourly
security analysis, systems restoration, and energy
intervals. These models can be classified into:
trading. Recent global privatisation and
deregulation initiatives in the sector has made
accurate forecasting vital not only for operational i). Static models which is described as a
but also for the profitability and sustainability of combination of explicit time function such as
privatised /deregulated utilities. sinusoids, exponential or polynomials, with
model parameters are calculated through
Because of its crucial function in the operation of the usage of a set of historical load data into
Electrical Power Systems, there are many authors linear regression or exponential smoothing
techniques [4],[5].
in the literature reporting a wide variety of methods
for short term load forecasting. Typically, the
models are classified into two basic models[1], ii). Dynamic models which consider that load
these are: is also affected by its most recent behaviour

UPEC 2007 - 1192

 as well as weather and random inputs and it is The error term ε has a mean value equal to
not only a function of the time of the day. This zero and constant variance.
model can be composed with autoregressive
moving average (ARMA) and state-space
models [1] and Artificial Neural Network (ANN)
Since parameters βi are unknown, they
[6] should be estimated from observations of y
and xi . Let bi (i = 0,1,2,...k ) be the estimates
The investigation we report in this paper is part of
our wider investigations to find an appropriate set in terms of β i (i = 0,1,2,...k ) .
of methodologies for the STLF the demand of the
South Sulewesi’s Electrical Power System. A brief Hence the predicted value of y is:
analysis and evaluation of five short term load
forecasting techniques, that we will be
investigating including Regression based yˆ = bo + b1 x1 + b2 x 2 + ... + bk x k (2)
approaches, Stochastic Time Series, General
Exponential Smoothing, State Space Method,and
The difference between the actual load value
Expert Systems, was presented in [7]. In this
paper we report on our investigations on the of y and the predicted value ŷ would, on
regression based approach to short term system average, tend toward 0, for this reason it can
load forecasting. This approach to forecasting has be assumed that the error term in equation (1)
been used by different authors including [8]. In the has an average, or expected, value of 0 if the
following we briefly explain the MLR technique and probability distributions for the dependent
detail our model and methodology. variable y at the various level of the
independent variable are normally distributed
(bell shaped). the error term can therefore be
omited in calculating parameters.
2 MULTIPLE LINEAR REGRESSIONS
Then, the least square estimates method is
Regression analysis is a modelling technique for used to minimize the sum of squared residuals
analysing the relationship between a continuous (SSE) to obtain the parameters bi :
(real-valued) dependent variable y and one or
more independent variables x1 , x 2 , L x k . The goal
in regression analysis is to identify a function that
B = [bo b1
T
(
b2 ...bk ] = X X
T
)−1 T
X Y (3)
describes, as closely as possible, the relationship
between these variables so that the value of the Where Y and X are the following column
dependent variables can be predicted using a vector and matrix:
range of independent variables values

In the multiple linear regression method, the load  y1  1 x11 x12 L x1k 
is found in terms of explanatory (independent) y  1 x x 22 L x 2 k 
variable such as weather and other variables Y =  2  And X =  21

which influence the electrical load. The load model M M M M M 

using this method is expressed in the form as    
 yn  1 x n1 xn 2 L x nk 

y = β 0 + β 1 x1 + β 2 x 2 + ... + β k x k + ε (1)

Where y is the load, xi is the affecting factors, After parameters are calculated, this model
can be used for prediction. Assuming that all
βi is regression parameters with respect to xi , the independent variables have been correctly
identified and therefor the standard error will
and ε is an error term. be small. The standard error is obtained by the
equation below:

UPEC 2007 - 1193

 SSE load temp
s= (4)
n − (k + 1) 350,0 40
300,0 35
250,0 30

Load (MW)

Temp (C)
25
SSE = ∑ ( yi (t ) − yˆ i (t ) )
2 200,0
(5) 150,0
20
15
100,0 10
50,0 5
0,0 0
y i (t ) : observed, yˆ i (t ) : estimated 1 3 5 7 9 11 13 15 17 19 21 23
Time
n

∑ ( y (t ) − yˆ (t ))
i i
2

R2 = 1− t =1
n
(6) Figure 1 Daily load and temperature curve

∑ ( y (t ) − y (t ))
t =1
i i
2
For this forecasting study, data during the
rainy season and dry season were used
separately. In the MLR application, the hourly
y i (t ) : the average value of yi (t ) load is modelled as: (i) Intercept component
which is assumed constant for different time
Experience about the load to be modelled helps intervals of the day, (ii) Time of observation
an initial identification of the candidate of which is represent the load characteristic of
explanatory (independent) variables. To determine the day. (iii) Temperature sensitive component
the significance of regression parameters, the F which is function of difference of temperature
statistical test is performed and to determine the at time t and the average temperature in time
significance of each of these coefficients; is intervals. The relationship between the
performed by calculating t ratios. A goodness of fit temperature sensitive component, time of
measurement is represented by the R2 statistic observation and the load fluctuation are not
which ranges from 0 to 1 and indicates the linear but in form of polynomial term.
proportion of the total variation in the dependent
variable Y around its average that is counted for The model that is used divides the load curve
by the independent variable in the estimated into 3 intervals, 1 to 6 am, 7am to 17pm and
2 18 to 24 pm.
regression function. The closer the R statistic to
the value 1, the better the estimated regression
function fits the data. For the interval 1 to 6 am, 1 ≤ t ≤ 6 , the
hourly load can be modelled as follows

yˆ i (t ) = b0 + b1 (Ti (t ) − Ti (t − 1))
3 IMPLEMENTATIONS
+ b2 (Ti (t − 1) − Ti (t − 2)) (7)
The MLR technique as has been discussed + b3t + b4 t 2
previously is implemented to predict the hourly
load of South Sulawesi Electricity system.
For interval 7am to 17 pm, 7 ≤ t ≤ 17 , the
South Sulawesi is one of the provinces of Republic load curve is modelled by the form:
of Indonesia. Indonesia is located on the equator
with tropical weather, and have only dry and rainy
season throughout the year. The energy sales
composition of the consumer base in 2005 was
49% residential, 16% commercial, 26% Industrial
and 9% public, with the typical daily load (in MW)
and temperature curve as follows:

UPEC 2007 - 1194

yˆ i (t ) = b0 + b1Ti (t ) + b2 (Ti (t ) − Tiav ) Table 1. Weekday MLR dry season model
parameter estimates for different time intervals
+ b3 (Ti (t ) − Tiav ) 2 of the day

+ b4 (Ti (t ) − Tiav ) 3 Day time interval

+ b5 t + b6 (Tava − Tavb ) (8) Parameter 1-6 7-17 18-24

+ b7 (Ti (t ) − Ti (t − 1)) b0 194,25 203,43 -30050,28

+ b8 (Ti (t − 1) − Ti (t − 2)) b1 4,33 0,00 -1,13

b2 -14,71 7,28 4277,08
+ b9 (Ti (t − 2) − Ti (t − 3))
b3 -15,79 -0,02 -198,86

And for interval 18 to 24 pm, 18 ≤ t ≤ 24 , the b4 1,94 -0,04 3,05

model is different because the load increase b5 - 0,00 -
sharply due to the decrease of sky brightness, b6 - 38,58 -
hence the consumer start to turn on their light. The
formula has the form: b7 - -4,04 -
b8 - -0,73 -
2
yˆ i (t ) = b0 + b1Ti (t ) + b2 t + b3t b9 - -0,22 -
(9) 2
+ b4 t 3 R 0,9028 0,9935 0,9898

Where:
Table 2. Weekday MLR rainy season model
yˆ i (t ) = predicted load at hour t in the interval i of parameter estimates for different time intervals
of the day
the day

Day time interval

b0 = Intercept component (regression constant
Parameter 1-6 7-17 18-24
coefficient)
b0 228,13 167,24 -29972,81
b1 8,25 0,00 0,19
b1 , L , b9 = Regression parameters of
b2 4,07 4,33 4214,20
temperature sensitive component and time of b3 -31,69 0,00 -193,82
observations
b4 4,24 -0,02 2,94
0
Ti (t ) = Temperature ( C) at time t in the interval i b5 - 2,77 -

of the day b6 - 1,50 -

b7 - 0,67 -
Tiav = Temperature average (0C) in the interval i of b8 - 3,73 -
the day. b9 - 4,12 -
2
R 0,9814 0,9821 0,9850
Tava = Average temperature of previous 24 hours
0
to the time t ( C)

Tavb = Tava lagged 3 hours, (0C) For this review, the model is derived both for
the dry season and the rainy season, with
parameters shown in table 1 and 2
t = Time of observation. respectively. These parameters regression
have been calculated using weekday hourly
data for every time interval. The division of the

UPEC 2007 - 1195

day into three unequal interval time zones is
Real 13Apr04 Forecast 13Apr04
based on the characteristics of load compare to
the temperature fluctuation
350,0
300,0
250,0
4 RESULT

L o a d (M W )
200,0

In order to investigate how effective the 150,0

implementations of the MLR method will be for 100,0
STLF in an actual real life network, this technique
50,0
has been applied to predict the daily load (up to 24
hours) on dry and rainy season days, in the South 0,0
Sulewesi Network. The error was calculated as 1 3 5 7 9 11 13 15 17 19 21 23
the mean average percentage error (MAPE). as Time
follows;

n
yt − yˆ t Figure 3. Rainy season load curve
∑
t =1 yt
MAPE = 100 ⋅ (10) Table 3. Actual, forecast load and mean
n average percentage error (MAPE) for dry
and rainy season
-The MAPE in the dry season, is 3,52 %
Dry Season Rainy Season
-The MAPE in the rainy season is 4,34 % Time Actual Forecast Error Actual Forecast Error
1 190,1 186,7 1,78 194,5 193,7 0,41
Figure 2 and 3 show the actual and predicted 2 176,5 166,4 5,75 181,5 177,6 2,15
curve in dry season and rainy season respectively. 3 168,3 171,9 2,10 176,4 166,7 5,49
Table 3 describes the actual load, forecast and 4 163,3 165,7 1,43 174,9 161,1 7,94
5 182,2 164,2 9,92 182,4 170,9 6,32
MAPE for dry and rainy season,
6 179,1 178,2 0,50 192,7 188,3 2,28
7 167,9 169,3 0,85 171,3 165,3 3,48
Real 5Oct04 Forecast 05Oct04 8 185,4 174,6 5,85 187,7 181,8 3,17
350 9 203,4 186,3 8,42 192,8 205,4 6,51
10 214,1 200,6 6,30 207,9 218,5 5,09
300
11 213,2 207,0 2,92 215,5 212,2 1,52
250 12 212,0 218,6 3,10 212,2 210,2 0,96
13 213,1 220,8 3,63 212,6 213,0 0,19
Load (MW)

200 14 216,9 222,1 2,41 219,8 215,7 1,88

150
15 224,7 217,8 3,07 214,6 213,6 0,48
16 215,3 211,9 1,54 211,3 214,8 1,62
100 17 207,4 202,6 2,31 171,4 216,3 26,19
18 263,4 267,2 1,46 219,0 254,0 16,00
50
19 332,9 320,9 3,63 307,1 319,2 3,94
0 20 320,1 324,1 1,26 322,6 332,2 2,97
1 3 5 7 9 11 13 15 17 19 21 23
21 299,1 295,0 1,37 311,5 310,8 0,24
Time
22 271,0 254,9 5,96 270,1 272,5 0,88
23 234,4 218,5 6,80 231,3 235,0 1,58
Figure 2. Dry season load curve
24 210,4 205,7 2,22 210,0 215,9 2,81
MAPE 3,52 MAPE 4,34

UPEC 2007 - 1196

 5 DISCUSSION The accuracy of predictions made using
regression models depends on how well the
The result of load forecasting contains several regression function fits the data, there should
component of errors, namely, modelling error be regular checks to see how well a
(error introduced to regression), error caused by regression function fits a given data set. This
system disturbances such as load shedding and can be done through regular updates or
irregular events and also errors of temperature monitoring to ensure that the error values are
forecast. always below a pre-spesified error threshold

This model is very sensitive to the fluctuation of A pre analysis of the load is necessary for
temperature. It needs a very accurate temperature successful MLR forecast as the result of the
forecast, as a small change of temperature will analysis showed, because in South Sulawesi
cause a significant change in load prediction. This system, the load is not only affected by the
forecasting model uses the next day temperature fluctuation of temperature but also by other
forecast as an input which will introduce further weather factors such as humidity and cloud
errors, As there were no temperature forecast data cover. It is the daylight change that changes
available, temperature forecast was not included the lighting levels in residential and
in this regression analysis. Further studies should commercial spaces and demand of consumers
include this in real life applications. according to the brightness of the day.

The load characteristic is not only affected by the In this paper we have reported our preliminary
temperature, but also affected by the other investigations the use of MLR for STLF of the
weather factors such as humidity, cloud cover and South Sulawesi Electrical System, and in our
brightness of the day. Knowledge based on pre- studies to date we have only modelled
analysis of the load, identified that in South weekdays in wet and dry seasons.
Sulawesi system, the dominant portion of the load Investigations are underway to model
is residential (lighting) and this load is dependent weekend and holidays.
mostly on the levels and duration of the daylight.
It can be seen from load curve that the load rises
sharply at 18 pm. Therefore the choice of time
observation (t) as an independent variable in 7 ACKNOWLEDGEMENTS
equation (7), (8) and (9) of the model, increased
the accuracy of the forecasting. The author would like to thank PT. PLN
(persero), the Indonesian state electricity
company for financing the study in University
of Abertay Dundee
6 CONCLUSIONS

To support the operation of electric power system

securely and economically, needs an accurate 8 REFERENCES
load forecasting method as the principal driving
element for all daily and weekly operations
1. Gross,G., Galiana, F.D, Short-Term Load
scheduling. Forecasting, Proceeding of IEEE,
December 1987.
The multi linear regression models for short term 2. Goh, T.N., Long, H.L., and Lee, Y.O., A
load forecasting are relatively easy to develop and New Approach to statistical Forecasting of
update regularly with widely available commercial daily Peak power Demand, Elect. Power
TM
computational software such as Microsoft Excel Syst. Res., Vol.10, No.2, pp 145-148,
as used in this investigations. The drawback of March 1986.
this model is its dependency on the accuracy of 3. Gupta, P.C., A Stochastic Approach to
previously recorded load (this will be a common Peak Power Demand Forecasting in
drawback for all STLF techniques) and Electric Utility Systems, IEEE Trans.
temperature data which will greatly affect the Power App. Syst., Vol. PAS-90,
forecast yield. March/April 1971.

UPEC 2007 - 1197

 4. Thomson, R.P., Weather Sensitive Electric
Demand and Energy Analysis on a large
geographically Diverse Power System –
Application to Short Term Hourly Electric
Demand Forecasting, IEEE Trans. Power App.
Syst., Vol. Pas-90, No 1, pp. 385-393,
Jan./Feb. 1976.
5. Christiaanse, W.R., Short Term Load
ForecastingUsing General Exponential
Smoothing, IEEE Trans. Power App. Syst.,
Vol. PAS-90, No. 2, pp 900-911, March/April
1971.
6. Park, D.C, El-Sharkawi, M.A, Marks II, R.J,
Atlas, L.E., Damborg, M,J., Electric Load
Forecasting Using An Artificial Neural
Network, IEEE Transaction on Power
Systems, Vol. 6, No. 2, pp 442-449, May
1991.
7. Moghram,I., Rahman, S., Analysis and
Evaluation of Five Short-Term Load
Forecasting Techniques, IEEE Transactions
on Power Systems, Vol.4, No.4, pp. 1484-
1491, October 1989.
8. Papalexopoulos, A.D., Hesterberg, T.C., A
Regression-Based Approach to Short-Term
Load Forecasting, IEEE Transactions on
Power Systems, Vol.5, No. 4, pp 1535-1547,
November 1990.
9. Ragsdale, C.T., Spreadsheet Modeling and
Decision Analysis a Practical Introduction to
Management Science, Course Technology
Inc, 1995.
10. Bowerman, B.L., O’Connel, R.T., Forecasting
and Time Series an Applied Approach,
rd
Duxbury, 3 Ed, 1993.

AUTHOR ADDRESS

The first author can be contacted at

Mphil Research Programme

Computing & Creative Technologies

University of Abertay Dundee

Bell Street, Dundee DD1 1HG

Email: 0604719@abertay.ac.uk

UPEC 2007 - 1198

View publication stats