1

A Nave Multiple Linear Regression Benchmark

for Short Term Load Forecasting
Tao Hong, Pu Wang, and H. Lee Willis, Fellow, IEEE

algorithms and techniques developed from other fields, of

AbstractBenchmarking issue in short term load forecasting which most are still on the theoretical level and of limited
has not received as much attention as it deserves. Although practical value to the industry.
dozens of techniques have been reported to be applied to short Benchmarking has been an issue in short term load
term load forecasting, most of them are still on the theoretical
forecasting (STLF) for decades. However, it has not received
level with insignificant practical value. None of them has been
established to produce benchmarking models for comparative as much attention as it deserves. As a result, there is not an
assessment. This paper proposes a nave multiple linear approach well established to produce benchmarking models
regression benchmark for short term load forecasting, which is for comparative assessment. Although regression analysis has
from the experience of helping a US utility develop the first in- been recognized as a legitimate tool to produce benchmarking
house short term load forecasts. The proposed model has been models, the use of regression analysis varies from one to
served as a benchmark for this utility since 2009, and was in
another [4, 7, 9, 11]. Most of the regression based approach
production use for a year with satisfying performance before a
major upgrade. It has also been used for a Canadian utility for segment the data and build separate models for different
load forecasting purposes. In addition, it was reproduced by a seasons, months, days of the week, or hours of the day.
group of graduate students from a creditable US university Different empirical rules, which may be specific to one utility,
following the documented procedure. have been involved to segment the data or fine tune the model,
so procedures are not easy to replicate to other utilities.
Index Termsload forecasting, power systems planning. Therefore, most of them are not eligible to be a benchmark.
This paper presents the benchmarking process, as part of
I. INTRODUCTION the process to start up the planning department, for a fast

F orecasts are always wrong, but some are useful. To an

electric utility, the useful forecasts should at least serve its
business needs, such as operations, planning, and energy
growing US utility. A multiple linear regression (MLR) based
load forecasting approach is demonstrated to produce the first
in-house forecasts for this utility. The resulting model was
purchasing, etc. Furthermore, the forecasts are supposed to be also used for the utility to determine whether or not to install
accurate, interpretable, and defensible. There is a common their own weather stations to collect weather data [6]. Due to
misunderstanding in the field of short term load forecasting: the generic set up of the regressors, the proposed model can be
overwhelmingly emphasizing on accuracy. In fact, as long as served as a nave MLR benchmark for a wide range of utilities
the utilities can be operated with competency in the market, that are accessible to two to three years of hourly temperature
the accuracy of load forecasts is not the most important factor and load data.
to utilities. In other words, minor improvement of forecasting This paper is organized as follows: Section II introduces
accuracy beyond certain level may not result in significant the theoretical background of MLR and general linear models
economic benefits [10]. Different utilities, depending upon the (GLM); Section III describes the details of the proposed
size, location and electricity consumption behaviors of its model; Section IV presents the forecasting accuracy and
customers, may have various comfortable zones of forecasting discuses the interpretation of the model; Section V concludes
accuracy, which may range from 2% to 6% with respect to the the paper with discussion of the future work.
mean absolute percentage error (MAPE) of one day ahead
hourly load forecast. II. MULTIPLE LINEAR REGRESSION ANALYSIS FOR STLF
There are roughly two approaches to improve the accuracy
A. General Linear Regression Models [8]
of the forecasts. One is to address and resolve the practical
issues, such as forecasting for large utilities with diverse The general linear regression model with normal error
meteorological regions [2], data cleansing problem [1], and terms can be defined as:
developing the forecast for some particular business needs Yi = 0 + 1 X i1 + 2 X i 2 + L + p 1 X i , p 1 + ei (1)
[12], etc, which in turn strengthen the defensibility of the where 0 L p1 are the parameters, X i1 L X i , p1 are the known
forecasts. The other one is to customize and combine new constants, ei is the independent normally distributed random
T. Hong is with Quanta Technology, LLC, Raleigh, NC 27607 USA (e- variable N (0, 2 ) . Then the response function is
mail: [email protected]). E[Y ] = 0 + 1 X 1 + 2 X 2 + L + p 1 X p1 (2)
P. Wang is with SAS Institute, Inc., Cary, NC 27513 USA
H. L. Willis is with Quanta Technology, LLC, Raleigh, NC 27607 USA.
 2

where X 1 L X p1 are p 1 predictor variables. Therefore, the It is a special case of (1), because it can be written as
definition (1) implies that the observations Yi are independent Yi = 0 + 1 X i1 + 2 X i 2 + 3 X i 3 + ei (10)
normal variables, with mean E[Yi ] as given by (2) and constant where X i1 = X i , X i 2 = X i2
, and X i 3 = . X i3

variance 2 . Notice that the squared or higher-ordered terms of a

qualitative predictor variable is equivalent to the 1st ordered
B. Quantitative and Qualitative Predictor Variables one.
In many cases, the predictor variables are quantitative. For
D. Transformed Variables
example, as the customer count (number of customers in the
utilitys service territory) increases, the load shows an In the polynomial regression models, the predictor
increasing pattern. If we model the load as a linear function of variables can be transformed to reach the standard form of the
the customer count, the customer count can be considered as a GLM. The response variable can also be transformed to model
quantitative predictor variable. some complex, curvilinear response functions. Considering
However, the definition of (1) does not limit the predictor the polynomial regression model of the relationship between
variables to the quantitative ones. Qualitative predictor the load and temperature, a model with a transformed Y
variables, sometimes called class variables or dummy variable can be written as:
ln(Yi ) = 0 + 1 X i1 + 2 X i 2 + 3 X i 3 + ei (11)
variables, such as weekday or weekend, can also be included
in the model. Indicator variables with values 0 and 1 can be Again, it is also a special case of (1). If we let Yi ' = ln(Yi ) ,
used to identify the classes of a quantitative variable. For the model (11) can be written as
instance, if the load (Y) is predicted based on whether it falls Yi ' = 0 + 1 X i1 + 2 X i 2 + 3 X i 3 + ei (12)
in a weekday or weekend, a qualitative predictor variable X 1
can be defined as follows: E. Interaction Effects
X 1 = 1, if the day is a weekday When the effects of one predictor variables depend on the
(3) level(s) of some other predictor variable(s), interaction effects
X 1 = 0, if the day is a weekend can be included in the GLM. Such models can also be called
Then the response function is nonadditive regression models. The interaction terms can be
E[Y ] = 0 + 1 X 1 (4) implemented by the multiplying two or more predictor
For the load in a weekday, X 1 = 1 , and (4) becomes variables. An example of the nonadditive regression model
E[Y ] = 0 + 1 (5) with two predictor variables X 1 and X 2 is the following:
For the load in a weekend, X 1 = 0 , and (4) becomes Yi = 0 + 1 X i1 + 2 X i 2 + 3 X i1 X i 2 + ei (13)
E[Y ] = 0 (6) It is still a special case of the GLM. By letting
X i 3 = X i1 X i 2 , the model (13) can be written as
In general, a qualitative variable with c classes can be
represented by c-1 indicator variables. For example, a Yi = 0 + 1 X i1 + 2 X i 2 + 3 X i 3 + ei (14)
qualitative variable day of the week with 7 classes (Sunday,
F. Linear Model vs. Linear Response Surface
Monday, , Saturday) can be represented as follows by 6
indicator variables: There should be no ambiguousness that the GLM can be
X 1 = 1, if the day is Sunday used to generate a large variety of nonlinear response surfaces.
In other words, linear models are not restricted to linear
X 1 = 0, otherwise
response surfaces. The term linear in GLM refers to the
X 2 = 1, if the day is Monday
parameters. A regression model is linear in the parameters
X 2 = 0, otherwise (7) when it can be written as:


L Y = XB + E (15)
X 6 = 1, if the day is Friday where Y is a vector of responses, B is a vector of parameters, X
is a matrix of constants, and E is a vector of independent
X 6 = 0, otherwise
normal random variables with zero expectation and variance-
Then the response function of the regression model with day
of the week as the predictor variable is covariance matrix 2 I .
E[Y ] = 0 + 1 X 1 + 2 X 2 + L + 6 X 6 (8)
III. COMPONENTS FOR THE NAVE BENCHMARK
C. Polynomial Regression
A. Data
Polynomial regression models contain polynomial(s) of the
The benchmarking approach proposed in this paper
predictor variable(s) making the response function curvilinear.
requires two sets of data: hourly load at system level and
For example, if the load (Y) is predicted by a polynomial
hourly temperature, which most utilities have access to.
regression model with one predictor variable temperature
Depending upon the maturity of the data management, the
( X i ), and the order of this polynomial is three [3], the
data quality may vary. To replicate the benchmarking model,
following model can be considered: the utility ought to have at least two years of history data with
Yi = 0 + 1 X i + 2 X i2 + 3 X i3 + ei (9) reasonably good quality. In this paper, four years (2005-2008)
 3

of hourly load (in kW) and temperature (in F) data from a polynomials were suggested in [3]. Fig. 3 shows the load
medium utility are used in the case study. The first three years temperature scatter plot of the utility in this study. The
are used as the history data, while the last year is the holdout piecewise functions need the cut-off temperature(s), which
sample. Fig. 1 shows the load series from 2005 to 2008, while may not be exactly the same in different service territories.
Fig. 2 shows the corresponding temperature series. The southern part of the United States may have a different
cut-off temperature from the northern part, because the
B. Linear Trend
comfortable temperature zone may be different for people
We define a quantitative variable (Trend) to capture the living in different regions. Therefore, the 3rd ordered
locally increasing (or decreasing) trend by assigning a natural polynomials of the temperature are used to predict the load for
number to each hour in ascending order. For instance, the the benchmarking purpose.
Trend variable of the first hour in 2005 is 1, the second hour in
2005 is 2, and the last hour of 2008 is 35064. It should be D. Calendar Variables
noticed that such a trend only belongs to the utilities with a It is well known that there are three seasonal blocks in the
stable service territory and local economics. The linear trend load series: day, week, and year. There can be different
can be interpreted as the linear approximation to the load treatments to each block depending upon the load
series: when the length of load history is relatively short consumption behavior of the particular service territory. For
comparing to the macroeconomic changes, such as recession example, the 7 days of a week can be modeled by a qualitative
and booming, the overall trend of the load during the interval variable with 2 classes (weekdays and weekends), 3 classes
of interest can be linearly approximated. The benchmarking (weekdays, two weekend classes), etc. In different countries,
model discussed in this paper is not directly applicable to the the weekends may be defined differently. For instance,
significant business events, such as merging two utilities and Thursday and Friday are weekends in Iran. As a benchmarking
splitting one utility into two. model, the qualitative variables (Hour, Day, and Month) with
24, 7, and 12 classes, are used to model the 24 hours of a day,
C. Temperature
7 days of a week, and 12 months of a year respectively.
The relationship between the load and temperature has been Namely each seasonal block is decomposed into the unit of the
intensively studied during the past several decades. For highest resolution.
instance, piecewise linear function was indicated in [2];
piecewise quadratic function was used in [5]; the 3rd ordered

Fig. 1. Load series (2005-2008).

 4

Fig. 2. Temperature series (2005-2008).

Fig. 3. Load-temperature scatter plot (2005-2007).

 5

E. Interaction Effects statistical software packages, such as SAS 9.2 with STAT.
Normally an afternoon is warmer than a midnight; a
summer is warmer than a winter. In other words, temperature IV. RESULTS AND DISCUSSIONS
is not independent of the hour of the day or the month of the This section presents the forecasting performance of the
year. Therefore, the interaction effects between the above benchmarking model GLMLF-B. The following error
temperature (appeared in the form of 3rd ordered polynomial analysis is often used in forecasting:
as discussed above) and the calendar variables Hour and 1) Error, the difference between a quantity and its estimated
Month should be in the model. Since it is not evidential that or measured quantity.
there is a relationship between days of the week and the 2) Absolute Error, the absolute value of error.
temperature, the interaction between temperature and days of 3) Absolute Percentage Error, 100% times the relative error
the week is not included in the model. (the error divided by the true value).
The hours in different days of a week may result in The distribution of the above errors can be characterized by
different load due to human activities. For instance, there may mean, standard deviation, minimum, Q1 (first quartile),
be lower load in the morning of the weekends than the other median, Q3 (third quartile), and maximum values.
mornings, because people do not have to get up as early as Several engineering concepts have been involved when the
weekdays to go to work, which results in less load at home load forecasts are communicated:
and office buildings in the weekend mornings. Therefore, the 1) Hourly load, the energy consumed in an hour. Sometimes
interaction effect between Hour and Day should be included in it is calculated by averaging several instantaneous
the benchmark model. measurements occurred in an hour.
When a qualitative predictor variable interacts with a 2) Energy, the summation of the hourly load within a
quantitative one, this quantitative term is not required to specific period.
appear in the model as a single regressor (main effect). When 3) Peak/valley load, the maximum/minimum of the hourly
two qualitative predictor variables interact together, either of load within a specific period.
them is not required in the model as a main effect. Therefore, 4) Peak/valley hour load, the load occurs during the hour
the benchmarking model GLMLF-B includes the follows: where actual peak/valley load occurs.
1) Quantitative variables: Trend, and TMP (the current hour
temperature); TABLE I
2) Class variables: Hour, Day, Month; RESULTS (MAPE, %) OF THE NAVE MODEL
Forecasting horizon (# of days)
3) Main effects: Trend, Month; 1 2 3 4 5 6 7
4) Interaction effects (also known as cross effects): Hourly load 4.98 5.00 5.01 5.01 5.02 5.03 5.04
DayHour, MonthTMP, MonthTMP2, MonthTMP3, Daily peak hour 4.27 4.29 4.29 4.29 4.30 4.30 4.30
HourTMP, HourTMP2, HourTMP3, where the cross sign Daily valley hour 5.44 5.46 5.47 5.49 5.50 5.50 5.52
Daily peak 3.94 3.96 3.97 3.97 3.98 3.99 3.99
represents the interaction effect; Daily valley 4.93 4.95 4.96 4.98 4.99 5.00 5.02
5) Intercept. Daily Energy 3.49 3.51 3.52 3.53 3.53 3.54 3.55
The model can be written as:
E(Load) = 0 + 1Trend + 2DayHour + 3Month +
4MonthTMP + 5MonthTMP2 + 6MonthTMP3 +
7HourTMP + 8 HourTMP2 + 9 HourTMP3 , (16)
With the above information, the model GLMLF-B as
specified in (16) can be easily implemented in commercial

Fig. 4. Forecasting performance during the peak week of 2008.

 6

The various combinations of the above error analysis and [7] O. Hyde and P. F. Hodnett, "An adaptable automated procedure for
short-term electricity load forecasting," IEEE Transactions on Power
engineering concepts, together with the updating cycle and Systems, vol. 12, pp. 84-94, 1997.
forecast horizon, form the diagnostic statistics for each load [8] H. M. Kutner, J. C. Nachtsheim, J. Neter, and W. Li, Applied Linear
forecast. In this paper, to present the performance of a short Statistical Models, 5th ed: McGraw-Hill/Irwin, 2004.
[9] A. D. Papalexopoulos and T. C. Hesterberg, "A regression-based
term load forecast, the mean absolute percentage errors approach to short-term system load forecasting," IEEE Transactions on
(MAPE) of the hourly load, daily energy, daily peak/valley Power Systems, vol. 5, pp. 1535-1547, 1990.
load, and daily peak/valley hour load can be performed with [10] D. K. Ranaweera, G. G. Karady, and R. G. Farmer, "Economic impact
the updating cycle ranging from one day to seven days, as analysis of load forecasting," IEEE Transactions on Power Systems, vol.
12, pp. 1388-1392, 1997.
listed in Table I. Rolling regression was applied to generate [11] S. Ruzic, A. Vuckovic, and N. Nikolic, "Weather sensitive method for
the forecasts for different horizons. For instance, when the short term load forecasting in Electric Power Utility of Serbia," IEEE
forecasting horizon is one day, the model is updated every day Transactions on Power Systems, vol. 18, pp. 1581-1586, 2003.
[12] T. Saksornchai, W.-J. Lee, K. Methaprayoon, J. R. Liao, and R. J. Ross,
and forecasts the loads of the next day. In every update, the "Improve the unit commitment scheduling by using the neural-network-
newly available loads and temperatures are added to the based short-term load forecasting," Industry Applications, IEEE
history data. The forecasted and actual loads of the peak week Transactions on, vol. 41, pp. 169-179, 2005.
of 2008 are shown in Fig. 4.
The results listed in Table I and Fig. 4 dont seem to be
VII. BIOGRAPHIES
attractive comparing with the ones reported in the literature. In
Tao Hong is a Principal Engineer in Quanta Technology. His major areas of
practice, due to the weather forecast errors, the actual MAPEs expertise are in operations research and its related applications in power
are even worse. A major reason is that they are produced out distribution engineering and T&D planning. He has applied various statistics
of a nave model, which does not capture the specific local and optimization techniques to T&D loss studies, development of long term
and short term load forecasting algorithms and tools. His work has been
electricity consumption behaviors, such as weekend effect, applied to many utilities in US and Canada. He received his Bachelor of
holiday effect, etc. However, it should be noticed that the Engineering degree in Automation from Tsinghua University, Beijing, a
predictability of the load series from different utilities varies. Master of Science degree in Electrical Engineering, a Master of Science
For instance, the utility was fairly comfortable with this degree with co-majors in Operations Research and Industrial Engineering, a
PhD degree with co-majors in Operations Research and Electrical
forecasting performance, which was much superior to the ones Engineering, all from North Carolina State University. Dr. Tao Hong is a
produced by several internal and external parties. The winner of the poster competition in SAS 12th Annual Data Mining
proposed model served as a production model in this utility for Conference in 2009 for his work Behavior Mining of Electric Load
Consumption: A Regression Approach.
a year, before it was replaced by a customized one. Pu Wang is a research statistician developer in SAS Institute. She is
specialized in operations research and statistical analysis with the applications
V. CONCLUSION in merchandise intelligence solutions. She has been doing research and
development of emerging techniques in market response modeling and
This paper proposes a nave MLP based benchmarking demand forecasting for fashion goods retailers. She received her Bachelor of
approach for STLF. The proposed model is designed to be Engineering degree in Industrial Engineering from Tsinghua University,
Beijing, where she was awarded with a first-class scholarship for the academic
generic and applicable to a wide range of utilities. The work excellence. She received the Master and PhD degrees in Industrial
presented in this paper is based on the experience of helping a Engineering from North Carolina State University.
US utility develop the first in-house short term load forecasts. H. Lee Willis, PE, (F92) is a Senior Vice President and Executive Advisor in
Quanta Technology. He has more than 35 years of electric T&D systems
The benchmarking process has been implemented in this planning and engineering experience. He has been transmission planning
utility. Future work of this paper is to customize the proposed manager at a major investor-owned utility, an executive with a major
nave benchmark to capture the local electricity consumption equipment supplier, and a senior consultant who has directly performed or
supervised over 400 system planning and asset management projects for
behavior of this utility. utilities around the world. Lee pioneered many of the modern planning and
asset management methods now considered industry best practice, including
VI. REFERENCES spatial simulation load forecasting for T&D planning, load-reach voltage
planning of feeder system layouts, and options-based asset management
[1] J. Chen, W. Li, A. Lau, J. Cao, and K. Wang, "Automated Load Curve
prioritization. He is a Fellow of the IEEE and a past chairman of its Power
Data Cleansing in Power Systems," Smart Grid, IEEE Transactions on,
System Planning and Implementation committee. From 1999 to 2005 he
vol. 1, pp. 213-221, 2010.
served on the US National Research Council, which advises the US Congress
[2] S. Fan, K. Methaprayoon, and W.-J. Lee, "Multiregion Load Forecasting
and Dept. of Commerce on the nations civilian technology needs. Lee has
for System With Large Geographical Area," IEEE Transactions on
published more than 240 papers and articles on power systems, forecasting
Industry Applications, vol. 45, pp. 1452-1459, 2009.
and planning, and utility practices during his career, including 78 papers in
[3] M. T. Hagan and S. M. Behr, "The Time Series Approach to Short Term
peer-reviewed and refereed technical journals, one theme cover article in the
Load Forecasting," IEEE Transactions on Power Systems, vol. 2, pp.
IEEE-wide Proceedings of the IEEE, and five cover articles for the IEEE
785-791, 1987.
Computer Applications in Power journal. He is the author of eight books on
[4] T. Haida and S. Muto, "Regression based peak load forecasting using a
power system planning including the Power Distribution Planning Reference
transformation technique," IEEE Transactions on Power Systems, vol. 9,
Book and Spatial Electric Load Forecasting.
pp. 1788-1794, 1994.
[5] T. Hong, M. Gui, M. E. Baran, and H. L. Willis, "Modeling and
forecasting hourly electric load by multiple linear regression with
interactions," presented at 2010 IEEE Power and Energy Society
General Meeting, Minneapolis, Minnesota, USA, July 25-29, 2010.
[6] T. Hong, P. Wang, A. Pahwa, M. Gui, and S. M. Hsiang, "Cost of
temperature history data uncertainties in short term electric load
forecasting," presented at 2010 IEEE 11th International Conference on
Probabilistic Methods Applied to Power Systems (PMAPS2010),
Singapore, June 14-17, 2010.