Technical

Modelling weather-sensitive
electrical loads
Here Véronique Bugnion, Aram Sogomonian and Glen Swindle introduce a new
methodology for forecasting and jointly simulating temperature and electrical load

T
his article presents an integrated approach to modelling temperature, with commercially available neural-net and similar day-forecasting technol-
electrical loads and power prices. This modelling framework will be ogy, and has the advantage of being transparent and easily calibrated1.
used to analyse and manage the risks associated with serving vari- The approach involves two primary steps: regression of load sensitivity
able quantities of load. The objective of the analysis is a better understanding to temperature; and modelling of the residuals using factor analysis. The
and control of the earnings volatility associated with variable load risk. second step is the most challenging and interesting.
The short-term variability of electrical loads is driven by temperature,
season and time of day. Serving full-requirement load obligations involves Load sensitivity to temperature
assuming the uncertainty in the amount of power that will have to be deliv- The scatter plots shown in figures 1 to 3 exhibit Pennsylvania-New Jersey-
ered on a given day and hour. Forward power purchases can only hedge the Maryland (PJM) system loads versus the hourly temperature at Baltimore
quantity expected to be delivered, leaving the supplier with the risk associ- International airport (BWI) for January 1, 1998 to December 31, 2001. The
ated with serving the deviations – whether positive or negative – from the qualitative nature of the plot is well known – loads grow at both low tem-
expected load. This risk is termed variable load risk. perature and high temperature. Figure 1 shows the relationship for week-
The financial risks associated with serving variable load obligations are days at 4am, figure 2 the situation on weekends at 2pm and figure 3 the
compounded by the correlation between load and prices. High loads are situation on 2pm on weekdays. The plots suggest that the response to tem-
often accompanied by high prices and low loads by low prices. If a load sup- perature is in general different for each hour of the day. In addition, as is
plier is left short in a high-price environment or long when temperatures and well known, the response varies by day of the week, with weekends and
loads are moderate, its earnings will suffer. holidays exhibiting dramatically lower loads.
Here we introduce a new methodology for forecasting and jointly simu- With such obvious structure, it would seem that straightforward regression
lating temperatures and electrical loads. The mathematical analysis is pre- of load as a polynomial function of temperature and date – for load growth –
sented in conceptual terms. The applications chosen to illustrate the would be viable. However, such approaches result in undesirable structures at
methodology are load forecasting used to value full-requirement obligations the limits of the range of historical temperature data2. In such situations, stan-
and short-term load forecasts based on weather predictions. A later article dard statistical techniques involve data transformation.
will focus on simulating prices in the power markets and will present an A key assumption underpinning the following analysis is that the statis-
integrated approach to managing variable load risk and reducing the earn- tics of load – such as mean and variance – approach asymptotic values at very
ings volatility of companies serving full-requirement load obligations. low and very high temperatures. The justification for this assumption is of
course that there are a finite number of air conditioners, heaters and power-
Short-term load forecasting consuming devices in general. This transformation maps temperatures T onto
~ ~
This section sets out our analysis of the temperature-load relationship and its T ∈[0,1] and loads L onto L∈(–∞, ∞), thereby facilitating a simple polyno-
application to short-term load forecasting. The method compares favourably mial regression of the form

Figure 1: BWI temp. v. PJM load – 4am, weekdays Figure 2: BWI temp. v. PJM load – 2pm, weekdays

55 55

50 50

45 45
MWh (in '000s)

MWh (in '000s)

40 40

35 35

30 30

25 25

20 20

15 15
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Temperature (˚F) Temperature (˚F)

84 I July 2002
 Figure 3: BWI temp. v. PJM load – 2pm, weekends Figure 4: BWI temp. v. PJM load at 4pm on weekdays

55 55
Polynomial fit
50
50
45
MWh (in '000s)

MWh (in '000s)

45
40

35 40

30
35
25
30
20

15 25
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Temperature (˚F) Temperature (˚F)

J with the principal component analysis. Figures 6 and 7 show the eigenvalues
L
= ∑ α jT
j ➝
{λ j}24j=1 and the first few eigenvectors {v j} 24j=1 from principal components analy-
j =1 sis (PCA) for both weekday and weekend/holidays.
and eliminating undesired polynomial fluctuations beyond historically ob- The eigenvalues can be interpreted as the relative importance of the cor-
served extreme values of temperature. responding eigenvector. The rapid decay of the eigenvalues indicates that
The estimation of the statistical parameters describing the load’s relation the analysis of the residuals can be greatly simplified by keeping only the
to temperature takes place in several steps: first few. The rapid decay of the eigenvalues implies that we can describe
92% of the load variance using the first three factors and 98% using the
~ ~
● Preliminary polynomial regression of (T , L ). first six factors. To exploit this reduction in dimension, only the first six
● Linear or other regression of the weather-adjusted load residuals on date factors are kept to simulate and forecast loads in the analysis that follows.
to detect any trend. The first six eigenvalues are rescaled so that the resulting variance of hourly
~ ~
● Final polynomial regression of detrended (T , L ). residuals in this truncated factor representation is the same as that of the
original series.
This regression is done for each hour – 1 through 24 – of weekdays and The similarity of the eigenvalues and eigenvectors of the weekday and
weekends – including holidays – respectively. The quality of the polyno- weekend residuals is an indication that the relevant temperature and load
mial regression depends on the weather sensitivity of the underlying load; growth dependencies of the weekdays and weekends have been adequate-
loads dominated by residential and commercial customers will typically ly described by the regression. It also suggests the possibility that the fac-
yield a better fit to temperature than loads with a bigger proportion of in- tor residuals of the two series could perhaps be aggregated into a single
dustrial customers. time series, facilitating a simpler analysis. However, in what follows we
The logarithm of the squared residuals is analysed in a similar way. The continue the analysis separately for the two day-classes.
logarithm prevents the variance estimates from being negative. Figure 4 We now define the historical ‘amplitudes’ of each of the factors by tak-
shows the results for the hour ending 4pm on weekdays using a fourth- ing the inner product of the time series of the normalised variance with the
order regression polynomial.
The hourly historical growth rates of individual load classes present in-
Figure 5: PJM load index hourly growth rate, weekdays
teresting characteristics. The growth rates for the PJM load is shown in fig-
ure 5. Loads in the off-peak early morning hours have been growing the
fastest, at more than 2% a year. The other block of hours with rapid growth 2.5
is in the afternoon and early evening. The midday and late evening hours
exhibit the slowest growth rate, 1.7–1.8%/year.
2.3
Factor analysis of residuals
This critical phase of our analysis is what facilitates realistic simulation of 2.1
% / year

load dynamics. Our approach starts by viewing residuals of the previous

date/temperature regression analysis as a daily multivariate time series of di-
mension 243. We normalise these residuals by the variance before proceeding 1.9

1 While parameter calibration in neural-net models represents in some fashion a

relationship between the factors driving electricity demand – such as temperature or 1.7
humidity – and the demand itself, regression-type models exhibit this relationship in a
much more transparent fashion.
2 A polynomial regression will result in a parabolic-type fit, which exhibits explosive

load growth for very hot and very cold temperatures.

1.5
4 8 12 16 20 24
3 An alternative is to use an hourly time series representation, but this involves

substantial complexities when dealing with weekend and weekday effects as well as Hour
daily correlation – issues that turn out to be benign in our current approach.

[Link] I 85
Technical

eigenvectors: aj (n) = ς(n)·vn. This amounts simply to a change of basis from Example: forecasting BGE loads
the original 24-hour basis to that of the six factors that have been kept. We The load regression has been tested on Baltimore Gas & Electric loads. The
can now analyse each of the series aj (n) as an independent one-dimensional underlying temperature is a day-ahead forecast for Baltimore airport. Com-
time series. The assumption of independence is for convenience and is at paring the forecast to realised loads assesses the load forecast’s accuracy.
least partially justified by the fact that the PCA yields uncorrelated resid- The same benchmark is used to measure the performance of commercially
uals. The procedure used to analyse these time series resembles the way available similar-day and neural-net forecasts4.
temperature time series are analysed and requires at least a full year of load A sample forecast for August 2001 is shown in figure 8 – the green
data. This method is described in detail in the next section. In summary, curve shows the actual loads, the blue line is the neural-net’s day-ahead
the procedure searches for any seasonality in the mean and variance of the forecast, the red curve is Constellation Power Source’s (CPS’s) forecast5.
time series. The CPS forecast outperforms the neural-net forecast on August 23 and
The final representation of load dynamics used for forecasting and 25, while the neural-net is the better model on the 29. Both models over-
simulation is the sum of the hourly polynomial fit of the load to temper- estimate loads on the two weekend days (August 25 and 26) and on Au-
ature E[L(n)|T(n)] and the fit of the load standard deviation gust 28, which saw very high loads. Note that much of the discrepancy
E[V(n)| T(n)], the hourly trend δ(n) and the sum over the first six factors between forecast and actual loads can be attributed to inaccuracies in the
of the seasonal factor amplitudes: temperature forecast.
The historical forecast error was measured from April 2001 to August
2002. It was measured as the absolute value of the difference between
L(n) = E [L (n ) | T (n) ] +
6
δ ( n) + E [V (n) | T ( n) ]⋅ ∑ m j ( f n )⋅ ν j
4 The ANNSTLF (Khotanzad & Abaye (1997)) and Neural-Net and Similar-Day forecast
are Electric Power Research Institute (EPRI) products ([Link]).
j =1 5 The CPS forecast was calibrated over the January 1996 to July 2001 period.

Figure 6: PCA eigenvalues of the PJM load index, weekdays (left) and weekends and holidays (right)

15 15
Eigenvalue

10 10
Eigenvalue

5 5

0 5 10 15 20 25 0 5 10 15 20 25
Eigenvalue number Eigenvalue number

Figure 7:PJM load index PCA eigenvalues – first four eigenvectors, weekdays (left) and weekends and hols (right)

First Second Third Fourth First Second Third Fourth

0.4 0.4

0.2 0.2
Eigenvector
Eigenvector

0 0

–0.2 –0.2

–0.4 –0.4

0 5 10 15 20 25 0 5 10 15 20 25
Hour Hour

86 I July 2002
Technical

Figure 8: Day-ahead load forecast for BG&E Group Performance of short-term load forecasting models

EPRI ANNSTLF neural-net forecast CPS forecast Neural-net Similar-day CPS forecast
EPRI similar-day forecast Actual loads (
E L − L forc ) 156 MWh 183 MWh 158 MWh
6  
(L − L )
2
E  213 MWh 242 MWh 215 MWh
 forc

5.5

5 forecast and actual load and the standard deviation of that difference. The
MWh ('000s)

table confirms that the three models have very similar performances. The
4.5 advantage of the regression is the transparency of the calibration process
and the ease of implementation and maintenance when compared to neu-
4
ral-net forecasts.
3.5
It is also worth noting that, unlike the similar-day and neural-net ap-
proaches, no other weather variables other than temperature forecasts are
3 used in the regression approach. The regression approach easily lends it-
self to further analysis of the residuals, to include the effects of summer-
2.5 time humidity or wintertime wind chill on electricity demand. This will
Aug 23 Aug 25 Aug 27 Aug 29 only improve the model’s performance.

Simulating full-requirement load obligations

Full-requirement obligations involve serving load, which can fluctuate on
Figure 9: Seasonal fit to temperature at BWI an hourly basis. The origin and statistical attributes of these fluctuations can
vary significantly between load classes. For example, residential and com-
90 mercial loads exhibit dramatic seasonal dependencies as well as daily and
hourly fluctuations, much of which can be explained by weather. In contrast,
80 the power required by a single large industrial user is often insensitive to
temperature. In this section, we describe a systematic approach to the mod-
Temperature (˚F)

70 elling of the stochastic dynamics of load obligations with significant sea-

sonal and weather dependencies6.
60 This approach to analysing full-requirement obligations requires mod-
elling of both temperature and loads. The load simulation module is simi-
50
lar to the model described in the previous section. The main difference is in
the additional stochastic nature of temperature and load simulations.
40

Temperature
30
There is an abundance of literature on the development of realistic weather
simulations. But many approaches suffer from unintended and undesired
20
0 50 100 150 200 250 300 350 non-stationarity (Dornier (2000); Moreno (2000); Carmona (1999); and
Day in year Storch & Zwiers (1999)). Our approach is based on the following model:

T (n ) = δ (n ) + µ ( f n )+ σ (f n )X (n )

where T(n) denotes temperature on day n, fn is the fraction of the year cor-
Figure 10: Seasonal fit to temperature variance at BWI
responding to day n, δ denotes a drift in mean temperature, µ and σ are the
mean and standard deviation of temperature, and X(n) is a stationary unit-
150 variance autoregressive moving average (ARMA) process. By construction,
T(n) has mean µ and standard deviation σ. Also by construction, µ and σ
are periodic functions and are arguably reasonably smooth, which makes a
Fourier representation convenient (see, for example, Hamilton (1994)). The
Temperature (˚F)

100 estimation method of choice at CPS is:

● a preliminary estimation of Fourier coefficients for µ and σ, with the

number of modes truncated by an optimal stopping rule;
● a standard regression estimate for the drift after de-seasonalising using
50
the preliminary estimate for µ;
● a final Fourier estimation of µ and σ after removing any trend; and
● an ARMA analysis of the residuals.

0
0 50 100 150 200 250 300 350 6 We do not address attrition, namely the long-timescale migration of load due to
Day in year embedded optionality afforded to customers in many full-requirement contracts.
Attrition is similar to mortgage pre-payment and involves econometric analysis well
removed from the scope of this paper.

88 I July 2002
 Figure 11: Simulated average BGE load, Jan 1–15, ‘03 Figure 12: Simulated average BGE load, Jul 1–15, ‘03

3.8 4.5

3.6
4.0
3.4
MWh ('000s)

MWh ('000s)
3.2 3.5

3.0
3.0
2.8
2.5
2.6

2.4 2.0
Jan 1 Jan 3 Jan 5 Jan 7 Jan 9 Jan 11 Jan 13 Jan 15 Jul 1 Jul 3 Jul 5 Jul 7 Jul 9 Jul 11 Jul 13 Jul 15

Figures 9 and 10 show the results of step 3 above for Baltimore Airport. ● the factor analysis also captures the seasonality in the temperature-load
The Fourrier coefficients for both figures were calibrated over the relationship; and
1960–2001 period. ● the hourly trend analysis captures historical changes in the growth rate
of load in peak and off-peak hours.
Example: long-term load simulations
The statistical parameters describing temperature are used to simulate a A later article will describe the simulation of clearing prices for organ-
number of realisations of a temperature time series at a given location. These ised power pools. The combination of load and price simulations will be
simulated temperatures are used to drive the loads used to value full-re- used to calculate the earnings distribution associated with serving load ob-
quirement obligations. The load simulation adds a stochastic component to ligations. The effectiveness of weather and load hedges in reducing earn-
the process used for short-term load forecasting (X(n) is a standard ARMA ings volatility will be the main focus of the article. EPRM
process and σj (n) is the seasonal fit to the variance of the jth factor).
Véronique Bugnion is an associate in the strategies group at
L(n) = E [L (n ) | T (n) ]+ Constellation Power Source (CPS), an independent power producer
6 and marketer based in Baltimore, Maryland
δ ( n) + E [V (n) | T ( n) ]⋅ ∑ ( m j ( f n )+ σ j ( fn ) X ( n)) ⋅ ν j e-mail: [Link]@[Link]
j =1 Aram Sogomonian is manager for business development at CPS.
e-mail: [Link]@[Link]
Figures 11 and 12 show the average of 500 load simulations for Balti- Glen Swindle is vice-president and head of strategies at CPS.
more Gas & Electric for the summer and winter of 2003. July loads exhib- e-mail: [Link]@[Link]
it a distinct single peak, and July 4 is correctly simulated as a ‘Sunday’ load
shape. Winter simulations have a structure with both a morning and an af- The authors would like to thank Michael Booth and Martin Lawera for
ternoon peak. many interesting and useful discussions on load forecasting methods
The relevance of the stochastic model above is not merely to generate and related topics.
expected future load, but also to provide a vehicle for pricing and assessing
the hedging effectiveness of load derivatives. The approach to pricing is a Note: please see page 8 for guidelines for submitting technical articles.
straightforward simulation of the process above. While this does not account
for market price of risk, it at least provide an effective methodology for ac- Bibliography
tuarial valuation of load products. Finally, a significant benefit from the re- Carmona, R, Calibrating degree-day options, Talk given at the third seminar on
duction in dimensionality afforded by the factor representation of load stochastic analysis, Random Fields and Applications, Ascona, Switzerland,
fluctuations is that we can easily perform joint simulation of multiple loads September 23, 1999.
in the same region.
Dornier, F and Queruel, M, Caution to the wind, Weather Risk, August 2000.
Conclusion
This article presents an integrated approach to modelling temperature and Hamilton, J, Time Series Analysis, Princeton University Press, 1994.
electrical loads. This approach has proven successful in forecasting short-
term loads and in simulating future loads for the purposes of establishing Khotanzad, A and Abaye, A, ANNSTLF – a neural networks-based electric load
properties of full requirements obligations. This method has a number of in- forecasting system, IEEE Transactions on Neural Networks, 8, 843–845, 1997.
teresting features:
Moreno, M, Riding the temp, Futures and Options World, December 2000.
● both temperature and load regressions are transparent and easily un-
derstood; Storch, H and Zwiers, F, Statistical Analysis in Climate Research, Press Syndicate
● the factor analysis allows loads to maintain a coherent daily structure; of the University of Cambridge, 1999.

[Link] I 89