LOAD FORECASTING
LOAD FORECASTING
ØCLASSIFICATION OF LOAD.
ØCHARACTERISTIC OF LOAD.
ØAPPROACHES TO LOAD FORECASTING.
ØFORECASTING METHODOLOGY.
1.EXTRAPOLATION .
2.CORRELATION.
REQUIREMENTS FOR FORECAST
ØCATEGORIZATION OF CONSUMPTION SECTORS.
ØCOLLECTION & COMPILATION OF PAST DATA.
ØDECOMPOSITION OF DATA.
ØVALIDATION OF PAST DATA & ANALYSIS.
ØSELECTION OF SUITABLE FORECAST
TECHNIQUE.
ØCOLLECTION OF DATA OF ECONOMIC VARIABLES
FOR ECONOMETRIC MODELLING.
KEY ISSUES
ØENERGY SHORTAGE DUE TO RESTRICTIONS &
UNSCHEDULED CUTS.
ØENERGY EFFICIENCY & CONSERVATION
MEASURES.
ØSPECIAL IMPETUS ON RURAL ELECTRIFICATION &
IRRIGATION.
ØTHRUST TO ECONOMICALLY WEAKER STATES &
REGIONS
Ø
FACTORS
INFLUENCING ELECTRICITY DEMAND
ØDEMOGRAPHIC GROWTH.
ØGEOGRAPHIC VARIATIONS.
ØMETEOROLOGICAL VARIATIONS.
ØNATIONAL POLICIES FOR SECTORAL GROWTH
§NATIONAL ELECTRICITY POLICY
§INFRASTRUCTURE DEVELOPMENT PLANS.
ØECONOMIC GROWTH.
ØDIVERSITY FACTORS-INTER STATE, INTER
REGIONAL.
ØINTERNATIONAL ECONOMY LIKE FUEL /OIL PRICE,
ETC.
ELECTRICITY DEMAND FORECASTING:
A NECESSITY
a)ELECTRICITY SECTOR PLANNING: NATIONAL LEVEL
STATE LEVEL
•GENERATION
•TRANSMISSION
•DISTRIBUTION
b)DEVELOPMENT PLANS OF ‘ELECTRICITY-
CONSUMPTION SECTORS’
- STAKE- HOLDERS
- INFRASTRUCTURE SECTORS
c)REVIEW OF ELECTRICITY SUPPLY POSITION:
NATIONAL AND STATE LEVEL
FORECASTING
METHODOLOGY
Short term forecasting
Medium Term forecasting
Long term forecasting
FORECASTING
TECHNIQUES
Mainly Classified in
Extrapolation
Correlation
Combination of both
Further Classified in
Deterministic
Probabilistic
Stochastic
EXTRAPOLATION
Extrapolation techniques involve fitting trend
curves to basic historical data adjusted to
reflect the growth trend .
Forecast obtain by trend curve
Deterministic extrapolation
Random errors not accounted
Standard analytical functions
1.Straight line: Y=a+bx
2.Parabola: Y=a+bx+cx2
3.S curve: Y=a+bx+cx2+dx3
4.Exponantial: Y=cedx
5.Gompertz: Y=In-1 (a+cedx )
Probabilistic Extrapolation
Uncertainty of extrapolated results is quantified using
statistical mean & variance.
Uncertainty arises in historical data & analytical
model.
Stochatic Extrapolation Model
Generate forecast from random inputs derived from
historical data investigated in 1975
Not wide spread
Based on
- Random change in the trend component, -
Random slope in seasonal component, -
Associated weighting factor,
- A general noise component.
Perform by transformation of basic model
Because basic model & data nonstationary
Obtain by exciting the transformed model by random
inputs with known statistics
CORRELATION
Relate system loads to various
demographic and economic factor
Advantageous for understanding
interrelationship between load growth &
other measurable factors
Disadvantage- results from the need to
forecast demographic & economic
factors, which can be more difficult than
forecasting system load.
Typically used factors
Population employment
building permits appliance saturation
business indicators weather data
No one forecasting method, it must be
emphasized is effective in all
situations.
Curve fitting method is adequate for
some utilities & completely worthless
for others.
Forecasting – As tools to aid the planner
Gives good judgment & experience.
Never be completely replaced.
ENERGY FORECASTING
Energy forecast classified in
○ Residential
○ Commercial
○ Industrial
Each class separately forecast because
of different characteristics
Residential Sales
Forecasts
Population Method
Residential energy requirements are dependent on
for
1. Residential Consumers week
2. Population per consumer month
3. Per capita energy consumption etc.
Forecast could be obtain by multiply three
factors
Simple curve fitting or Regression analysis is
used
Synthetic method
- Requires detail look at each consumer
- Major factors
1. Saturation level of major appliances
2. Average energy consumption per appliance
3. Residential consumers
- These factors extrapolated and multiplied.
- Projection phase involves data gather from
customers, Builders.
- Rapid growth is an extremely important step
Commercial Sales
Forecast
Commercial establishments are usually
service oriented
Hence growth patterns are close to
residential sales
By using method of extrapolation ratios of
commercial to residential sales in to
the future & then multiply by residential
sales forecast
Another approach to extrapolate
historical commercial sales.
Industrial sales forecast
Industrial sales are tied very closely to the overall economy
& overall economy not stable over selected periods.
Approaches in view
1. Multiply forecasted production levels by forecasted
energy consumption per unit of production
2. Multiply forecasted number of industrial workers by
forecasted energy consumption per worker.
Approaches depends upon type &location of industry
For Accurate forecasting – Historical data is decompose
into subclasses for broad spectrum of industries
Final step is tuning the forecast
PEAK DEMAND
FORECASTING
Extrapolating historical demand data over the last 5 years for accurate
peak demand forecast.
Can get weekly, monthly, yearly peak demand forecast by changing
data sampling rate
The basic approach
Ø Determine seasonal weather load model
Ø Separate weather sensitive or non weather sensitive component of
demand
Ø Forecast mean, variance of NWS component of demand
Ø Extrapolate weather load model & forecast mean, variance of NWS
component
Ø Determine mean, variance, and density function of total weekly forecast
Ø Calculate density function of monthly or annual forecast.
o Seasonal variation in peak demand due to weather conditions
o For elimination of seasonal variation method proposed by Shiskin,
database of at least 12 yrs is required
Weather Load Model
W= Ks(T-Ts) if T > Ts Ts, Tw-Threshold Temp.
W= -Kw(T-Tw) if T<Tw Ks, Kw- slopes
=0 if Tw < T < Ts
Ø Drawing of scatter diagram is Req.- Daily peaks verses
weather variables
e. g. Dry bulb Temp., Humidity etc
linier reggration is used
Three Bus Layout
Separating weather sensitive &
weather non-sensitive
Components
NON WEATHER SENSITIVE
FORECAST (NWSF)
Discounted multiple regression(DMR) to fit a time
polynomial to the NWS component of historical weekly
peak demand
NWS weekly peak demands described by deterministic
time polynomial + random variable
Variations described by
ξ(t)=a1+a2 = a1f1(t) +a2f2(t)
f1(t) =1.0 Fitting functions
f2(t)= t
a1, a2 – DMR Coefficients
From a1, a2 we can calculate NWS component of peak
demand at some future week.
Discounted Multiple
Regression
Historical weekly peak demand data as
l(t)= ξ(t)+η(t) where, a’=1.n coefficient =a’f(t)
+η(t) f(t)=n.1 fitting function
vector
η(t)=random variable
l(t)&ξ(t) - variations
Assume, η(t)=0
ε(η(t))=ň(t)=0 property 1
Indicates sum of η(t) over all time, average=0
Addition of gaussian noise component
f(η(t))=(1/ση√2pi)exp(-1/2*(η(t)/ ση)2) Property 2
It says noise component is Gaussian distributed about
its mean in bell shape.
Probability density range:- η(t) to dη(t)
e.g. F(∞)=-∞ ∫∞ f(η(t))dη(t)=1.0
Based on property 1 average value 0
ε(η(t) η(t+Ɣ))=0 Ɣ≠0 property 3
= ση Ɣ=0
σ 2 = Variance large σ 2indicates noise vary about its
η η
mean
σ 2 =0 means η=ň
η
Σ defines 99% confidence interval
η
Random weekly peak historical data gives values of
coefficients on the fitting functions.
That is done by minimizes squared difference
between weekly peak demand l (t) & true peak
demand ξ(t)
n
function J= ∑1/2W2(t)[l(t)-ξ(t)]2
t=1
W(t) = √ᵦN-t where, t is discrete variable
ᵦ is discount factor
ξ(t)=(a1 a2) f1(t) matrix form
f2(t)
l(t)= l(t) W(t)=W(t)
J=1/2W2(1)[l(1)-a’f(1)]2+1/2W2(2)[l(2)-a’f(2)]2 +…
J=1/2[L(N)-a’R(N)]W(N)W’(N)[L(N)-a’R(N)]’
Where, L(N)=1*N Vector whose elements are l(t)
a’= 1*n Fitting function coefficient vector
R(N)=n*N fitting function matrix whose columns
are f(t)
By taking partial derivative
â= [L(N)W(N)W’(N)R’(N)][R(N)W(N)W’(N)R’(N)]-1
Letting, F(N)=[L(N)W(N)W’(N)R’(N)]
G(N)=)[R(N)W(N)W’(N)R’(N)]
â= F(N) G(N)-1
Recursive formula F(N)
Fik =∑N fi(t)ᵦN-t fk(t)
t=0
= fi(N)fk(N)+ ∑N-1 fi(t)ᵦN-1-t fk(t)
But ∑N-1 fi(t)ᵦN-1-t fk(t) = Fik (N-1)
t=1
Thus recursive relationship for the ikth element of F
Fik = fi(1) fk(1)
Initial values of Fik for fitting function describing 3 bus
system
F11 = f12(1) =0.1
F12 = f1(1) f2(1)= 0.1
F21 = f2(1) f1(1)=0.1
F22 = f22(1) =0.1
Recursive formula G(N)
Gn =∑N l(t)ᵦN-t fn(t)
t=1
= l n(N)fn(N)+ ᵦ ∑N-1 l (t)ᵦN-1-t fn(t)
t=1
= l n(N)fn(N)+ ᵦ Gn(N-1)
ξ(t) = â(n)f(t)………equation for NWS peak demand
Mean & Variance
Coefficients
The Variance of a (N) qualifies the range over
1
which the elements of a(N) varies as data are
simultaneously processed.
Determination of coefficient variance
con(â(N))= ε([(â(N)-ẫ(N)] [â(N)-ẫ(N)]’)
ẫ(N) mean value of (â(N)
â(N) = [εL(N)] [L(N)W(N)W’(N)R’(N)F(N)-1 ]
since [εŋ(N)] =0
ẫ(N) = F(N)-1 R(N)W(N)W’(N)ξ’(N)
con(â(N))= ε(â(N)â’(N)-ẫ(N)â’(N)+ẫ(N)ẫ’(N))
= ση2 F(N)-1 R(N)W(N)W’(N)W(N)W’(N)R(N)F(N)-1
ση2 = (N-n)-1 ∑nt-1 r2(t)
ξ(t) used for finding NWS component when
t=104
FORECAST VARIANCE
variance qualifies uncertainty of forecast
ε(ξ(t))= â’(N)f(t)
var(ξ(t))= ε[ξ(t)ξ’(t)][ξ(t) ξ’(t)]
= f’(t)con[â’(N)f(t)]
= σξ2
