Shortterm Load Forecasting based on Neural network and Local Regression Jie Baoa Ellen Maxonb Vasant

1
Short-term Load Forecasting based on Neural
network and Local RegressionJie Baoa Ellen
Maxonb Vasant Honavarc acArtificial
Intelligence Lab, Dept of Computer Science Iowa
State University, Ames, IA, 50010
baojie_at_cs.iastate.edubPower Domain, Inc 7575
Palos Verdes Reno, Nevada 89502,
ceo_at_powerdomain.com
Abstract Accurate short term load forecasts
(STLF) are of significant financial implication
for the utility industry as well as the
consumers. We used a multi-stage approach to
solving the STLF problem. Initially, when little
is known about for the exact form of the load
function f, a neural network is used to uncover
predictive relationships between input and
output. The resulting neural network is examined
to identify factors that are most meaningful in
STLF. A linear model is constructed using the
identified factors leading to a Local Regression
model (with Moving Average as a special case). We
then explore a Neural Network (NN) model that
refines the linear regression model further, with
the result from Local Regression (LR) as guidance
for the training of the neural network. Best
day-ahead result from a combination model of
LRNN is 2.70 error The results demonstrate the
usefulness of a multi-step iterative refinement
approach to model building in solving complex
problems such as STLF .

• 1 Whats Load Forecasting?
• Tell the Future Given history record,
  forecast future electrical load
• An central problem in the operation and
  planning of electrical power generation.
• To minimize the operating cost, electric
  supplier will use forecasted load to control the
  number of running generator unit.
• Short-term load forecasting (STLF) is for hour
  to hour forecasting and important to daily
  maintaining of power plant

2 What may be the influential factors in STLF
Day1
Day4
Week1
Week5
Summer
winter
winter

• Calendar
• Seasonal variation
• Daily variation
• Weekly Cyclic
• Holidays
• Weather Temperature Cloud cover or sunshine
  Humidity
• Economical or environmental
• Unforeseeable random events

Winter
3 How to do the forecasting ?
Summer
Find the function f - it is complex and unknown,
but really existent and consistent Lf(t)
f(L(t),L(t-1),L(t-k),c,w,e)
Overlapped influence Of calendar and weather
Temperature and Humidity
Load History LCalendar cWeather
wOthers e
Forecasted Load Lf
f
Temperature shifting
?
Linear regression of weeklyaverage temperature
and load
Normalized average load and temperature (dashed
line) for every week
4 Black box model Neural Network - with
little expertise knowledge is known
5 White box model 1 Moving Average -
with knowing something
Its a common and naïve way as treating the
problem as a time series forecasting problem, and
function approximation.Ensemble network is used
to reduce model bias

• Leant from the NN model (by trial-and-error and
  apply/hide method )
• Overlapped calendar circle - daily , weekly ,
  yearly
• Temperature is the most influential weather
  factor and almost linear!
• Forecasted temperature helps
• Error pattern kept in adjacent

Long-range influence Li is load at past
reference point, eg. a week ago , wi is its
weight - all calendar information is
well kept. Short-range influence Temperature
shifting and error feed back is
applied Day-ahead model Best result 3.54
error
INPUTs hour , real/forecasted temperature ,
load history, weekend, weekday OUPUT load of the
future Day-ahead model Best result 5.45
error Hour-ahead model Best Result 1.53
Advantages Little knowledge is required
General algorithm ready Weakness
Time-consuming learning Not Stable, Cant
online update Season sensitive

Advantages Simple, Stable, Quick Neednt
Learning Weakness Temperature coefficient is
not constant
7 Hybrid model Neural Networkplus Local
Regression - there still factors we dont
clear
6 White box model 2 Local Regression
- with load pattern is somehow clear

• LR gives a roughly estimation and then NN refines
  it
• Input the result of LR into NN, along with other
  information
• result of local regression
• load and temperature of the past
• forecasted temperature of the future
• Other information

Comparison NN 5.45 MA 3.54 LR
3.22 NNLR 2.97 NNLR no errFeedBack
2.70
Can we use local information other than global
regression in MA? Day-ahead 3.22 2-hour
ahead 2.14
Lr(n) and k(n) are the constant and linear order
coefficients for given time n
Advantages Simple but works well Can online
updating No learning required Uniform model
for all seasons Weakness Only use week ahead
info
More on http//www.public.iastate.edu/baojie/pub/
2002-10-25_stlf.ppt.pdf