Hybrid Load Forecasting Method With Analysis of Temperature Sensitivities Kyung-Bin Song, Seong-Kwan Ha, Jung-Wook Park, Dong-Jin Kweon, Kyu-Ho Kim IEEE transactions on POWER SYSTEMS, May 2006
Feb 23, 2016
Hybrid Load Forecasting Method With Analysis of Temperature Sensitivities
Kyung-Bin Song,Seong-Kwan Ha,Jung-Wook Park,Dong-Jin Kweon,
Kyu-Ho Kim
IEEE transactions on POWER SYSTEMS, May 2006
Outline Introduction Hybrid Short-Term Load Forecasting (STLF) Methods
Exponential Smoothing Fuzzy Linear Regression Temperature Sensitivities
Results Conclusion
Introduction Conventional methods on STLF are not suitable
for higher load forecasting errors in some particular time zones, like weekends. So intelligence techniques like fuzzy and neural network are recently as an alternative in forecasting.
In this paper, a new hybrid load forecasting consist of fuzzy linear regression and the general exponential smoothing with analysis of temperature sensitivities is proposed. Each has their usage on different situation.
Hybrid Short-Term Load Forecasting General patterns of weekdays are almost
identical, but on weekends and Mondays are depend on consumers lifestyle. Especially, the load patterns of a summer season are strongly affected by the temperature.
In this paper, Mondays and weekends are processed together for better accuracy. And hybrid STLF with temperature is used in summer.
Flow chart
Methods – Exponential Smoothing General exponential smoothing is developed to
provide load forecasting of weekdays without regarding to weather conditions and special vacations. In other words, these methods only use in spring, fall and winter.
A smoothing time series model :
F is the forecasted value, X is real value, N is the number of observation data
Methods – Exponential Smoothing Assume the observation data varies slowly, using time
series model :
substitute with
for general :
Using maximum load during three days prior to the
predicted day, final equation :
Methods – Exponential Smoothing example -
Predicted day : March 26(Tue)Three observation days maximum load : March 25(Mon), 22(Fri), 21(Thu) : 1/3
the equation of forecasted load:
Normalized The average normalized value for 24 hourly loads
=
, are the maximum and hourly load on previous same day as the predicted day.
The hourly load on the predicted day :
This operation attached to all the methods
Method - Fuzzy linear regression
Linear regression is a statistical method to model the relationship between two variables by fitting a linear equation to observed data
Fuzzy linear regression model can expressed as :
All variables are fuzzy numbers, and are fuzzy addition and multiplication
Triangular fuzzy number
Method - Fuzzy linear regression
Fuzzy number addition
multiplication
For simplicity, assume all the variables are symmetric fuzzy numbers, then , , , , and the result of fuzzy linear regression analysis using shape preserving operation :
Method – temperature sensitivities
Method – temperature sensitivities Step 1 : Select 3 years highest temperature/load
correlation coefficient input data among past 5 years of predicted year.
Step 2 : Divide predicted period into two parts, July and August, which have different temperature sensitivities.
Step 3 : calculate variation of temperature between the predicted day and one previous day
and are maximum temperature of predicted day and previous day
Method – temperature sensitivities Step 4 : calculate the variation of load
and are the maximum loads
Step 5 : compute the temperature sensitivities using and
Method – temperature sensitivities Step 6 : check the signs of in July and August, and
compute the average of each positive and negative value
Step 7 : forecast the maximum load on predicted day
,
is the forecasted maximum load in the one day before predicted day
Results The algorithm is applied to forecast a
week(different seasons) of 1996
Comparative algorithm is the “top-down developed method”, called Load Forecasting Engineering System(LOFES), by Korea Electric Power Research Institute(KEPRI), which use time-domain regressive analysis with past load profile and weather data.
Results March 26 to April 1(spring)
Results October 29 to November 4(fall)
Results July 9 to July 15
Conclusion This paper proposed a new hybrid short-term load
forecasting algorithm. It combines exponential smoothing, fuzzy linear regression and temperature sensitivities.
The temperature sensitivity is a new approach and it actually improves the load forecast accuracy in summer