Enhanced Power System Security Assessment Through Intelligent Decision Trees - Complex System Modelling and Control Through Intelligent Soft Computations - page 334

cation accuracy and error rates are used as the performance measures of a
decision tree. There are two kinds of errors: false alarms —

Classi

acceptable cases clas-

si unacceptable cases as acceptable. Errors can be

calculated by testing the obtained decision model on the training set, which is
usually an over-estimate. There are training set sampling methods such as holdout
procedures, cross-validation, bootstrap etc. (Witten and Frank 2000 ) to make the
error estimation unbiased. It is even better if the testing is performed using an
independent test dataset. There are numerous references that explain the process of
building a decision tree from a database with algorithms such as ID3, J48 etc.
CART, Answer Tree, Orange, WEKA etc. are some software available for building
decision trees.
Many utilities have taken and are continuing to take a serious interest in
implementing learning algorithm such as decision tree in their decision making
environment. French transmission operator RTE has been using decision tree based
security assessment methods to de

ed as unacceptable; and risks —

ne operational security rules, especially
regarding voltage collapse prevention (Lebrevelec et al. 1998 , 1999 ; Schlumberger
et al. 1999 , 2002 ; Pierre et al. 1999 ; Martigne et al. 2001 ; Paul and Bell 2004 ;
Henry et al. 1999 , 2004a , 2006 ; Cholley et al. 1998 ). They provide operators a
better knowledge of the distance from instability for a post-contingency scenario in
terms of pre-contingency conditions, and thus save a great amount of money by
preserving the reliability while enabling more informed operational control closer to
the stability limits. So the central topic of this chapter will be: what is the significant
component of this decision tree induction process, and how to improve it for the
betterment of the planning solutions that are needed under realistic operating
conditions?
The remaining parts of this chapter are organized as follows. Section 2 provides
the background of this work in terms of motivation behind this research, related
past work, and the objective of this work. Section 3 describes the concept of
“ in the context of this work. Section 4 presents the technical

approach of the proposed high information contained training database generation.
Section 5 demonstrates the application in deriving operational rules for voltage
stability problem in Brittany region of RTE

information content ”

s system, and presents results and
discussions. Section 6 presents conclusions and future research directions.

'

2 Motivation, Related Work, and Objective

The most vital and sensitive part of MCS based reliability studies is the stage of
database generation. The con dence we will have in the results generally re fl ects

the con dence we have in the set of system states generated. The generated data-

base does in cation performance of the decision tree against

realistic scenarios, selection of critical attributes and their threshold values, and size
of the operating rules.

fl uence the classi