Enhanced Power System Security Assessment Through Intelligent Decision Trees - Complex System Modelling and Control Through Intelligent Soft Computations - page 353

“ p ” cally, two aspects are discussed, namely (a) computational require-

ments and accuracy, and (b) economic bene

. Speci

ts.
(a) Computation, Accuracy and Tree Size: Fig. 13 a

d show the total SEO load
probability distribution from sampled operating conditions as the sliding factor
p increases from the base value in f(x) to 1 (bias towards boundary).
Table 5 shows the results when validated using the test database, which con

-

rms
that as the sampling of operating conditions is biased towards the boundary region,
the entropy of the database increases (a quantitative indicator of information con-
tent) and even with lesser database size higher accuracy for decision tree is
obtained. The error rates, namely false alarms and risks are both simultaneously
reduced to a great degree. It was also found that as the sampling is biased more
towards the boundary region, the size of the decision tree required for good clas-
si

cation also decreased. This is due to the ability of database to capture high
information content (i.e., the variability of performance measure across the security
boundary) even with smaller number of instances.
(b) Economically bene cient

sampling in producing economical rules. The table shows that for the various
possibilities of the decision tree top node attribute among the most in

cial rules: Table 6 presents the in fl uence of ef

uential
attributes, the database generated from within boundary region with p =1

fl

nds
rules with attribute thresholds that are always less conservative than from the
database that was generated with p = 0, i.e., from entire operational state space.

Fig. 13 Effect of p on sampled total SEO load probability distribution. a p = 0.25. b p = 0.50.
c p = 0.75. d p = 1.0