Enhanced Power System Security Assessment Through Intelligent Decision Trees - Complex System Modelling and Control Through Intelligent Soft Computations - page 355

Fig. 14 Economical bene t
of operational rules from
ef cient sampling

Table 7 Comparison
between different sampling
strategies

Sampling
strategy

Size Accuracy False

alarm

Risk

IS (MVN — pruned) 2,879 80.6142 0.142 0.228

IS (MVD — pruned) 2,852 87.0951 0.094 0.178

IS (MVD) 2,852 97.5202 0.021 0.03

It can be seen from Table 7 that the database produced by importance sampling
of correlated-MVD state space de nitely shows better performance. When the trees

are pruned for operator s convenience of usage the accuracy decreases, which can

be improved using the right-hand side loop as shown in the Fig. 3 . It also performs
better than sampling from MVN load space, which is conventional assumption in
many studies due to trivial modeling requirements.
The signi

'

cance of sampling from correlated-MVD, i.e., capturing the inter-load
dependencies, than from MVN is even strongly vindicated by Fig. 15 that shows the
top 5 critical attribute locations produced by decision trees from respective dat-
abases. The contingency event is shown by a red star. The location of 5 critical
monitoring attributes as well as their sequence in the tree matters. Compared to
MVN, all the 5 top locations found by correlated-MVD sampling strategy are very
interesting ones, with the top node being reactive reserve at a big nuclear plant
Chinon, the node in the next level of the tree is closer to the contingency location,
the next nodes (3 and 4) in the tree deals with the two SVC locations in Brittany,
and

finally the attribute of node 5 is right at the contingency location.