Enhanced Power System Security Assessment Through Intelligent Decision Trees - Complex System Modelling and Control Through Intelligent Soft Computations - page 336

large computational cost. This work proposes a computationally ef cient method

based on Latin Hypercube sampling (LHS) to characterize the operational param-
eter space.
The next section introduces the concept of

“ high information content ” and the

measure that can be used to quantify it.

3 High Information Content Region

The decision tree learning algorithm requires a database that has good represen-
tation of all the class values, so that it can effectively classify new instances and not
overlook the less representative classes. So, for a two-class problem, a good rep-
resentation of operating conditions on both sides of the class boundary is required.
Also, not every operating condition on both sides of the class boundary contributes
equally to the operating rule derivation process. This is further demonstrated using
Fig. 2 with the help of its four parts a-d, which explain the importance of sampling
the most

uential operating conditions for the purpose of rule making. For
instance, consider sampling some operating conditions de

in fl

ned in terms of varia-
tions in Loads A and B as shown in Fig. 2 a. Perform contingency analysis to nd

the post-contingency voltage stability performance (yellow dots have acceptable,
and red dots have unacceptable performances). A suitable rule can be de

ned by
line R that effectively partitions the operating region with acceptable post contin-
gency performance from unacceptable performance. We refer to this line as the
security boundary. Now, if more operating conditions are sampled as shown in
Fig. 2 b, the samples drawn near to the security boundary in

uences the rule making
process more than the samples away from the boundary. This is evident from the
consequent rule change (shifting line R) that is necessary as shown in Fig. 2 c. So it
is very essential

fl

that the database contains operating conditions nearer to the

security boundary with finer granularity, since they convey more information on the

variability of the performance measure, which thereby enables a clear cut decision
making on the acceptability of any operating condition. Furthermore, if the some of
the operating conditions with unacceptable performance near the rule line R in
Fig. 2 c are less likely to occur in reality, then the rule line R may be shifted slightly
upwards to exploit more operating conditions for economic reasons, as shown in
Fig. 2 d. Hence the desired in

uential operating conditions are obtained by sampling
according to the probability distribution of the boundary region, which is the shaded
region in Fig. 2 d where there is a high uncertainty in the acceptability of any
operating condition. This will also ensure a very good representation of both the
classes in the database at a reduced computational cost compared to sampling from
the entire operational parameter state space probability distribution.
In this work Entropy, the most commonly used information theoretic measure
for the information contained in a distribution, is used to quantify information
content in a database (Unger et al. 1990 ). It is a function of class proportions, when

fl