Notwithstanding developments in design, processing and quality control techniques, flaws in insulation of electrical power apparatus such as blow-holes, cavities, surface imperfections, etc are unavoidable. Manufacturers and utilities employ a variety of prognostic tools to assess the condition of insulation amongst which Partial Discharge (PD) measurement has become an indispensable non-intrusive technique. Since discrimination of PD is a fundamental yet an indispensable prerequisite for diagnosis and since multiple PD sources are frequently encountered in practice, such studies have grabbed the attention of researchers. In view of the fact that PD is inherently a complex non-Markovian process and a majority of modern PD measurement systems acquire data for a substantial duration, the database is large leading to a plethora of complications during classification such as ill-conditioning, over-fitting etc. Though an array of intelligent and scientific computational techniques such as Neural Networks (NN), Hidden Markov Models, Wavelet Transformation etc have been utilized for single and overlapped PD sources with a fair degree of success, the inability to develop a comprehensive and validated scheme for identification continues to be elusive. In this research, exhaustive studies have been carried out on benchmark models that replicate single and multiple PD sources with a focus on three major aspects during classification which utilizes PNN versions that employ a variety of partition based clustering techniques namely algorithms based on distance and similarity, square error and density function. The first aspect focuses on ascertaining the ability of non-clustering based PNN versions in handling large ill-conditioned datasets during training in comparison with the various partitioning based algorithms. The second is on the analysis of the ability of square-error, distance and density based clustering techniques in providing frugal sets of representative centers during training. The third facet is to ascertain the role played by the preprocessing techniques in dealing with issues related to the curse of dimensionality during classification. Further, exhaustive analysis is carried out to determine the role played by the free parameter (variance parameter) in distinguishing various classes of PD, number of iterations and its impact on computational cost during the training phase in NNs which utilize the clustering algorithms and the choice of the number of codebook vectors in classifying the patterns. © 2005 - 2012 JATIT & LLS.