Clustering of data into multiple models for real time applications is an interesting and important problem. This is usually referred to as the multiple model learning (MML) problem. This MML problem has been investigated for several engineering systems. Various non-parametric algorithms have been proposed in literature to identify multiple models using support vector classifier, k-nearest neighbors and so on. Prediction error (PE) based approaches that perform clustering in the space of parameters instead of the actual data have also been proposed. In this article, we propose an integration of PE based clustering with statistical testing for a more robust solution to the MML problem. The efficacy of the proposed approach is demonstrated using several case studies of static and dynamic data. The ability to identify redundant variables using the proposed approach is tested by identifying the variables that affect energy utilization in residential buildings. The utility of this approach in identification of piece-wise dynamic models that can be used in control applications is also demonstrated by identifying non-linear dynamics of a non-isothermal continuous stirred tank reactor (CSTR). The performance of the proposed approach in comparison to other approaches in the literature is also studied. © 2018 Elsevier Ltd