Estimating electromechanical oscillations of power systems plays a crucial role to infer about their stability. In this paper it is proposed a new method in which dominant dynamics of power systems are inferred by analyzing pole estimates produced by several regularized robust recursive least squares (R3LS) implementations operating individually — each one with a different set of autoregressive moving average exogenous (ARMAX) model orders. The proposed method is named “Clustered R3LS”, since it properly adapts the so-called K-means clustering algorithm to automatically model regions of pole agglomeration produced by these individual R3LS implementations. An additional strategy for ignoring (discarding) poles identified as spurious is also part of the proposed Clustered R3LS algorithm. From a practical point of view, Clustered R3LS is here shown to more effectively track dominant dynamics of power systems when compared to Classical R3LS, which is based on analyzing only pole estimates produced by a single R3LS implementation operating with a fixed set of ARMAX model orders that is usually hard to tune. Three case studies are here used to validate the proposed Clustered R3LS method for both ambient and ringdown data. The first and second case studies are based on two well known benchmark models proposed for the analysis of oscillatory dynamics in power systems, whereas the third case study considers actual power system data extracted during a large-scale event from the Brazilian Interconnected Power (BIP) system.