The existing power system utilizes communication infrastructure for fast and reliable transfer of control and protection inputs. This dependency of a power system on communication infrastructure for critical applications makes it vulnerable to cyber attacks. Cyber-induced outages trigger both randomized and intentional switchings in a power system producing relatively similar dynamics as natural events making their classification difficult. This article proposes a principal component analysis (PCA) assisted sequential deep learning approach for online classification of cyber outages and natural events in a power system. This article provides objective-driven models for false setting injection (FSI) and false command injection (FCI) type attacks. The proposed classification method uses PCA to deduce truncated z-score sequences [or principal test sequences (PTSs)] capturing distinct spatio-temporal progression patterns of natural disturbances and cyber outages. The PTSs in the training sets are shuffled and sampled using a stratified random sampling technique and classified using an ensemble long short-term memory network. The proposed method is tested for simulation examples of FSI and FCI attacks in the standard IEEE118-bus test system, where it showed improved accuracy and time performance.