The increased availability of large scale computing hardware brings the analysis of fully three-dimensional non-autonomous flow cases within reach. In these flow scenarios, the simplifying assumption of temporal homogeneity is not applicable and with it many data-driven analysis techniques that rely on it. Within the well-established modal decomposition framework of Proper Orthogonal Decomposition (POD), we can treat time in the same way as the spatial dimensions and apply the method to statistical ensembles of non-autonomous flows in order to extract coherent structures in space and time from the resulting experimental or numerical data, leading to the space-time POD formulation. This extension of the existing method is demonstrated on the model problem of the complex Ginzburg--Landau equation, modified to include non-autonomous parameter variations. Subsequently, the space-time POD analysis is carried out on a numerical dataset of 25 realisations of the onset of leading edge dynamic stall on a NACA0009 airfoil section subject to low levels of background disturbances. The space-time POD, combined with extended POD, is used to extract the spatio-temporal structure of energetic wavetrains during the bursting of the laminar separation bubble close to the leading edge, which are found to be statistically relevant phenomena in the context of incipient dynamic stall. The potential of the space-time POD methodology to objectively extract coherent structures from ensembles of non-autonomous data is demonstrated.