To support N-1 pre-fault transient stability assessment, this paper introduces a new data collection method in a data-driven algorithm incorporating the knowledge of power system dynamics. The domain knowledge on how the disturbance effect will propagate from the fault location to the rest of the network is leveraged to recognise the dominant conditions that determine the stability of a system. Accordingly, we introduce a new concept called Fault-Affected Area, which provides crucial information regarding the unstable region of operation. This information is embedded in an augmented dataset to train an ensemble model using an instance transfer learning framework. The test results on the IEEE 39-bus system verify that this model can accurately predict the stability of previously unseen operational scenarios while reducing the risk of false prediction of unstable instances compared to standard approaches.