The topological configuration of a bulk power grid is often altered by network investment upgrades, forecasted disasters and random faults, as well as planned operator-triggered transmission line maintenance and controlled switching actions. Such topological variations can drastically change the measurement data distribution from phasor measurement units (PMUs), which may in turn compromise the accuracy of the artificial intelligence (AI)-aided monitoring and control applications using the measurements. For instance, data-driven transient stability assessment (TSA) models that were trained with static network topologies may no longer be accurate for monitoring power grid stability as the network topology changes. Not only would the number of possible topology changes be too vast to train all possible scenarios, but also the training process will render computationally intensive. This paper proposes a model-based transfer learning (TL) approach that integrates a convolutional neural network and a long short-term memory network (ConvLSTM), to efficiently train a new stability prediction model that predicts the system operating states (SOSs) and identify critical generators (CGs) in case of instability when the system undergoes enduring topological changes. Numerical analyses on three test cases including the IEEE 39-bus test system, the IEEE 118-bus test system, and the large-scale 2000-bus synthetic power grid in the state of Texas verify the efficiency of the proposed approach and highlight benefits in training time and accuracy, when compared to the state-of-the-art alternatives. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —As the national power grid goes through transitions towards digitalization and modernization with emerging technologies, maintaining its reliability and resiliency against environmental stressors and cyber attacks remains an urgent need. The power system topology is expected to change more frequently, sometimes to accommodate the proliferation of heterogeneous distributed energy resources and tackle their intermittence, sometimes event-driven due to disruptive events (e.g., faults), and sometimes operator-triggered for maintenance activities and responsive control to return the system back to its normal operating condition. This article is motivated by the need for an efficient and computationally-attractive approach for online situational awareness and real-time transient stability monitoring and assessment of the power grid under an enduring topological change, where the main goal is to identify the power system operating states (SOSs) and critical generators (CGs) in case of instability. Instead of training a new model for each topological change, this paper proposes an adaptive power system transient stability assessment (TSA) method that uses transfer learning (TL) which in return reduces the training time yet with less data than a newly-trained model for each topology change scenario.