Developing energy-aware applications is an important approach to promoting sustainable computing. This paper addresses a fundamental and challenging problem faced by developers and deployers: how to \emph{port} a power model built for one machine to another? We present LucRETius, a novel approach toward portable power modeling through transfer learning, where a pre-trained power model on one machine can serve as a \emph{teacher} to help construct a \emph{student} model on another machine rapidly. The key insight that enables transfer learning is that the layer of application runtimes can abstract away the machine-specific details, so that two machines—despite different hardware and software system configurations—are unified with a common set of runtime events that can serve as features for transfer learning. We evaluate LucRETius through bi-directional transfers across 4 machines, and we show the power models built by LucRETius have a median percentage error of 1.01%-1.75% when used for predicting the energy consumption of 36 real-world JVM-based applications. Compared with training from scratch, LucRETius can lead to a speed up of $8.07!\times$.