Advancements in robotic neurorehabilitation have made it imperative to enhance the safety and personalization of physical human-robot interactions (pHRI). Estimation and management of energy transfer between humans and robots is essential for enhancing safety during the rehabilitation. Traditional control methods, which rely on coordinate-based monitoring of robot velocity and external forces, often fail in unstructured environments due to their susceptibility to sensor noise and limited adaptability to individual patient needs. This paper introduces the concept of transactive energy, a coordinate-invariant entity that captures the energy dynamics between the human and the robot during robot-assisted rehabilitation and can be used for personalized robot control. However, estimation of such energy transfer is a complex process and therefore, we have developed a transformer-based model to predict the transactive potential energy. The proposed model is implemented on an ankle rehabilitation robot which is a compliant parallel robot and provides the required three rotational degrees of freedom (DOF). The model learns from the data obtained from the experiments carried out using the ankle robot with five stroke patients on two types of controllers: an impedance controller operated in zero impedance control mode and a trajectory tracking controller. This study provides a baseline, for future research on energy-based control mechanisms in pHRI applications, by utilizing the advanced deep learning models.