Physics-based hybrid machine learning models are gaining popularity due to their ability to capture both known physics and unknown dynamics introduced by sensory data. However, such models are often large and computationally intensive compared to purely data-driven models. To address this challenge, this study proposes a physics-guided feature transfer (PgFT) framework that selectively extract the physical and meaningful features from a pretrained hybrid physics-data model to a functional learner model. The functional learner is a light-sized predictor lessening learning burdens with guarantying high accuracy. The proposed framework is modeled for predicting the battery performance of an electric vehicle (EV) under real-world operation scenarios. Despite the accounted dynamic noise and significant nonlinearities in each scenario, the functional learner model—trained using the PgFT strategy—successfully learns the key physical behaviors and adapts an EV dynamically.
Noureldin et al. (Sat,) studied this question.