Integrating Physics-Based and Deep Learning for Energy Management Optimization Under Measurement Uncertainty: the Case of Fuel Cell Hybrid Electric Vehicles

Combining physics-based and deep learning models creates synergy in applying Energy Management Strategies (EMS) to achieve high-efficiency operating conditions for Fuel Cell Hybrid Electric Vehicles (FCHEVs). However, ensuring an effective learning process while considering uncertainties and fluctuations in operating conditions is challenging due to the complexity of physics-based models with diverse subsystems and dynamics. This study aims to emulate this hybridization in four main steps: data generation, preprocessing, learning, and uncertainty measurement. Firstly, numerous known driving cycles were used to generate a massive amount of data (i.e., 1,355,305 × 16 points) through precise physics-based modeling, ensuring sufficient volume, velocity, and variety of data emulating real-world scenarios accurately. Secondly, the generated data undergoes a well-designed preprocessing process involving a variety of consecutive scaling, denoising, outlier removal, and smoothing-based filtering algorithms and methods to ensure satisfactory quality for learning models. Next, the clean data undergoes a methodically designed deep learning process employing Projected Layer Long Short-Term Memory network (PLSTM) to enhance EMS generalization in FCHEVs. Finally, to mitigate the effects of measurement uncertainty, PLSTM undergoes automatic hyperparameter tuning using Bayesian algorithms, aiming to minimize uncertainty under the Bootstrap Confidence Interval (BCI) features enhancements. The performance of PLSTM is validated through comprehensive visual illustrations, multiple generalizations errors, coefficient of determination, and BCI uncertainty metrics. Further comparative analysis with its variants proves the generalizability and certainty of the methodology in addressing EMS challenges in FCHEVs.