Machine Learning-Based Methodology for Intelligent Energy Management Strategy in Heavy-Duty Fuel Cell Hybrid Electric Vehicles with Pantograph

Abstract

This work presents the proposal and validation of a novel methodology for enhancing energy management strategies in heavy-duty FCHEVs with overhead current collector (pantograph) by means of ML-based predictions for the characterization of the driver-vehicle system from a holistic approach, correlating its energy and power profiles with characteristics of the route where it transits, specifically the speed profile and the height profile for the complete route. The base concept is the possibility of characterizing the vehicle’s energy use from an approach that also considers the driver behavior and road characteristic. This data-driven characterization using historic and real-time data stream from the vehicle allowed for a ML-model to be trained to make predictions using limited information from the upcoming route. The predictions created with the described method included energy demand, power base-values and optimal SoC profiles. These predictions were then used in an energy management strategy by means of a heuristic controller that received and used the optimal SoC profile and the power demand base-value of the complete route, thus allowing the controller to perform in accordance to current and upcoming vehicle energy demand. The methodology begins with the clusterization of vehicle and road data to define zonetypes for assigning labels to individual samples by means of unsupervised machine learning. Next, the labeled data is used to train a supervised machine learning classification algorithm which is then used to make predictions about the upcoming route. The clusterization and zone type predictions are then used for discretizing the complete route in a step called zonification, where the route is divided into sections with their own characteristics, providing a base on which the energy management strategy can be adjusted and executed accordingly. The data used for these tasks included vehicle dynamics data and energy demand profiles, as well as road information. The information used regarding the road was the expected speed profile and the elevation profile of the route. Both of these features can be obtained from external sources like vehicle to X communication or third-party navigation services and cartography. The ML-enhanced EMS controller was then validated through simulation using real data from 5 different routes in Germany and its performance was compared to that of other 3 controllers which made use of different approaches for the actuation and control of the onboard energy systems. The results were consistent in demonstrating the superior performance of the controller making use of the ML predictions, obtaining the best scores in FC degradation and H2 mass consumption indexes, with 25% and 27% less than the next best performer on each index respectively.