Authors

H. Lee, M. S. Song, T. W. Yoon, H. J. Lee, J. S. Lee, Hyundai Motor Company, Hwaseong, Korea

Year

2026

Print Info

Production/Publication ÖVK

Summary

To maximize the energy efficiency of Hybrid Electric Vehicles (HEVs), precise power management strategies reflecting future driving conditions are essential. This study proposes two advanced short-term power prediction models—a Neural Network and a Physics-Guided Neural Network (PGNN)—to enhance the lower layer of a Hierarchical Predictive Control (HPC) system. Unlike conventional linear regression methods that fail to capture non-linear interactions between variables such as road gradient and relative speed, the proposed models were trained using supervised learning on real-world driving data to predict power demand for 1 to 5 seconds ahead. Comparative analysis based on statistical metrics and driving scenarios demonstrated that the Neural Network model yielded superior accuracy, particularly in longer horizons where the linear model's performance degraded. While the PGNN offered generalization benefits, the Neural Network proved more effective for immediate application. To address the computational constraints of vehicle controllers and the "black-box" nature of AI, the trained Neural Network was calibrated into an interpretable Lookup Table (LUT) structure. This approach ensures real-time responsiveness without heavy computational load. Simulation results confirmed that the improved HPC logic achieved approximately 17% higher prediction accuracy for the 5-second horizon than existing methods, enabling faster, pre-emptive engine On/Off responses and significantly improved energy efficiency and SOC optimization.

ISBN

978-3-9504969-5-6

DOI

https://doi.org/10.62626/u94c-7og6 Copy link

Order Members Login