Scholarlink

This paper presents an intelligent braking and energy recovery control strategy for autonomous electric vehicles using deep recurrent neural architectures. Vehicle dynamics, adhesion feedback, and regenerative power prediction of braking are merged for optimizing deceleration stability and efficiency. It was evaluated with real-world driving conditions, including front/rear wheel velocity responses, changing passenger load conditions (100–900 kg), and adhesion command–execution tracking. For the first case, it demonstrates steady deceleration from 80 km/h to rest in 5 s with a low front–rear slip, whereas for the second, it exhibits dynamical oscillations picked up by RNN-based estimations with fast braking events. Generated power varies non-linearly with increased vehicle weight, maximum output approximately 1.2 kW for 900 kg, affirming increased load increases recovery potential. Adhesion control loop exhibits rapid convergence (≈1.4 s) between executed and commanded displacements, affirming proposed controller responsiveness. Overall, deep learning architecture achieves a good compromise between brake stability as well as regeneration efficiency, but a reasonable starting point for real-time energy maximization for further autonomous vehicles.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive Deep Learning-Based Energy Control for Autonomous Vehicles Under Variable Passenger Loads and Dynamic Adhesion Conditions

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nidal Ghalim,
  • Souad Touairi,
  • Hanaa Ouaomar,
  • Nourreeddine Kouider

摘要

This paper presents an intelligent braking and energy recovery control strategy for autonomous electric vehicles using deep recurrent neural architectures. Vehicle dynamics, adhesion feedback, and regenerative power prediction of braking are merged for optimizing deceleration stability and efficiency. It was evaluated with real-world driving conditions, including front/rear wheel velocity responses, changing passenger load conditions (100–900 kg), and adhesion command–execution tracking. For the first case, it demonstrates steady deceleration from 80 km/h to rest in 5 s with a low front–rear slip, whereas for the second, it exhibits dynamical oscillations picked up by RNN-based estimations with fast braking events. Generated power varies non-linearly with increased vehicle weight, maximum output approximately 1.2 kW for 900 kg, affirming increased load increases recovery potential. Adhesion control loop exhibits rapid convergence (≈1.4 s) between executed and commanded displacements, affirming proposed controller responsiveness. Overall, deep learning architecture achieves a good compromise between brake stability as well as regeneration efficiency, but a reasonable starting point for real-time energy maximization for further autonomous vehicles.