A New Energy Vehicle Autonomous Driving Decision Control Method Integrating Neural Networks and Reinforcement Learning

Abstract

Autonomous driving has emerged as a capable technology to improve the energy efficiency and safety of New Energy Vehicles (NEVs) by enabling intelligent driving decisions. Conventional energy-efficient autonomous driving methods are often limited to longitudinal velocity planning and fixed-route strategies, which fail to achieve optimal adaptability in dynamic traffic conditions. To overcome these challenges, this research proposes an autonomous driving decision control method integrating Neural Networks (NNs) with Reinforcement Learning (RL). In the proposed framework, convolutional neural networks (CNNs) extract spatial-temporal features from multi-sensor inputs, including camera, Light Detection and Ranging (LiDAR), and radar data, to capture lane geometry, vehicle distances, and obstacle dynamics. The data were preprocessed using Min-Max normalization to scale all features to a consistent range. These features form the state space for a Capuchin Search-driven Scalable Deep Q-Network (CapSA-SDQN)-based RL agent, which simultaneously manages lane-changing and vehicle-following behaviors and optimizes decision control in complex and dynamic traffic environments. The decision control method further incorporates a rule-based safety checker that is embedded downstream to guarantee safe maneuver execution. The strategy is trained and tested in the simulation platform under diverse traffic scenarios, including congested urban intersections, highway merging, and dynamic overtaking. Results show that the proposed CapSA-SDQN approach achieves significant improvements in average speed (46.72) and reduced speed standard deviation (2.15). The findings demonstrate that integrating neural networks with reinforcement learning provides a scalable solution for NEV autonomous driving in real-world environments.