HPC-SafeRL: A predictive context-aware safe reinforcement learning framework for autonomous lane changing in multiple scenarios

Autonomous lane-changing decision-making requires a balanced treatment of safety, comfort, and efficiency under dynamic traffic interactions. Existing SafeRL methods mainly emphasize constraint satisfaction or reward-penalty design, while MPC-RL approaches often rely on separate planning or model-based optimization modules and provide limited adaptation to context-dependent objective conflicts. To address these limitations, this paper proposes Hybrid-action Predictive Context-aware Safe Reinforcement Learning (HPC-SafeRL), which integrates Hybrid Action Representation, CFC-based predictive rollout, Safety Mask filtering, and fuzzy logic-based dynamic reward weighting within a unified TD3-based decision framework. The novelty of HPC-SafeRL lies in coupling hybrid action abstraction with short-term predictive safety evaluation and context-aware reward adaptation, enabling the agent to coordinate discrete lane-changing decisions and continuous control parameters while adjusting the safety-comfort-efficiency trade-off across scenarios. SUMO experiments show that HPC-SafeRL achieves a balanced performance profile. Compared with MP-DQN and PPO-Lagrangian, it reduces mean collision risk by 51.6% and 60.8%, respectively, while maintaining a higher average speed of 29.35 m/s. Ablation results further show that removing the prediction module increases the collision rate from 16.62% to 35.43% and reduces the average speed from 29.35 to 16.57 m/s, while removing dynamic reward weighting increases the collision rate to 32.03%. These results demonstrate that HPC-SafeRL improves proactive risk handling, action stability, and adaptive multi-objective decision-making for autonomous lane changing.