<b>Stability Analysis of Intelligent Connected Vehicle Traffic Flow at Mountain Tunnel Entrances Using the MF3DQN-TF Framework with Attention Mechanism and Nonlinear Coupling Effects</b>

The nonlinear coupling effect triggered by the abrupt change in light environment and complex longitudinal slope at the entrance section of mountain tunnels deteriorates the stability of intelligent and connected mixed traffic flow. Traditional control methods exhibit significant limitations in terms of dynamic environmental adaptability, data privacy protection, and cross-scenario generalization ability. Addressing the multi-physics field coupling mechanism of light-slope-vehicle, this study proposes the MF3DQN-TF intelligent control framework, achieving breakthroughs through triple innovations. Firstly, a dynamic coupling equation between light gradient attenuation and longitudinal slope resistance is established, and a six-dimensional joint state vector is constructed to quantify the intensity of the “black hole effect”. A dual-channel attention mechanism assigns a decision weight of 0.35 to environmental abrupt change factors, compressing the propagation distance of speed oscillations to within 50 meters. Secondly, a federated transfer learning architecture is designed, integrating gradient obfuscation differential privacy and low-rank tensor decomposition techniques. The communication overhead is reduced to 42MB per round, maintaining 92% control stability even in an environment with a 20% packet loss rate. Finally, a dynamic gated transfer mechanism is developed, achieving cross-scenario knowledge distillation such as in rainstorm tunnels through maximum mean discrepancy loss, reducing the target domain convergence steps by 61.6%. Simulation verification shows that this framework reduces the standard deviation of speed to 5.1 km/h, reduces high-risk events with TTC &lt; 2s to 3 per 10,000 vehicle-kilometers, stabilizes the peak brake temperature within the safety threshold of 498℃, and controls the lateral offset to within 0.18 meters. This achievement marks a paradigm shift in tunnel traffic control from single-vehicle control to environment-vehicle collaborative governance, providing theoretical support and technical paradigm for all-weather safe passage on high-risk road sections. Keywords: Mountain Tunnel Traffic Llow, Nonlinear Coupling Effect, Federated Reinforcement Learning, Attention Mechanism, Dynamic Collaborative Control, Robust Optimization山区隧道入口段光环境突变与复杂纵坡引发的非线性耦合效应（Nonlinear Coupling Effect）劣化了智能网联混合交通流的稳定性。传统控制方法在动态环境适应性、数据隐私保护及跨场景泛化能力方面存在显著局限。针对光-坡-车多物理场耦合机制，本研究提出MF3DQN-TF智能控制框架，通过三重创新实现突破。其一，建立光梯度衰减（light gradient attenuation）与纵坡阻力（longitudinal slope resistance）的动态耦合方程，构建六维联合状态向量（six-dimensional joint state vector）以量化“黑洞效应（black hole effect）”强度；双通道注意力机制（dual-channel attention mechanism）为环境突变因素分配0.35的决策权重，将速度振荡传播距离压缩至50米以内。其二，设计联邦迁移学习（federated transfer learning）架构，集成梯度混淆差分隐私（gradient obfuscation differential privacy）与低秩张量分解技术（low-rank tensor decomposition techniques）；通信开销（communication overhead）降至每轮42MB，即使在20%丢包率（packet loss rate）环境下仍保持92%的控制稳定性。其三，开发动态门控迁移机制（dynamic gated transfer mechanism），通过最大均值差异损失（maximum mean discrepancy loss）实现暴雨隧道等跨场景知识蒸馏（knowledge distillation），将目标域收敛步数（target domain convergence steps）减少61.6%。仿真验证表明，该框架将速度标准差（standard deviation of speed）降至5.1 km/h，TTC<2s的高风险事件（high-risk events）降至每万车公里3起，峰值制动温度（peak brake temperature）稳定在498℃安全阈值（safety threshold）内，横向偏移量（lateral offset）控制在0.18米以内。该成果标志着隧道交通控制从单车控制向环境-车辆协同治理（environment-vehicle collaborative governance）的范式转变（paradigm shift），为高风险路段全天候安全通行（all-weather safe passage）提供理论支撑与技术范式（technical paradigm）。关键词：山区隧道交通流（Mountain Tunnel Traffic Flow）、非线性耦合效应、联邦强化学习（Federated Reinforcement Learning）、注意力机制（Attention Mechanism）、动态协同控制（Dynamic Collaborative Control）、鲁棒优化（Robust Optimization）