The study found that the efficiency benchmark value reveals a negative correlation between energy consumption and traffic capacity, indicating that the energy consumption index soars to 1.25 under mixed disturbance scenarios, representing a 52.4% increase compared to the baseline condition. This phenomenon stems from the strong coupling effect of longitudinal slope resistance and braking frequency. Through Hamiltonian functional optimization, the system compresses energy consumption to a feasible range of 1.12 while maintaining a flow throughput of 1,520 vehicles per hour (as indicated by the data in Table 9). This multi-objective balancing strategy successfully achieves Pareto optimality between traffic efficiency and energy consumption.
The study found a breakthrough improvement in the stability achieved by the control algorithm. Data showed that the MF3DQN-TF framework reduced the standard deviation of speed to 5.1 km/h, a decrease of 41.4% compared to PID control. This achievement stems from the dynamic weighting of environmental mutation factors by the dual-channel attention mechanism (with an illumination gradient weight of 0.35), which compressed the propagation distance of speed oscillations in the black hole effect zone to within 50 m. Especially under conditions where the longitudinal slope gradient is greater than 6%, the Lyapunov exponent remains stable in the safe range of -0.35, confirming that the nonlinear coupled dynamics model successfully suppresses the chain instability effect (as indicated by the data in Table 10).All statistical tests were conducted using a two sample t-test (α = 0.05), with data being the mean ± 95% confidence interval of 10 independent repeated experiments, speed in km/h, and communication in MB.
The experiment confirmed that the safety performance has reached a new industry benchmark, showing that the incidence of high-risk events has been reduced to 3 incidents per 10,000 vehicle-kilometers, a decrease of 62.5% compared to centralized DQN. This breakthrough relies on the multi-objective optimization mechanism of the federated reward function, which controls the peak brake temperature below the critical value of 498 °C through the collaborative constraints of gradient resistance and illumination gradient (as indicated by the data in Table 11). Crucially, the excellent performance of a lateral offset of 0.18 m verifies the precise compensation capability of the aerodynamic model for crosswind disturbances.
The study found that the traffic efficiency achieved energy consumption collaborative optimization, where the MF3DQN-TF framework achieved a throughput of 1920 vehicles/hour with an average travel time of 498 s, and the energy consumption index was 0.82, a reduction of 15.5% compared to the benchmark. This advantage stems from the real-time speed planning of the Hamiltonian function, which calculates through the optimal thrust curve on a 5.8% longitudinal slope section, improving the fuel efficiency of trucks to 7.1L/100 km. It is worth noting that when the CAV penetration rate is greater than 45%, the fleet coordination module triggers an excellent performance with a queue stability index of 0.92, confirming that the transfer-reinforcement fusion architecture successfully achieves knowledge transfer across scenarios (as indicated by the data in Table 12).
The study found that the attention weighting strategy significantly affects control effectiveness, demonstrating that the dual-channel mechanism achieves a standard deviation of 5.1 km/h with an illumination weight of 0.35, a reduction of 25% compared to the uniform weighting strategy. As shown in Fig. 6, this design enhances the speed reduction rate in high illumination gradient areas to 21.3% by dynamically assigning decision-making priorities to environmental and traffic characteristics, specifically addressing the response delay issue in the black hole effect zone encountered by traditional methods. Crucially, the risk-driven strategy compresses events with TTC < 2 s to only 3, confirming that active intervention by environmental mutation factors can reduce the rear-end collision risk by 52% (as indicated by data in Table 13).
Experiments have confirmed the existence of nonlinear correlation in key state responses. For every 2000 lx/s increase in illumination gradient, the system automatically increases the attention weight by 0.06, triggering a safety margin of an average speed reduction of 3.2 km/h. This gradient response mechanism originates from the physiological constraints of the pupil accommodation model. When the illumination mutation exceeds 8000 lx/s, the standard deviation of acceleration decreases from 0.33 to 0.12, stabilizing the increase in safety distance at over 18.2 m, accurately matching the light adaptation delay characteristics of human drivers (as indicated by the data in Table 14).
The study found that computational architecture optimization achieved a breakthrough in real-time performance, with a spatial attention mechanism achieving an inference latency of 27 ms and a computational efficiency of 195 GFLOPs, representing a 31.6% improvement over the fully connected scheme. This optimization reduces the number of parameters through neural architecture search, enabling the dynamic convolution module to achieve an update frequency of 27 Hz in in-vehicle ECUs, meeting the 50 ms control response requirement. Especially in long tunnel scenarios, the local attention mechanism controls the memory peak to 205 MB, successfully addressing the computational bottleneck of high-dimensional state space under resource constraints of edge devices (as indicated by the data in Table 15).
In the study of traffic flow stability in intelligent connected vehicles, the update frequency of the attention mechanism has a significant impact. For example, in the paper focusing on the entrance section of mountain tunnels, considering nonlinear coupling effects, an efficient attention mechanism is required to ensure the accuracy of the analysis. As shown in Fig. 7, different attention mechanisms have varying update frequencies. Mechanisms like Spatial, with higher update frequencies, may be better at capturing the dynamic characteristics of traffic flow in real time. This provides more sensitive attention focus and information processing for traffic flow stability simulation analysis in mountain tunnel entrance sections under the MF3DQN-TF framework, helping to accurately depict the complex behavior of traffic flow under nonlinear coupling conditions.
In the simulation of traffic flow stability for intelligent connected vehicles at the entrance section of mountain tunnels, the memory consumption of attention mechanisms affects system efficiency. This study adopts the MF3DQN-TF framework, which requires attention to memory consumption to adapt to computational resources. As shown in Fig. 8, different attention mechanisms have varying memory requirements, such as the Recurrent mechanism reaching 4.0 GB and the Spatial mechanism requiring 2.5 GB of memory. Selecting a mechanism with reasonable memory requirements can ensure the capture of nonlinear coupling characteristics of traffic flow while avoiding memory overload, thereby facilitating efficient simulation analysis of traffic flow stability at the entrance of mountain tunnels and balancing computational costs with analysis accuracy.
When studying the stability of intelligent connected vehicle traffic flow at the entrance of mountain tunnels using the MF3DQN-TF framework based on attention mechanisms, inference latency affects the real-time nature of the simulation. Figure 9 shows the inference latency of different attention mechanisms, such as the Dual Channel mechanism with a latency of 28 ms. Mountainous scenarios involve complex traffic flows, requiring low-latency mechanisms to ensure real-time decision-making. In this study, attention mechanisms with low inference latency can accelerate information processing, adapt to dynamic changes under nonlinear coupling of traffic flow, and assist the MF3DQN-TF framework in more efficiently simulating the stability of traffic flow at the entrance of mountain tunnels, thereby enhancing the simulation's support for real-time traffic control.
In the simulation study on the stability of intelligent connected vehicle traffic flow at the entrance section of mountain tunnels based on attention mechanisms and the MF3DQN-TF framework, computational load is a critical consideration. Figure 10 shows the computational load of different attention mechanisms, with the Spatial mechanism having a computational load of 195 M. Mountain tunnel scenarios involve complex traffic flow that requires handling nonlinear coupling effects. Attention mechanisms with low computational load can effectively extract traffic flow characteristics while reducing computational pressure, thereby enhancing the efficient operation of the MF3DQN-TF framework. This enables more precise simulation of traffic flow stability, aligning with the dual requirements of computational efficiency and analytical accuracy in real-world engineering applications.
Weight 0.35 is determined through grid search and verified in the [0.2, 0.5] interval. At 0.35, the risk is reduced by 52%
The modular ablation experiment demonstrates that each innovative component makes a significantly differentiated contribution to system performance: the absence of the attention mechanism leads to a sharp degradation in speed control accuracy, validating its core role in dynamically responding to environmental mutations; the removal of the federated learning module significantly increases communication load, highlighting the balanced advantage of the distributed architecture between privacy protection and communication efficiency; the disabling of the transfer learning module significantly prolongs the training convergence period, confirming the effectiveness of cross-scenario knowledge transfer. It is worth noting that the decoupling of the light-slope coupling model causes the most severe performance degradation, emphasizing that environmental physical modeling is the theoretical cornerstone of the framework. The experiment quantitatively reveals the functional boundaries and collaborative mechanisms of each module, providing a reproducible methodological reference for intelligent traffic control in complex environments (as indicated by the data in Table 16).
The study found that there is a clear inflection point in the privacy-performance balance. When the privacy budget ε = 3, the model accuracy reaches 96.2%, and the location information entropy increases to 8.2 bits. This optimization compresses the gradient leakage risk to a new industry low of 0.05% through precise control of the standard deviation of gradient confusion noise σ = 0.05, while reducing the communication overhead to 5.4 MB per round. This mechanism successfully resolves the fundamental contradiction between GDPR compliance requirements and model efficiency, establishing a new paradigm of privacy protection for distributed tunnel control systems (as shown in Fig. 11). Figure 11 shows the sensitivity of ε: as ε increases from 1 to 5, the model accuracy improves from 92.3% to 97.0%, and the positional information entropy increases from 5.8 bits to 9.0 bits. Balance privacy performance selection with ε = 3 (accuracy 96.2% + entropy 8.2 bits).
The experiment confirmed that the federated architecture achieved a breakthrough in convergence efficiency, with the MF3DQN-TF framework reaching a global loss of 0.128 in just 4800 steps, a speedup of 41.5% compared to FedAvg. This advantage stems from the dynamic gated transfer mechanism, which triggers the invocation of source domain knowledge through the mutation strength index, reducing client variance to an extremely low level of 0.15 (as indicated by the data in Table 17). Crucially, the excellent performance with a generalization error of 0.068 verifies the cross-scenario adaptability of the maximum mean difference loss (as shown in Fig. 12).
The study found that communication efficiency has achieved an order of magnitude improvement, with data showing that the single-round transmission volume has been compressed to 42 MB, a reduction of 46.2% compared to FedProx. As shown in Fig. 13, This breakthrough relies on low-rank tensor decomposition technology, reducing bandwidth requirements to the applicable 4G network level of 4.2Mbps. Especially in harsh channels with a packet loss rate of 12.5%, the aggregation delay remains stable within 75 ms, successfully adapting to the weak coverage environment of mountain tunnels (as indicated by the data in Table 18). As shown in Fig. 13, when rank r = 8, the communication volume is 42 MB (53.8% of FedProx). When r < 6, the accuracy decreases by more than 5%. When r > 10, the communication gain is less than 3%. Therefore, r = 8 is selected.
The experiment confirmed that the system possesses extreme operating condition tolerance, showing that the standard deviation of speed only increased by 31.4% to 6.7 km/h under a 50% packet loss rate. As can be seen from Table 19, this robustness stems from the attention weight caching mechanism, which automatically switches to the locally optimal strategy during communication interruptions, compressing the control recovery time to within 3.5 s. In disaster scenarios where the packet loss rate exceeds 90%, the lateral offset remains within the safety threshold of 0.45 m, verifying the environmental adaptability advantage of the light-slope coupling model.
The study found that there is a critical threshold for sensor noise robustness. When the noise standard deviation σ is 0.25, the positioning error increases to 0.25 m, which is still lower than the CAV car-following safety threshold of 0.3 m. This phenomenon stems from the dynamic fusion of multi-source evidence by the attention mechanism. By using Dempster-Shafer theory to increase the reliability weight of the lidar to 0.62, the acceleration response delay is stabilized within 5.7 ms (as indicated by the data in Table 20). This design successfully addresses the core risk of sensor failure at the boundary between light and dark, verifying the environmental adaptability of the heterogeneous sensor redundancy architecture.
Experiments have confirmed that the control efficiency maintains a safety margin under extreme scenarios. Under the condition of dense fog with visibility of 30 m, the standard deviation of speed increases only to 7.1 km/h, and the system degradation rate is controlled at 0.25. The key breakthrough lies in the federated knowledge distillation mechanism, which compresses the lateral offset under crosswind disturbance to 0.38 m through cross-scenario feature alignment (as indicated by the data in Table 21). Under the harsh condition of electromagnetic interference at 30 V/m, the Lie group optimization algorithm still maintains a trajectory tracking error of < 0.15 m, confirming that the light-slope-vehicle coupling model has the ability to ensure stability under all operating conditions.
In the simulation of traffic flow stability for intelligent connected vehicles at the entrance section of mountain tunnels, the MF3DQN-TF framework requires attention to the computational load of the attention mechanism. The spatial mechanism in the Fig. 14 has a computational load of 195 M, and its low load characteristics are suitable for analyzing the nonlinear coupling of traffic flow. This allows for feature extraction while reducing computational pressure, enabling the framework to efficiently simulate the complex traffic flow at the tunnel entrance section. It balances computational efficiency with the accuracy of stability analysis, providing a lightweight and effective attention mechanism option for traffic flow research in intelligent connected vehicle scenarios.
In the study of traffic flow stability for intelligent connected vehicles at the entrance section of mountain tunnels, system degradation under extreme conditions has a significant impact. The paper focuses on the MF3DQN-TF framework and nonlinear coupling effects. Figure 15 shows the system degradation rates under different extreme conditions. In some scenarios, the degradation rate reaches 0.52, far exceeding the acceptable threshold, reflecting the impact of extreme conditions on traffic flow stability simulations. The study requires the use of attention mechanisms for optimization, combining the framework to accurately capture traffic flow characteristics, reduce system degradation under extreme scenarios, and ensure the reliability of traffic flow stability analysis for intelligent connected vehicles at the entrance of mountain tunnels, providing support for addressing complex extreme operating conditions.