基于时间感知Transformer的智能网联汽车交通流预测系统

Intelligent Connected Vehicle Traffic Flow Prediction System Based on Time Aware Transformer

摘要:

由于交通数据具有高度非线性和时空耦合特性，传统时序模型难以有效捕捉长距离依赖关系，导致预测精度受限；同时，多源异构数据的时空对齐误差会进一步降低预测可靠性，尤其在复杂交通场景（如拥堵、突发事故）下，误差累积效应加剧，导致汽车交通流预测偏差较大。基于此，提出基于时间感知Transformer的智能网联汽车交通流预测系统设计方法。在硬件层面，通过优化多源数据采集设备（高精度GNSS/INS、多线激光雷达）确保原始数据质量，部署5G边缘计算节点实现低延迟预处理，有效缓解时空耦合特性带来的传输延迟问题。在软件层面:采用时空对齐模块（高斯滤波+坐标转换）消除多源数据时空偏差，解决误差累积问题；其次设计改进的时间感知Transformer模型，通过时间嵌入层捕捉周期规律，结合局部-全局混合注意力机制建模长距离依赖关系，并引入膨胀因果卷积增强非线性特征提取能力。动态适应交通流的周期性及突变特征，提升预测精准性。通过硬件与软件的协同运作，实现了交通流的精准预测。消融实验结果显示:时间感知Transformer模型应用后交通流预测误差最小值达到了0.1%，充分证实了改进模型的有效性。对比实验结果显示:设计系统应用后交通流数据时间戳同步误差最大值仅有0.08%，空间坐标偏移量最大值仅有0.16mm，交通流预测结果与实际结果趋于一致。

Abstract:

Due to the highly nonlinear and spatiotemporal coupling characteristics of traffic data, traditional time series models are difficult to effectively capture long-distance dependencies, resulting in limited prediction accuracy; At the same time, the spatiotemporal alignment error of multi-source heterogeneous data will further reduce the reliability of prediction, especially in complex traffic scenarios such as congestion and sudden accidents, where the cumulative effect of errors is exacerbated, leading to significant deviations in predicting automobile traffic flow. Based on this, a design method for an intelligent connected vehicle traffic flow prediction system based on time aware Transformer is proposed. At the hardware level, the original data quality is ensured by optimizing multi-source data acquisition equipment (high-precision GNSS/INS, multi line laser radar), and 5G edge computing nodes are deployed to achieve low latency preprocessing, effectively alleviating the transmission delay problem caused by spatio-temporal coupling characteristics. At the software level, a spatiotemporal alignment module (Gaussian filtering+coordinate transformation) is used to eliminate spatiotemporal biases in multi-source data and solve the problem of error accumulation; Next, design an improved time aware Transformer model that captures periodic patterns through a time embedding layer, models long-range dependencies using a local global hybrid attention mechanism, and introduces dilated causal convolution to enhance nonlinear feature extraction capabilities. Dynamically adapting to the cyclical and abrupt characteristics of traffic flow to improve prediction accuracy. Through the collaborative operation of hardware and software, accurate prediction of traffic flow has been achieved. The results of the ablation experiment showed that the minimum prediction error of traffic flow after the application of the time aware Transformer model reached 0.1%, fully confirming the effectiveness of the improved model. The comparative experimental results show that the maximum timestamp synchronization error of traffic flow data after the application of the designed system is only 0.08%, and the maximum spatial coordinate offset is only 0.16mm. The predicted traffic flow results are consistent with the actual results.