基于双流递归LSTM模型的高速公路多粒度交通状态预测研究

Research on Multi-Granularity Traffic State Prediction for Expressways Based on Dual-Stream Recursive LSTM Model

摘要: 针对高速公路改扩建施工期交通流呈现的高度非线性与多变量耦合特征，以及传统断面监测存在的感知盲区问题，本文提出了一种基于双流递归长短期记忆网络(Dual-Stream Recursive LSTM)的多粒度交通状态预测方法。首先，构建微观–中观–宏观分层空间感知体系，实现对100 m微观路段的精准映射与全域覆盖；其次，设计双流独立预测架构，通过解耦拥堵指数与平均车速的特征学习，有效解决了多变量量纲差异导致的梯度干扰难题；最后，引入递归滚动交互输入机制，利用变量间的物理交互约束修正长时序预测轨迹。基于常虎与莞深高速改扩建工程实测数据的实验表明：该方法在微观粒度下的拥堵指数预测精度(MAE)较单流模型提升约35%，显著增强了对拥堵起步阶段的感知敏锐度；同时，在未来60分钟的长时预测中，车速预测误差(RMSE)降低了27%，有效克服了传统递归预测的不稳定性。该模型在宏观至微观全域范围内均表现优异，可为施工期车道级精细化管控提供可靠的决策支持。

Abstract: Aiming at the highly nonlinear and multi-variable coupling characteristics of traffic flow during expressway reconstruction and expansion, as well as the blind spots in traditional section-based monitoring, this paper proposes a multi-granularity traffic state prediction method based on a Dual-Stream Recursive Long Short-Term Memory (LSTM) network. First, a hierarchical “Micro-Meso-Macro” spatial sensing system is constructed to achieve precise mapping and full-domain coverage of 100m micro-segments. Second, a dual-stream independent architecture is designed to decouple the feature learning of the congestion index and average speed, effectively solving the gradient interference problem caused by dimensional heterogeneity. Finally, a recursive rolling mutual-feeding mechanism is introduced to correct long-term prediction trajectories by leveraging physical interaction constraints between variables. Experiments based on field data from the Changhu and Guanshen expressway reconstruction projects demonstrate that the proposed method improves the prediction accuracy (MAE) of the congestion index at the micro-granularity level by approximately 35% compared to single-stream models, significantly enhancing sensitivity to congestion onset. Furthermore, in the 60-minute long-term prediction, the root mean square error (RMSE) of speed prediction is reduced by 27%, effectively overcoming the instability of traditional recursive prediction. The model exhibits superior performance across macro to micro spatial granularities, providing reliable decision support for lane-level refined control during construction.