Exploring traffic flow characteristics and predicting its variation patterns are the basis of Intelligent Transportation Systems. The intermittent characteristics and intense fluctuation on short-term scales make it a significant challenge on urban roads. A hybrid model, Genetic Algorithm with Attention-based Long Short-Term Memory (GA-LSTM), combining with spatial–temporal correlation analysis, is proposed in this study to predict traffic volumes on urban roads. The spatial correlation is captured by combining the volume transition matrix estimated from vehicle trajectories and network weight matrix quantified from different detectors. The temporal dependency is explored by the attention mechanism, and we introduce the Genetic Algorithm to optimize it. In the experiment, traffic flow data collected from License Plate Recognition (LPR), is utilized to validate the effectiveness of model. The comparison is conducted with several traditional models to show the superiority of the proposed model with higher accuracy and better stability.

探索交通流特征并预测其变化模式是智能交通系统的基础。间歇性特征和短期尺度上的剧烈波动使其成为城市道路上的重大挑战。本研究提出了一种混合模型，即基于注意力的长时记忆遗传算法（GA-LSTM），并结合时空相关分析，以预测城市道路上的交通量。通过将根据车辆轨迹估算的体积转换矩阵和根据不同检测器量化的网络权重矩阵进行组合，可以捕获空间相关性。通过注意机制探索时间依赖性，并引入遗传算法对其进行优化。在实验中，从车牌识别（LPR）收集的交通流数据，用于验证模型的有效性。与几种传统模型进行了比较，以显示所提出模型的优越性，具有更高的准确性和更好的稳定性。