Abstract: Urban elevated expressway traffic flows exhibit characteristics such as pronounced nonlinearity, significant temporal correlation, and complex fluctuations. To enable accurate forecasting of short-term traffic conditions, this study proposes a recursive prediction model for future traffic volumes based on the long short-term memory (LSTM) neural network. Taking the Jiaoning Road elevated expressway in Shinan District, Qingdao as the study subject, this paper conducts short-term traffic flow forecasting based on actual traffic flow data obtained from gantry monitoring. The model's predictive performance is comprehensively evaluated using metrics including the coefficient of determination, mean absolute error, mean deviation error, and root mean square error. Comparative analysis between the LSTM reference model and back propagation neural network optimized by genetic algorithm (BP-GA) and convolutional neural network (CNN) prediction models reveals that the LSTM model demonstrates superior predictive accuracy and robust generalisation capability. With a coefficient of determination reaching 0.606 and overall minimal prediction errors, its forecasting performance surpasses the other two models. This approach provides valuable support for predicting traffic flow at future time points.

Key words: LSTM algorithm; BP-GA algorithm; CNN algorithm; urban expressway; traffic flow prediction

摘要： 城市高架快速路交通流具有非线性强、时序相关性显著及波动性复杂等特点。为能够 准确预测未来短期的交通流状况，提出基于长短期记忆网络（LSTM）模型的未来交通流量的 递归预测。文章以青岛市市南区胶宁路高架快速路为研究对象，基于门架监测获取的实际交 通流数据，开展短时交通流预测研究。采用决定系数、平均绝对误差、平均偏差误差和均方 根误差等指标对模型预测性能进行综合评价。以 LSTM 为参照与遗传算法优化的反向传播神 经网络（BP-GA）、卷积神经网络（CNN）预测模型进行对比分析。结果显示，LSTM 模型具 有较高的预测精度和良好的泛化能力，其决定系数达到 0.606，预测误差整体较小，预测能力 优于其他两种模型，对未来时刻的交通流预测提供帮助。

关键词: LSTM 算法；BP-GA 算法；CNN 算法；城市快速路；交通流预测