短时交通流量预测对于提高实时交通数据信息的精准性及增加车辆道路行驶的效益性具有重要意义。为能准确预测未来短期交通流量情况，支持智能交通系统的应用和决策，提出一种基于CNN-LSTM-AM的短时交通流量预测模型。首先利用卷积神经网络（convolutional neural network, CNN）来对交通流序列进行信息捕捉，从而提取交通流数据的动态变化特征；其次将所提取的特征向量构成时间序列作为长短期记忆（long short-term memory, LSTM）网络的输入；最后根据注意力机制（attention mechanism, AM）来分配LSTM隐含层不同权重，增强重要特征的作用，完成交通流量预测。采用美国加利福尼亚州高速路网数据库PeMS里面的相关数据信息，通过实验与其他神经网络预测模型进行对比，结果显示，CNN-LSTM-AM模型的相对平均误差（mean absolute percentage error, MAPE）值为0.254 578%，R² = 0.583 152，预测能力优于其他对比模型。其所用方法可以对未来短时交通流量预测提供一种思路模型。

The ability to forecast short-term traffic flow is crucial for enhancing the precision of real-time traffic data and expanding the advantages of driving a car. A short-term traffic flow forecast model based on CNN-LSTM-AM was suggested to facilitate the application and decision-making of intelligent transportation systems by reliably predicting the short-term traffic flow scenario in the future. Initially, the traffic sequence of traffic sequences was captured using a convolutional neural network (CNN) in order to extract the dynamic changes of the traffic flow data. Second, the long short-term memory (LSTM) network used the remarkable vector of the time sequence extraction as its input. Ultimately, the attention mechanism (AM) was used to assign various weights to the LSTM hidden layer, emphasise the significance of key attributes, and finalise the traffic flow estimate. Using pertinent data from the PEMS expressway database in California, USA. Through experiments, it was contrasted with other neural network prediction models. The CNN-LSTM-AM model's mean absolute percentage error (MAPE) value is 0.254 578%, and its R² value is 0.583 152, according to the results. Compared to earlier comparable models, the model reported has a higher prediction capacity. The approach taken can offer a conceptual model for forecasting future short-term traffic flow.