交通流序列多为单步预测。为实现交通流序列的多步预测，提出一种基于编码器解码器（encoder-decoder， ED）框架的长短期记忆网络（long short-term memory， LSTM）模型，即ED LSTM模型。将自回归滑动平均、支持向量回归机、XGBOOST、循环神经网络、卷积神经网络、LSTM作为对照组进行实验验证。实验结果表明，当预测时间步长增加时，ED框架能够减缓模型性能的下降趋势，LSTM能够充分挖掘时间序列中的非线性关系。除此之外，在单变量输入的情况下，在PEMS-04数据集上，当预测时间步长为t+1到t+12的12个时间步时，ED LSTM模型的均方根误差（root mean squard error，RMSE）及平均绝对误差（mean absolute error，MAE）分别下降0.210~5.422、0.061~0.191。相较于单因素输入，多因素输入的ED LSTM模型在12个预测时间步长下，RMSE、MAE分别下降0.840、0.136。实验证明了ED LSTM模型能够有效地用于交通流序列的多步及单因素、多因素预测任务。

Most of the traffic flow sequences are single-step prediction. To realize multi-step prediction of traffic flow sequence, a long short-term memory (LSTM) model based on encoder-decoder (ED) framework was proposed. To verify the proposed encoder-decoder LSTM multi-step traffic flow prediction model (ED LSTM), autoregressive moving average, support vector regression machine, XGBOOST, recurrent neural network, convolutional neural network and LSTM were used as control groups for the experiment. Experimental results show that when the prediction time step increased, ED framework could slow down the decline of model performance, and LSTM could fully mine the nonlinear relationship in time series. In addition, under the condition of univariate input, the root mean squard error (RMSE) and mean absolute error (MAE) of ED LSTM model decreased by about 0.210-5.422 and 0.061-0.192, respectively, on PEMS-04 dataset with 12 time steps from t+1 to t+12. Compared with single-factor input, the ED LSTM model with multi-factor input decreased RMSE and MAE by about 0.840 and 0.136 respectively under 12 prediction time steps, demonstrating that ED LSTM model can be effectively applied to multi-step and single-factor and multi-factor forecasting of traffic flow series.

U491

公安部公安理论及软科学研究计划项目（2020LLYJGADX020）；中国人民公安大学拔尖创新人才培养经费支持研究生科研创新项目成果（2021yjsky014）；中国人民公安大学公共安全行为科学实验室开放课题基金资助（2020SYS15）。