Abstract: In order to overcome that the existing short-term traffic flow prediction methods failed to fully consider the randomness and nonlinearity of traffic flow, a short-term traffic flow prediction model construction method based on ensemble empirical mode decomposition algorithm (EEMD) combined with wavelet threshold was proposed. Firstly, EEMD was used to decompose the original traffic flow data into N intrinsic mode functions (IMF) and a residual component (Res). Secondly, wavelet analysis was used to process the eigenmode function of noisy signal. Finally, two types of model construction methods were proposed. Method 1: the N IMF and Res processed by wavelet analysis were reconstructed and input into long short-term memory (LSTM) model, sequence to sequence (Seq2seq) model and Seq2seq-Attention model respectively, and their output were the final predicted values. Method 2: the N IMF and Res processed by wavelet analysis were input into long short-term memory (LSTM) model, sequence to sequence (Seq2seq) model and Seq2seq-Attention model respectively, and the output of these models was the predicted value of each component, which was the final predicted value after reconstruction. The two types of model prediction methods were compared with the initial LSTM model, Seq2seq model and Seq2seq-Attention model as well as the combined prediction models based on wavelet analysis and LSTM model, Seq2seq model and Seq2seq-Attention model. The results show that the two types of model construction methods can significantly improve the prediction performance of the initial prediction model. Compared with the combined prediction model based on wavelet analysis, its prediction performance has been improved; and compared with method ①, method ② has a more significant effect on improving the performance of the model.

Key words: traffic engineering; short-term traffic flow prediction; ensemble empirical mode decomposition（EEMD）; wavelet threshold

摘要： 为克服现有短时交通流预测方法未能充分考虑交通流的随机性、非线性特征，提出一种基于集合经验模态分解（empirical mode decomposition, EEMD）结合小波分析的短时交通流预测模型构建方法。首先，利用EEMD将原交通流数据分解为N个本征模态函数（intrinsic mode fuction, IMF）和1个趋势项（residual, Res）；其次，使用小波分析对含噪声信号的本征模态函数进行小波分析处理；最后，提出两类模型构建方法：①将经过小波分析处理后的N个IMF和Res进行重构，将其分别输入长短期记忆网络模型（long short-term memory, LSTM）、序列模型（sequence to sequence, Seq2seq）和引入注意力机制序列模型（seq2Seq attention），模型输出即为最终预测值（方法1）；②将小波分析处理后的IMF和Res分别输入LSTM模型、Seq2seq模型和Seq2Seq Attention模型，模型输出为各分量预测值，将其重构后即为最终预测值（方法2）。将两类模型预测方法分别与初始LSTM、Seq2seq和Seq2Seq Attention模型以及基于小波分析与LSTM、Seq2seq和Seq2Seq Attention模型的组合预测模型进行对比实验，结果表明：两类模型构建方法能够显著提升初始预测模型的预测性能，相较于基于小波分析的组合预测模型，其预测性能均有所提升，且相较于方法1，方法2对模型性能的提升效果更加显著。

关键词: 交通工程；短时交通流预测；集合经验模态分解；小波分析