基于改进CNN-LSTM组合模型的分时段短时交通流预测

Short-Term Traffic Flow Prediction Method of Different Periods Based on Improved CNN-LSTM

针对现有预测模型不能充分提取交通流时空特征的问题,提出一种基于改进卷积神经网络(convolutional neural network,CNN)和长短时记忆(long short-term memory,LSTM)神经网络的短时交通流预测方法。首先,采用分层提取方法使设计的网络结构和一维卷积核函数自动提取交通流序列的空间特征;其次,优化LSTM网络模块来减少网络对数据的长时间依赖;最后,在端对端模型的训练过程中,引入改进后的自适应矩估计(rectified adaptive moment estimation,RAdam)优化算法,加快权重的拟合并提高网络输出的准确性和鲁棒性。实验结果表明:在工作日与周末分时段,所提出的模型相比堆栈自编码(stacked auto-encoders,SAEs)网络预测模型,性能分别提升3.55%与8.82%,运行时间分别缩减6.2%与6.9%;相比长短时记忆网络-支持向量回归(long-short term memory-support vector regression,LSTM-SVR)预测模型,性能分别提升0.29%与1.79%,运行时间分别缩减9.0%与9.7%。所提模型能够更加适用于不同时段下的短时交通流预测。

Aiming at solving the problem that existing prediction models could not fully extract the spatio-temporal features in traffic flow,we proposed an improved convolutional neural network(CNN)with long short-term memory neural network(LSTM)for short-term traffic flow prediction.First of all,a layered extraction method was used to design the network structure and one-dimensional convolution kernel which enabled automatic extraction of spatial features of traffic flow sequences.Second,the LSTM network modules were optimized to reduce the long-term dependence of network on the data.Finally,the optimization algorithm for rectified adaptive moment estimation(RAdam)was introduced to the end-to-end model training process,which accelerated fitting effects of the weight and improved the accuracy and robustness of network output.Experimental results showed that compared with the prediction model of stacked auto-encoders(SAEs)network,performance of the proposed model was enhanced by 3.55%and 8.82%on weekdays and weekends with model running times reduced by 6.2%and 6.9%,respectively.Compared with the prediction model of long-short term memory-support vector regression(LSTM-SVR),its performance was enhanced by 0.29%and 1.79%with model running times reduced by 9.0%and 9.7%,respectively.Therefore,the proposed model was more applicable to the short-term traffic flow prediction of different time periods.