基于SSA-CNN-BiLSTM组合模型的短时交通流量预测

Short-Term Traffic Flow Prediction Based on SSA-CNN-BiLSTM Combination Model

为改善城市道路交通拥堵状况,并为智能交通系统决策提供辅助手段,针对短时交通流的非线性和时序性特点,构建了一种基于麻雀搜索算法(SSA)优化的卷积神经网络(CNN)联合双向长短时记忆神经网络(BiLSTM)的组合模型以预测短时交通流量。首先,对原始交通流数据进行异常值清洗、小波阈值去噪和归一化处理。然后,利用SSA算法对CNN与BiLSTM组合网络中的隐藏层单元数、初始学习率和L2正则化系数三个超参数迭代寻优。最后,将搜索得到的最优超参数组合输入搭建好的组合网络中进行训练和预测。实验结果显示:与粒子群优化(PSO)和灰狼优化(GWO)算法相比,SSA算法在网络超参数寻优过程中的收敛速度更快,全局寻优能力更强;与3种对比模型(CNNBiLSTM、BiLSTM和LSTM)相比,在5 min时间尺度划分下,SSA-CNN-BiLSTM组合模型的均方根误差(RMSE)分别降低了5.46、12.78、20.38,平均绝对百分比误差(MAPE)分别降低了0.49%、2.24%、3.11%;在15 min时间尺度划分下,SSA-CNN-BiLSTM组合模型的RMSE分别降低了9.70、28.42、41.18,MAPE分别降低了0.50%、1.98%、2.59%。研究表明,相比既有算法,该短时交通流量预测组合模型在精度和稳定性上都有所提升,可通过提供更精准的短时交通出行信息来改善道路交通状况。

In order to solve the problem of urban road traffic congestion and provide auxiliary means for decision-making of intelligent transportation systems,a combination model of SSA-CNN-BiLSTM was constructed to predict short-term traffic flow,taking into account the nonlinear and temporal characteristics of short-term traffic flow.Firstly,the original traffic flow data was subjected to outlier cleaning,wavelet threshold denoising,and normalization processing.Secondly,the SSA algorithm was used to iteratively optimize the three hyperparameters of the number of hidden layer units,initial learning rate,and L2 regularization coefficient in the CNN-BiLSTM composite network.Finally,the optimal hyperparameter combination obtained from the search was input into the constructed combination network for training and prediction.The experimental results showed that compared with PSO and GWO algorithm,SSA algorithm had a faster convergence speed and stronger global optimization ability in the process of network hyperparameter optimization.Compared with the three comparison models of CNN-BiLSTM,BiLSTM,and LSTM,the SSA-CNN-BiLSTM combination model showed a reduction of 5.46%,12.78%,and 20.38%in RMSE,and a decrease of 0.49%,2.24%,and 3.11%in MAPE respectively at a 5-minute time scale;and under the 15-minute time scale division the RMSE of SSA-CNN-BiLSTM model was decreased by 9.70,28.42,and 41.18,while the MAPE was decreased by 0.50%,1.98%,and 2.59%,respectively.Research has shown that the accuracy and stability of the short-term traffic flow prediction model of SSA-CNN-BiLSTM have been improved compared to existing algorithms,and road traffic conditions can be improved by providing more accurate shortterm traffic travel information.