天津职业技术师范大学学报

针对交通流量序列的时间依赖性、空间相关性及易受外部因素的干扰等问题，提出一种基于多尺度卷积神经网络和融合注意力机制的双向长短期记忆神经网络自适应融合预测模型（MCNN-ABiLSTM模型）。通过串联的多尺度结构增强卷积神经网络的特征提取能力，融合注意力机制的双向长短期记忆网络提升对时序特征的连续性、周期性的挖掘能力，将2个分支特征自适应融合以提升交通流量预测的准确性。同时，通过计算各路口时序流量的皮尔逊相关系数分析交通流量的空间相关性，并提出改进粒子群算法（IPSO）设置外部因素标签值。实验结果表明，MCNN-ABiLSTM模型比其他基线模型预测准确性更高，RMSE、MAE以及MAPE均有明显下降。

An adaptive fusion prediction model of bidirectional long short-term memory neural network based on multiscale convolutional neural network and fusion attention mechanism(MCNN-ABiLSTM) is proposed for the time dependency,spatial correlation,and susceptibility to external factors in traffic flow sequences. By concatenating a multi-scale structure to enhance the feature extraction capability of the convolutional neural network and integrating a bidirectional long short-term memory network with a fusion attention mechanism to improve the continuity and periodicity mining ability of temporal features,the adaptive fusion of two branch features is performed to enhance the accuracy of traffic flow prediction. Additionally,the spatial correlation of traffic flow is analyzed by calculating the Pearson correlation coefficients of time series traffic volumes at each crossing. Furthermore,an improved Particle Swarm Optimization(IPSO)algorithm is proposed to set the label values of external factors. The experimental results indicate that the MCNN-ABiLSTM model outperforms other baseline models in terms of prediction accuracy,with significant reductions in RMSE,MAE,and MAPE.