Traffic Forecasting via Dilated Temporal Convolution With Peak-Sensitive Loss
Traffic Forecasting via Dilated Temporal Convolution With Peak-Sensitive Loss
复制标题
DOI：
10.1109/mits.2021.3119869
复制
发表时间：
2023-01
影响因子：
3.6
通讯作者：
Ge Guo;W. Yuan;Jinyuan Liu;Yisheng Lv;Wei Liu
中科院分区：
文献类型：
--
作者：
Ge Guo;W. Yuan;Jinyuan Liu;Yisheng Lv;Wei Liu
基于深度学习的交通预测方法可以捕捉交通数据中复杂的时空特征和环境因素。然而，它们在少数峰周围的性能不令人满意，并且对于大范围空间相关性的建模效率不高。本文给出了一个峰值感知深度学习架构，用于交通预测，该架构涉及一个称为峰值敏感损失的成本敏感损失函数。由于对均方损失和平均绝对百分比损失的平方等常用度量采用了不同的代价，因此该方法可以提高性能。构建了一种基于扩展卷积网络（DCN）和时间卷积网络（TCN）的时空卷积体系结构，通过DCN对空间特征（宽范围和短范围）建模，通过TCN学习时间特征。用实际数据集验证了该模型的有效性。
Deep learning-based traffic forecasting methods can capture intricate spatiotemporal features in traffic data and environmental factors. However, they have unsatisfactory performance around the minority peaks and are inefficient for modeling wide-range spatial correlations. This article gives a peak-aware deep learning architecture for traffic forecasting by involving a cost-sensitive loss function called peak-sensitive loss. This method can improve the performance since different costs are employed on the prevalent metrics such as mean-square loss and square of mean absolute percentage loss. A spatiotemporal convolutional architecture based on a dilated convolutional network (DCN) and a temporal convolutional network (TCN) is constructed that models the spatial features (both wide and short range) by the DCN and learns the time characteristics by the TCN. The effectiveness of the model is demonstrated with real-world data sets.