Highway Traffic Flow Forecasting with Multidimensional Signal Feature Decomposition and Patch Time Series Convolutional Neural Network

Accurate prediction of traffic flow is the key to highways control. However, traditional time series forecasting methods cannot meet the accuracy requirements of long-term forecasting. This paper proposes a multi-channel univariate long-term highway inbound traffic flow forecasting framework with multidimensional signal feature decomposition of time series and a patch time series depth-separable convolutional neural network. Firstly, we propose a multidimensional decomposition block consisting of a principal feature decomposition block based on the Fourier transform, a backbone and noise decomposition block based on the Stationary Wavelet Transform, a cyclic signal enhancer based on threshold comparator, and a trend extraction block based on average pooling. Secondly, we propose to change the depth-separable convolution layer mode and stack multiple depth-separable convolution layers so as to capture the developmental characteristics of the time series signal. Furthermore, a feed-forward neural network layer is set up between the depth-separable convolution layers. Then, true time series decomposition is used in the training phase to compute the multidimensional feature loss, with the aim of improving the shortcoming of the tensor decomposition that does not allow for gradient propagation. Finally, weight aggregation is used to transform the multidimensional data into univariate time series data. Experimental results on real highway inbound traffic flow datasets show that the proposed method achieves better performance than the baseline model and effectively improves the prediction accuracy.