Exploring the role of wavelet decomposition order in deep learning-based network-wide traffic prediction

The performance level of traffic control systems highly relies on accurate short-term network-wide traffic prediction. Signal processing techniques have been widely integrated with deep learning algorithms for this purpose, but no study has focused on how the parameter of wavelet decomposition order (level) affects the prediction robustness given the training data limitations. This study proposes a hybrid framework for short-term network-wide traffic state prediction, which applies Multi-Layer Perceptron (MLP) neural networks to implicitly capture the traffic network co-movement patterns. Then, the framework is complemented by a Seasonal Auto-Regressive Integrated Moving Average (SARIMA) model, extracting the location-specific features, including the local seasonality and stochastic disturbances. Besides, the hybrid framework is used to explore the association between the traffic prediction accuracy and wavelet decomposition order using an order-adaptive Discrete Haar Wavelet Transform (DHWT). The proposed method was validated over four open-access datasets with different training data characteristics in Paris and Madrid urban areas. The results indicated that the hybrid framework significantly improved the predictive accuracy of the benchmark deep learning algorithms. The forecasts for the low-resolution dataset experienced a noticeable improvement once higher wavelet orders were applied. The statistical analysis revealed the moderating effects of the training sample size and data spatial resolution on the link between the wavelet decomposition orders and the model predictive performance. Nevertheless, a combination of pre-processing order of two, and post-processing order of three led to satisfactory results in the cases where sufficient historical data were available.