ARDE-N-BEATS: An Evolutionary Deep Learning Framework for Urban Traffic Flow Prediction
Xiaocai Zhang, Zhixun Zhao, Jinyan Li
Abstract
Accurate and reliable traffic flow prediction is difficult due to the highly nonlinear, complex, and stochastic natures of urban traffic flow data, but its solutions are critically important for intelligent transportation systems (ITSs) and Internet of Things (IoT). In this study, a novel deep learning framework, named adaptive reinitialized differential evolution (ARDE)-neural basis expansion analysis for time-series forecasting (N-BEATS), is proposed to address this challenge. With the framework of ARDE-N-BEATS, first, an N-BEATS-based deep learning architecture is formulated for modeling traffic flow data. Second, a novel enhanced evolutionary algorithm, termed ARDE, is presented for optimizing the hyperparameter and structure of N-BEATS. Compared to the vanilla differential evolution (DE) algorithm, ARDE exhibits faster convergence and stronger searching capabilities. Experiments on three real-world traffic flow data sets from Dublin and San Francisco demonstrate that ARDE-N-BEATS can achieve high accuracy of at least 94% for most of the predictions, and outperforms the existing counterpart methods. A comparison between different hyperparameter optimization approaches further reveals that ARDE provides better or very competitive predictions and saves as high as 78.90% of computational expense.