Transferability improvement in short-term traffic prediction using stacked LSTM network

Abstract Short-term traffic flow forecasting is a key element in Intelligent Transport Systems (ITS) to provide proactive state information road network operators. A variety of methods predict variables the short-term can be found literature, ranging from time-series algorithms, machine learning tools and deep selective hybrid these approaches. Despite advances prediction techniques, challenging problem that affects application such real world prevalence insufficient data across an entire network. It rare extensive historical training required for model are available all links city. In order address this insufficiency problem, paper applies transfer techniques prediction. All used were collected Highways England networks UK. The results show through improving transferability learning-based models, computational burden due process significantly reduced accuracy under deficient scenarios improved one-step ahead However, gradually decreases multi-step also proposed method highly dependent upon consistency between datasets but less on geographical attributes links.