Probabilistic Deep Learning for Traffic Density Prediction

Abstract

Real time, accurate predictions of recurrent and nonrecurrent traffic congestion are essential for optimizing transportation systems and ensuring a smooth user experience. Traditional models often focus on long-term point estimates, limiting their use in scenarios requiring short-term predictions or probabilistic assessments (e.g., traffic signal optimization, dynamic tolling, and emergency response). This study explores probabilistic deep learning (DL) for real time traffic density distribution prediction. This study demonstrates that an adapted multi-quantile recurrent neural network (MQRNN), termed MQRNN-monotonic, outperforms traditional time series methods, particularly when handling nonrecurrent disruptions. A novel loss function is introduced to address quantile crossing issues, ensuring valid distributional predictions. Experiments on two highway data sets show that probabilistic DL for traffic density prediction yields well-calibrated and sharp dynamic traffic congestion distributions. This study offers a promising new approach to real time traffic density forecasting, paving the way for transportation systems that respond faster and smarter to changing road conditions, making traffic smoother, more sustainable, and more predictable for everyone.