Probabilistic traffic breakdown forecasting through Bayesian approximation using variational LSTMs

Abstract

Robust artificial intelligence models have been criticized for their lack of uncertainty control and inability to explain feature importance, which has limited their adoption. However, probabilistic machine learning and explainable artificial intelligence have shown great scientific and technical advances, and have slowly permeated other areas, such as Traffic Engineering. This thesis fulfils a literature gap related to probabilistic traffic breakdown forecasting. We propose a traffic breakdown probability calculation methodology based on probabilistic speed predictions. Since the probabilistic characteristic is absent in traditional formulations of neural networks, we suggest using Variational LSTMs to make the speed forecasts. This Recurrent Neural Network uses Dropout to produce a Bayesian approximation and generate probabilistic outputs. This thesis also investigates the effects of inclement weather on traffic breakdown probability and methods for identifying traffic breakdowns. The proposed methodology produces great control over the probability of congestion, which could not be achieved using deterministic models, resulting in important theoretical and practical contributions.