We develop a deep learning model to predict traffic flows. The maincontribution is development of an architecture that combines a linear modelthat is fitted using $\ell_1$ regularization and a sequence of $\tanh$ layers.The challenge of predicting traffic flows are the sharp nonlinearities due totransitions between free flow, breakdown, recovery and congestion. We show thatdeep learning architectures can capture these nonlinear spatio-temporaleffects. The first layer identifies spatio-temporal relations among predictorsand other layers model nonlinear relations. We illustrate our methodology onroad sensor data from Interstate I-55 and predict traffic flows during twospecial events; a Chicago Bears football game and an extreme snowstorm event.Both cases have sharp traffic flow regime changes, occurring very suddenly, andwe show how deep learning provides precise short term traffic flow predictions.
展开▼
