Region-wide congestion prediction and control using deep learning

Abstract Traffic congestion is forecast for neighborhoods within a region using a deep learning model. The model is based on Long Short-Term Memory (LSTM) neural network architecture. It forecasts a congestion score, defined as the ratio of the vehicle accumulation inside a neighborhood to its trip completion rate. Inputs include congestion scores measured at earlier times in neighborhoods within a region, and three other real-time measures of regional traffic. The ideas are tested using Newell’s simplified theory of kinematic waves. Simplified street networks are featured first. Initial tests demonstrate the suitability of the congestion score for characterizing neighborhood traffic conditions, and that the score can be predicted using the four inputs. Further tests of the simplified networks illustrate the value of the deep learning approach, as compared against the use of three benchmark models. A next round of tests shows that the model can be made robust, even to adverse settings. A final round of tests features a pared-down version of the freeway network in the San Francisco Bay Area. The final tests show that the model is scalable. The model is thereafter improved by representing the inputs through weighted undirected graphs that incorporate the route-choice of individuals, and learning features through graph convolutions. A framework for better interpreting the contributions of the model’s inputs to its output is developed. A demonstration of the model’s usefulness in designing traffic control schemes is presented as well.