Real Time Headway Predictions in Urban Rail Systems and Implications for Service Control: A Deep Learning Approach