Automated vehicles connected through vehicle-tovehicle communications can use onboard sensor information from adjacent vehicles to provide higher traffic safety or passenger comfort. In particular, automated vehicles forming a platoon can enhance traffic safety by communicating before braking hard. It can also improve fuel efficiency by enabling reduced aerodynamic drag through short gaps. However, packet losses may increase the delay between periodic beacons, especially for the rear vehicles in a platoon. If the connected vehicles can forecast link quality, they can assign different performance levels in terms of intervehicle distances and also facilitate the designing of safer braking strategies. This paper proposes a strategy for incorporating machine learning algorithms into, e.g., the lead vehicle of a platoon to enable online training and real-time prediction of communication delays incurred by connected vehicles during runtime. The prediction accuracy and its suitability for making safety-critical decisions during, e.g., emergency braking have been evaluated through rigorous simulations.