Model-Based Reinforcement Learning for Control of Strongly-Disturbed Unsteady Aerodynamic Flows