This paper presents a neural network approach for controlling train dynamics within a digital twin, addressing the challenge of creating an efficient training environment for a reinforcement learning (RL) agent. Our method uses high-fidelity longitudinal dynamics equations to model train movement, accounting for track gradients and speed restrictions. We formalize the problem as a Markov Decision Process (MDP) and investigate nine environment variants to identify optimal observation spaces and reward functions. Experiments using the Proximal Policy Optimization (PPO) algorithm show that using a normalized velocity error for the observation space yields superior performance. This work contributes a robust framework for developing and evaluating RL-based control systems for intelligent railway traffic management.