Reducing ecological footprint has become a priority for most industries including the railway industry which is one of the most energy-consuming industry. In that prospect, the optimisation of the driver’s control to reduce the consumed energy while respecting multiple constraints (punctuality, security, etc.) is a particularly important challenge. Several proposals for optimising these control functions can therefore be found in the literature, mainly based on offline methods to guide the driver toward optimality. Continuing these works, this presentation focuses on an online methodology to make use of the knowledge brought by on-board sensors. Focusing on a particular type of train, and for a particular track, this work breaks down into four phases: the implementation of a simplified parametric model of the train, the estimation of the parameters of this model by Bayesian inference from dedicated online measurements, the offline approximation of the manifold characterizing the set of optimal controls associated with the set of likely values of the model parameters, and then an online method for updating the true controls of the true train as it runs, by combining the information provided by this manifold with the position and speed measurements recorded in real time. Finally, the efficiency and the robustness of the method are numerically validated using nominal and non-nominal context.