The optimization of train speed with respect to energy criteria is a major ecological and economic challenge for the railway sector. This optimization is however complex due to the constraints of comfort and punctuality to be respected as well as the presence of uncertainties in the system (wind, humidity which modifies the characteristics of the wheel/rail contact, yield parameters...) This work presents an original method based on two simplified physical models, Bayesian inference of the parameters from on-line measurements, the definition of the optimization problem integrating the constraints and uncertainties and finally its resolution using an evolutionary algorithm. Finally, an application to a portion of a high speed line is proposed.