Boiler drum is a critical power plant component having nonlinearities and a non-minimum phase response but, despite this fact, control structures with PID controllers are usually used in industrial applications. In the last decades, important efforts have been made to improve the closed loop performance of the drum level controller by using on-line self-tuning approaches. In this paper, such a method of on-line adaptation is presented based on a neuro-predictive technique of the authors of [4], [5]. In order to implement the neuro-predictive procedure a neural model was developed for drum level process. By means of simulation, a comparison between the obtained drum level controller's operation and that of a classic controller of the type PI under similar conditions is carried out.