In this paper, models based on fuzzy logic and recurrent high- order neural networks are developed for forecasting the power output of a wind park. These models appear several advantages over conventional models; e.g. they permit to consider easily explanatory variables like wind speed, direction etc. Adaptive learning laws are derived for the estimation of the various parameters. The architecture of the models is optimised automatically using a non- linear optimisation technique that aims to substitute the usually applied trial-and-error method. The developed models have been tested using various time-series of wind power where they outperform persistence as well as conventional models. Finally, the paper presents issues on the on-line implementation of a forecasting model into an advanced control system for the optimal operation and management of a real autonomous wind-diesel power system.