Producing accurate and reliable wind forecasts can be very helpful during offshore navigation, both for on board safety and for taking tactical decisions during sailing yacht races. In this work different models based on artificial neural networks are used for short-term wind forecasts. A time-series approach is used, and the forecast is based on past values of wind velocity. The wind velocities measured at previous instants are input of the proposed algorithm, which predicts the velocities for future instants. The peculiarity of this method is that no other physical values are needed to obtain the forecast. A computer program was implemented and tested on different time series: daily-averages, ten-minutes and threesecond measurements. Using daily-averaged data, the algorithm was able to accurately forecast when the wind was going to increase or decrease for the following day. Ten-minute data allowed forecasting up to 4 steps ahead with good accuracy, while three-second data allowed accurately forecast up to 20 steps ahead, subjected to an adequate training of the network. Also, updating the network with real-time velocity measurements, an iterative algorithm was achieved allowing a continuous forecast.