Short-term wind information is important for operating small wind plants and for planning other distributed energy resources. In this study, we work with one year of measurements from the Termas station and build a simple, reproducible workflow that goes from raw data to ten-minute wind speed forecasts. The original file contains 193,493 records with non-uniform timestamps and some suspect values near zero. We first standardise the timestamps, compute 10-minute averages and select only complete days with 144 samples, obtaining 52 days and 7,488 valid points. Days with extremely low daily means are removed and very small speeds are corrected to reduce obvious sensor and logging errors.

The cleaned series is then described with daily and global Weibull distributions. For the global fit, we obtain a shape parameter k = 1.27 and a scale parameter c = 1.40 m/s, which point to a low to moderate wind regime with frequent calm periods. For the forecasting step, we form supervised samples using 24-hour sliding windows (144 time steps) to predict the next 10-minute value. A single-layer SimpleRNN network is trained using a chronological split into training, validation, test and a final 10% hold-out set.

On the hold-out set, the network reaches a mean absolute error of 0.28 m/s and an R² of 0.78. These results suggest that even a modest recurrent model can follow the short-term variability of the wind speed at this site and that the proposed workflow can be reused with longer records or other stations.