Abstract. Wind turbines and wind farms lack information about upstream wind conditions which are ultimately converted into electricity. Remote sensing instruments such as compact pulsed scanning wind lidars can observe the incoming wind field at large distances (up to 10 km) ahead of a wind farm and provide spatial and temporal information about the inflow on operational timeframes not feasible with numerical weather models. On very-short horizons (below 1-hour lead times), the persistence method is commonly used, which fails to capture the unsteady state of the atmosphere and can introduce costly errors into the power system by means of imbalances.

A method of measuring, processing, and predicting site-specific 1–60 minute ahead wind speeds is proposed using machine learning methods applied to lidar observations from a field experiment in western Denmark. A direct multi-step forecast strategy is implemented using Stochastic Gradient Descent Regression (SGDR) with model weights updated following each repeating lidar scan. Overall, the proposed method demonstrates improved skill over persistence, with a reduction of root-mean-squared (RMS) wind speed errors ranging from 21 % (1-min ahead), to 10.9 % (5-mins ahead), 9.2 % (10-mins ahead), 7.1 % (30-mins ahead), and 6.2 % (60-mins ahead) while maintaining normally distributed errors.