Forecasting intermittent wind energy sources presents significant challenges in the operation of the power systems. Numerical weather prediction models do not explicitly provide turbine-hub wind speed forecasts and tend to exhibit significant biases complicating their integration into energy modelling systems. Some degree of statistical calibration is often desirable to address both systematic model errors and derive risk information related to wind ramp events, large and rapid variations in wind power production causing disruptions across electricity markets.

In this study, we compare several statistical post-processing methods aiming to improve on the accuracy of wind speed and wind power forecasts and to derive risk of wind ramp events from a range of global and regional numerical weather prediction models for up to 48 hours ahead. We demonstrate application of machine learning methods for statistical enhancement and tuning of numerical forecasts against the observed wind generation records as well as an analogue approach for refining the training sample and deriving risk scenarios. Results for select wind farms in Southern and Western Australia with a focus on significant ramping events over 2018-2019 period are presented.