Conform with the Wind : Processing short-term ensemble forecasts with conformal based methods for probabilistic wind-speed forecasting

Abstract: Forecasting wind has always been an interesting subject, and as large parts of the world are relying more on wind for power production it is becoming even more important to have reliable forecasts. Probabilistic forecasts, where distributions are predicted in contrast to deterministic forecasts, are impor- tant for informed decision making. We apply two methods based on conformal prediction for processing ensemble forecasts to well calibrated probability dis- tributions. Conformal prediction is a relatively modern method for quantified uncertainty analysis within machine learning. These methods are compared to the quantile regression forest algorithm, which has been well tested in literature for probabilistic ensemble post processing. Ensemble forecasting is a method based on running several numerical weather prediction models simultaneously, creating an array of forecasts. The conformal methods rely on an additional point forecast, supplied by another model, for producing the distributions while the quantile regression forest works directly on the ensemble. The methods were tested using a teaching schedule which determines the best configuration of parameters, from a predefined set, based on the continuous ranked proba- bility score metric before making each prediction. This is a way of simulating how the methods perform over time. For the conformal methods we employ a normalized version of conformal predictive distribution systems and a non- exchangeable conformal prediction method. For the non-exchangeable case we suggest a method of stacking confidence intervals to produce distributions. We also suggest a normalized version of this algorithm. Both methods show promis- ing results, both able to produce significantly better distributions than the raw ensemble and as good or better calibrated distributions compared to the quantile regression forest. Though the conformal methods are supplied external forecasts and the quantile regression forest is not using an optimal configuration. The conformal methods also produce well calibrated predictions consistently over different setups of the algorithms.