A Comparison of Statistical Methods to Generate Short-Term Probabilistic Forecasts for Wind Power Production Purposes in Iceland

Sammanfattning: Accurate forecasts of wind speed and power production are of great value for wind power producers. In Southwest Iceland, wind power installations are being planned by various entities. This study aims to create optimal wind speed and wind power production forecasts for wind power production in Southwest Iceland by applying statistical post-processing methods to a deterministic HARMONIE-AROME forecast at a single point in space. Three such methods were implemented for a 22 month-long set of forecast-observation samples in 1h resolution: Temporal Smoothing (TS), Observational Distributions on Discrete Intervals (ODDI - a relatively simple classification algorithm) and Quantile Regression Forest (QRF - a relatively complicated Machine Learning Algorithm). Wind power forecasts were derived directly from forecasts of wind speed using an idealized power curve. Four different metrics were given equal weight in the evaluation of the methods: Root Mean Square Error (RMSE), Miss Rate of the 95-percent forecast interval (MR95), Mean Median Forecast Interval Width (MMFIW - a metric to measure the forecast sharpness) and Continuous Ranked Probability Score (CRPS). Of the three methods, TS performed inadequately while ODDI and QRF performed significantly better, and similarly to each other. Both ODDI and QRF predict wind speed and power production slightly more accurately than deterministic AROME in terms of their Root Mean Square Error. In addition to an overall evaluation of all three methods, ODDI and QRF were evaluated conditionally. The results indicate that QRF performs significantly better  than ODDI at forecasting wind speed and wind power at wind speeds above 13 m/s. Else, no strong discrepancies were found between their conditional performance. The results of this study are limited by a relatively scarce data set and correspondingly short time series. The results indicate that applying statistical post-processing methods of varying complexity to deterministic wind speed forecasts is a viable approach to gaining a probabilistic insight into the wind power potential at a given location.