The AnEn technique has been widely used for both meteorology and renewable energy applications. It is an effective method to generate skillful and reliable probabilistic predictions of meteorological variables, wind and solar power for short-term forecasts up to 72 hours. It is based on an historical dataset including measurements paired with corresponding deterministic predictions. For each forecast lead time and location, AnEn is created using the measurements corresponding to the past deterministic predictions that are more similar to the current forecast. Until recently the AnEn technique proposed by Delle Monache et al. (2013) has been used to generate predictions at specific locations, where observations are available. By using an analysis field as the ground-truth AnEn is extended here over a 2D grid, where each grid point is considered as a different location and treated independently.
In this novel application AnEn forecasts are generated using data from the European Centre for Medium-Range Weather Forecasts (ECMWF) as a deterministic model and the ERA-interim reanalysis dataset as the ground-truth. Also, data from the ECMWF Ensemble Prediction System (ECMWF-EPS) are used as a reference for AnEn. Data are collected over two domains centered over Northern Italy and Colorado, covering an area of about 500x500 km2 for a period of 3 years, from March 2012 to March 2015. Both ECMWF and ERA-interim datasets have a horizontal resolution of 0.125°, while ECMWF-EPS has a horizontal resolution of 0.25°. Reanalysis data cover the whole period and are available at 00, 06, 12, 18 UTC, while 0-144 hour forecast from the ECMWF deterministic model are issued at 00 UTC of each day and are collected with three hour increments. The analogs are selected over a training period defined by the first 2 years of the dataset, while the remaining available year is used for the verification.
An in-depth analysis of AnEn skill and a comparison between AnEn and ECMWF-EPS forecasts will be presented. Several important attributes of probabilistic predictions are evaluated, including statistical consistency, reliability, and the spread/skill relationship, showing the effectiveness of AnEn to produce gridded probabilistic forecasts.