ECMWF Ensemble Prediction System (EPS) is an ensemble model specifically designed for medium range predictions. As a result, it is generally underdispersive for the very short and short range, making raw ECMWF EPS forecasts unreliable. These forecasts need to be calibrated if they are to be used effectively in this range. Different methods have been developed in the past to achieve this, specifically tailored for each output parameter. In particular, this issue also affects the direct normal irradiance (DNI), a key parameter used to predict the performance of concentrating solar power (CSP) plants.
A quantile regression method has been used to calibrate the DNI, based on the work done by (Beaullegue, 2016) for a different parameter, the global horizontal irradiance (GHI). A simple regression has been applied, to interpret better the results and check the dependency with the input variable, though some tests using regularization techniques have been done too. Rolling windows of 30 and 60 days have been used for the training. The method has been verified comparing short range hourly forecasts against observations from a station in Badajoz (Spain), for a three year period (from June 2015 to May 2018). The benchmarks used in this study are the 1-day persistence, a 60 member perfect ensemble, composed by the DNI measured for each minute in the hour, and the raw ECMWF forecasts.
DNI and GHI have a similar behaviour and statistical properties, but raw DNI forecasts are even more underdispersive than GHI ones. Calibrated DNI forecasts correct largely this problem, though not completely at dawn and dusk. They improve the Continuous Ranked Probability Score (CRPS) by a 20% compared with raw DNI forecasts. On the other hand, in a few cases calibration produces incoherent quantile crossings, and a rearrangement of the quantiles can be necessary.
Higher powers of the DNI have been also considered as predictands to model possible nonlinear dependencies. It has been found that despite not affecting significantly the summary scores, these predictands can lead to an abnormal behaviour of the calibrated forecasts in some specific days, which can negatively impact operational systems.
Quantile regression has shown to be a good method to calibrate DNI ensemble forecasts for the short range, improving significantly the reliability and accuracy of raw ECMWF forecasts. It can be used in combination with regularization methods, but abnormal behaviour can happen occasionally for some predictands, hence they should be chosen carefully.
Ben Bouallègue, Z. (2016). Statistical postprocessing of ensemble global radiation forecasts with penalized quantile regression. Meteorologische Zeitschrift, 26, 253 – 264.