Objective & Background
The most influential and least accurate variable in PV power forecasts is the solar irradiance. State-of-the-art PV power forecasts most often use deterministic forecast models of the solar irradiance. However, decision making would benefit from information about the prediction accuracy. Probabilistic forecasts provide information about the prediction accuracy. One way to obtain it is to use ensemble forecasts like the Integrated Forecast System (IFS)-ensemble of the European Centre for Medium-Range Weather Forecasts (ECMWF). However, such ensemble forecasts may lack reliability of the probabilistic information. Hence, post-processing with reference values, called calibration, is required. Here we address a location- independent calibration of the IFS solar irradiance ensemble for Germany, in order to be able to provide reliable forecasts for PV power plants without measurements available.
In this contribution we consider the solar irradiance at surface level of the IFS-ensemble and calibrate it with ground-based pyranometer measurements from 19 measurement stations of the German weather service (DWD). The aligned dataset of two years has a time resolution of 15min. As a first approach for calibration, we use variance-deficit (VD) calibration . We split the dataset by different forecast steps and calibrate the ensemble for each step separately. To determine the VD calibration coefficient we use the data of the previous month and find one common set of parameters for all stations. We analyse the dependence of the spread-skill relationship for different forecast steps and quantify the improvement by the calibration with the continuous rank probability score (CRPS). In addition, we simulate PV power values for the calibrated ensemble for six PV power plants. For the simulation, we convert the irradiance to the tilted module plane and run a parametric PV simulation model . We repeat the ensemble analysis for the PV power values and compare the results to the irradiance ensemble analysis.
From a comparison of rank histograms, we observe an increase in reliability of the VD-calibrated irradiance and PV power ensemble. The spread-skill diagram shows an overestimation of the uncertainty for small spreads, and a larger underestimation for large spreads. We find that the VD calibration lead to a reduction of the CRPS for the day-ahead forecasts from 22.7% to 21.1% (rel. -7%). For the ensemble of PV power values the CRPS reduces from 28.5% to 24.9% (rel. -12%).
The VD calibration increases the performance of the ensemble forecast both for the solar irradiance and the PV power. The spread-skill relationship indicates the potential for a future spread-depended VD calibration. Further the location-independent calibration shall be compared to a location-specific calibration.