摘要
| Figures/TablesFigures/Tables | ReferencesReferences
Summary
Bias correction of varying complexity - from simple scaling and additive corrections to more advanced histogram equalisation (HE) corrections - is applied to high resolution (7 km) regional climate model (RCM) simulations. The aim of the study is to compare different methods that are easily implemented and applied to the data, and to assess the applicability and impact of the bias correction depending on the type of bias. The model bias is determined by comparison to a new gridded high resolution (1 km) data set of temperature and precipitation, which is also used as reference for the corrections. The performance of the different methods depends on the type of bias of the model, and on the investigated statistic. Whereas simpler methods correct the first moment of the distributions, they can have adverse effects on higher moments. The HE method corrects also higher moments, but approximations of the transfer function are necessary when applying the method to other data than the calibration data. Here, an empirical transfer function with linear fits to the tails is compared to a version where the complete function is approximated by a linear fit. The latter is thus limited to corrections of the first and second moments of the distribution. While making the transfer function more generally applicable, these approximations also limit the performance of the HE method. For the current model biases, the linear approximation is found suitable for precipitation, but for temperature it is not able to correct the whole distribution. The lower performance of the linear correction is most pronounced in summer, and is likely due to a difference in skewness between the model and observational data. Further limitations of the HE method are due to the need for long time series in order to have robust distributions for calculating the transfer function. Theoretical approximations of the required length of the calibration period were performed by using different sampling sizes drawn from a known distribution. The excerise show that about 30 year long time series are needed to have reasonable accuracy for the estimation of variance, when also corrections of the annual cycle is required.
