Analysis of the statistical bias correction of ERA5-Land on different time aggregations in Trentino-Alto Adige

Global and regional climate models are constantly improving the quality of their outputs with increasingly fine spatial and temporal resolutions. These products, which comprise, for instance, reanalysis, reforecast and forecast, can be used for several applications, such as boundary conditions for climate simulations, initial conditions for local weather forecasting, and reference datasets for environmental and energy uses. Nevertheless, many authors have pointed out that such climate models are not suitable for direct use in local applications due to the presence of biases between the model results and the metered data. At this aim, several statistical methodologies have been proposed to correct and downscale the climate models outputs and make it available also for local purposes. Therefore, the purpose of this contribution is to analyse the current state-of-the-art statistical bias correction methods on different time aggregation to assess the capabilities of these methods from monthly to hourly temporal scale.

This study is carried out on the Trentino- Alto Adige, which is an alpine region in north Italy equipped with several measuring weather stations, around 300. The temperature and precipitation observations have been then used to produce a reference dataset through the geostatistical interpolation method called kriging. Instead, ERA5-Land, the reanalysis of ECMWF, has been adopted for the bias correction analysis. Several methods have been tested comprising of univariate and multivariate method including: linear scaling, variance scaling, local intensity scaling, local power transformation, quantile mapping, quantile delta mapping, and multivariate bias correction methods such as MBCn, MBCp, and MBCr. The time scale investigated are monthly, daily and hourly aggregations.

The results show a general decreasing of the performance of all the bias correction methods with the increase in the time-frequency of the weather variables. In particular, the mean absolute error of the corrected daily temperature is 50% larger than the monthly one, and the same 50% increase in error is found between daily and hourly corrected data. The increase in error with decreasing temporal resolution is even more pronounced for the precipitation variable, which is known to be discontinuous with respect to temperature. Multivariate bias correction methods seem to have difficulty maintaining dependencies between variables in the case of high-frequency data.

Although the results on the hourly data are not so scarce, it is evident that more depth analysis of temporal high-resolution climate data is needed, including sub-hourly data in the future, and therefore become crucial to develop new methodologies capable of correcting sub-daily bias. In conclusion, with this work, the authors seek to support research in the direction of providing high-frequency weather data for local applications, which are crucial, for example, in hydrological simulations for the assessment of hydrogeological risks and the management of renewable energy in the electricity market.