Objective & Background.
Intrinsic variability and climate change pose serious challenges to the energy sector, which requires a growing amount of climate information at all time and spatial scales to put in action effective adaptation strategies. In this context, climate projections are a crucial tool to inform energy companies about the future state of the climate. However, when comparing projections to observed data, it becomes evident that the simplifications contained in climate models induce biases in the simulated climate variables. For this reason, bias correction (BC) is a necessary step to adjust model outputs. We assess the performance of Cumulative Density Function transform (CDF-t) BC, consisting of a remapping of the entire CDF, instead of the mean and variance only. As a main drawback, this method does not consider temporal dependence explicitly, and it is performed on each grid point separately, assuming spatial independence. We aim at assessing the capability of CDF-t BC to preserve temporal correlations at different time horizons, including seasonality and inter-annual variability. Moreover, we investigate whether neglecting spatial dependence introduces spurious spatial correlations in the corrected data.
We consider a dataset consisting of eleven EURO-CORDEX 3-hourly simulations at a spatial resolution of 0.25° over the period 1951-2100 for Europe. We assess the BC performance on four relevant variables for the energy sector, i.e. near-surface temperature, precipitation, wind speed and surface solar radiation. Projections are compared to observations provided in the E-OBS v18.0e dataset for the period 1951-2017. To measure the BC performance, we quantify the displacement of both raw and adjusted projections from the observations in terms of: trends; autocorrelation (lags equal to 1, 7, 28, 183, 365 days); global (spatial correlogram, global Moran's I) and local (local Moran's I) measures of spatial correlation. The latter is estimated for selected percentiles of the difference between raw and adjusted detrended projections, divided by season. In a second step, the methodology may be extended to the assessment of cross-variable correlations.
Currently available results concern near-surface temperature projections from the WRF381P model forced by the IPSL-CM5A-MR under the RCP 8.5 scenario. Regarding trends, differences from observations appear to be negligible for both raw and adjusted data. On the other hand, differences in autocorrelation are more important, especially at seasonal and inter-annual lags, but sensibly decreased by BC. Finally, spatial autocorrelation is significant, with global Moran's I higher than 0.90 for all cases and slowly decaying correlograms; both spatial correlograms and local Moran's I show different patterns depending on the selected season and percentile.
Preliminary results indicate that some dynamical properties of climate projections are affected by BC, and that the assumption of spatial independence is not realistic, introducing spurious spatial correlations. This may suggest the need to improve BC to preserve not only the CDF, but also temporal and spatial dependence present in climate variables.