Improvements of statistical downscaling methods and evaluation of their contributions to the uncertainty of hydrologic impacts in a changing climate

Abstract

The most important impacts of climate change will likely be linked to water resources. Hydropower companies throughout the world increasingly realize that they must deal with future climate change. To evaluate future impacts, realistic climate projections that encompass the uncertainty linked to climate change are needed. Given the relatively large biases of General Circulation Model (GCM) outputs, particularly for precipitation and to a lesser extent for temperatures at the regional scale, it is necessary to perform some postprocessing to improve these global-scale models for hydrologic and water resource management studies. The two most commonly used approaches, dynamical and statistical downscaling, each have significant advantages and drawbacks. It is not a simple task to select one over the other. This work aims at coupling global and regional climate models and statistical downscaling into a new hybrid method by merging stochastic weather generators with climate models that quantify the hydrological impacts of climate change for a Canadian river basin. The performances of stochastic weather generators were first improved. A statistical downscaling method combining attributes of both stochastic weather generator and change factor (CF) methods was then developed. Several aspects of statistical downscaling were also evaluated. Moreover, global uncertainty and the downscaling uncertainty were outlined in quantifying the hydrological impacts of climate change. A spectral correction method and integration scheme resulted in a weather generator that can accurately produce the low-frequency variability of precipitation and temperatures, as well as the auto- and cross-correlations of and between maximum and minimum temperature (Tmax and Tmin). A large number of atmospheric predictors were used to assess the ability of statistical methods to downscale precipitation to the station scale. The downscaling of daily precipitation occurrence was mostly unsuccessful with both linear regression methods and using discriminant analysis, even though the latter was much better. Explained variances were very low for regression-based downscaling of precipitation, although results were consistently improved as the climate model resolution was made progressively finer. Even when going to the 15-km resolution Canadian Reginal Climate Model (CRCM), the predictors still explained less than 50% of the total site precipitation variance. Despite the added complexity, the weather typing approach was not much better at downscaling precipitation than the approaches without classification. The weather generator was used as a downscaling tool to downscale outputs of the CRCM (45km scale) to catchment scale. Its performance was further compared with the CF method for quantifying the hydrological impacts of climate change. Both downscaling methods suggested increases in annual and seasonal discharges for the 2025-2084 period. The weather generator-based method predicts more increase in spring (AMJ) discharge, as well as smaller increases in summer-autumn (JASON) and winter (DJFM) discharges than the CF method. Moreover, both methods indicated increases in mean annual and seasonal low flows, while there are considerable differences between their predictions. All downscaling methods including dynamical and statistical approaches suggested general increases in winter discharge (November - April) and decreases in summer discharge for the 2071-2099 horizon. Winter flows would be especially large for regression-based methods, which also predicted the largest temperature increases in autumn and winter. Peak discharges would appear earlier for all downscaling methods, but their timing varies according to the downscaling method. A GCM was consistently a major uncertainty contributor when quantifying the hydrological impacts of climate change. However, other sources of uncertainty such as the choice of downscaling method and natural variability, as represented by GCM ensemble runs, also had a comparable and even larger uncertainty affect depending on the criteria. For example, the downscaling method was the largest source of uncertainty with respect to spring discharge magnitude, annual low flow and peak discharge; while GCM initial conditions (which were a member of the ensemble runs) dominated the uncertainty for the time to peak discharge and the time to the end of flood. Uncertainties linked to greenhouse gase emission scenarios (GGES) and hydrological model structure also played an important role in hydrological predictions, but these were somewhat less than those related to GCMs and the downscaling method. Uncertainties due to the hydrological model parameters had less impact than those of the other five sources. Overall, combining Regional Climate Models (RCMs) and statistical downscaling in a unified approach appeared to have significant advantages in quantifying the hydrological impacts of climate change. Any management and adaptation of water resource systems should consider the effects of future climate change, as well as all sources of uncertainty.

Date	28 Jul 2011
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	François Brissette (Supervisor)