A Generic Framework to Produce Probabilistic Rainfall Scenarios Incorporating Climate Change : Application to: Sustainable Urban Stormwater Management Strategies

The present work addresses the problematic of forecasting impacts of climate change on future rainfall regimes and their consequences on urban stormwater infrastructures. Researches carried out allowed to develop an integrated framework for producing high resolution probabilistic rainfall projections suitable for studying hydrological processes at the scale of urban drainage. Downscaling is at the core of the methodology as the predictions of the numerical General Circulation Models (GCMs) employed by the climate scientific community to model climate evolution are too coarse for hydrological impact studies. The proposed downscaling approach respects the scales of the physical processes characterizing precipitations and consists in three steps: i) Daily rainfall series at the location of interest are downscaled from coarse-gridded monthly GCMs projections (scale of weather events); ii) The generated daily series are further downscaled to the hourly time-step (scale of storms dynamics); iii) Finally, hourly series are disaggregated to sub-hourly level (scale of raincells). Daily downscaling is achieved by a statistical procedure, based on Generalized Linear Models (GLMs), seeking to relate large-scale atmospheric variables, corresponding to the scale of GCMs, to local daily rainfall series. The proposed methodology is assessed using three contrasted situations in Switzerland (Geneva, Sion and Säntis) and is shown to perform well in reproducing historical rainfall statistics (including extremes and inter-annual variability) in the present-day climate; furthermore, projections were shown to be consistent with the simulations of physically-based dynamical models (i.e. Regional Climate Models). Projections for the second part of the 21th century indicate considerably drier summers, but no significant tendency toward more extreme events was detected except for Säntis. Finally, extensions of the methodology were presented allowing to downscale other atmospheric variables than rainfall. Sub-daily rainfall downscaling is achieved using a stochastic hourly rainfall generator based on Poisson clusters model which aims at conceptualizing storm dynamics in a simple way. To provide sensible results such generators have to be fitted on historical rainfall statistics computed at different levels of temporal aggregation. In the present context, this raises a fundamental problem as the required fitting statistics at the sub-daily time-scale are not available for the future. Shortcomings of existing methods led us to develop a novel approach based on Multivariate Adaptive Regression Splines (MARS) which were so far seldom used in hydrology. The proposed MARS models are conditioned on climate and fit thus particularly well in the general downscaling framework. In addition, atmospheric predictors allow to account naturally for seasonal variations meaning that a single MARS model holds for the whole year, whereas existing models are specific to each month of the year a