Statistical downscaling of General Circulation Model outputs to catchment streamflows

Abstract

Since the latter half of the 20th century, many regions of Australia experienced a drop in average rainfall, causing low inflows to reservoirs. Until the recent heavy rainfalls of late 2010 and early 2011, Victoria suffered a severe drought commencing 1997. This resulted in a reduction of annual average inflows to Melbourne's main water supply reservoirs of about 38%, during the period 1997-2008. The Grampians Wimmera Mallee Water (GWMWater) supply system in north-western Victoria also experienced a drop in annual inflows to its reservoirs of about 75%, from the long term average since 1997. Already being the driest inhabited continent in the world, this drop in inflows to reservoirs was of significant concern to water managers across much of Australia. Such a significant deviation from the long term average highlights the importance of being able to reliably predict streamflows considering the likely future climate change and variability, which will ultimately aid in future planning of the water supply systems. General Circulation Models (GCMs) are the most advanced tools available for the simulation of future climate. However, the coarse spatial resolution of GCMs does not allow for hydroclimatic predictions at the catchment scale. Indeed, they are incapable of producing outputs at the fine spatial resolution needed for most hydrological studies. To address this issue, downscaling methods have been developed, which link coarse resolution GCM outputs to surface hydroclimatic variables at finer resolutions. Downscaling techniques are broadly classified as either dynamic or statistical. The computation cost associated with dynamic downscaling methods is much higher than that of statistical downscaling. Another major drawback of dynamic downscaling models is their high complexity. The aim of the present study was to develop a model capable of statistically downscaling monthly GCM outputs to catchment scale monthly streamflows, accounting for the climate change. The current study investigated only the calibration and validation of the abovementioned statistical downscaling model. This was demonstrated through a case study applied to the GWMWater supply system in north-western Victoria, Australia. It is a large scale complex multi-reservoir system that is operated to meet a range of economic, social, and environmental interests. Support Vector Machine (SVM), a statistical downscaling technique, was used in the current streamflow downscaling exercise. The selection of SVM for downscaling was based on its better capability in capturing complex non-linear relationships between GCM outputs and catchment level variables, than artificial neural networks (ANN) and multi-linear regression (MLR), as observed in the past studies. National Center for Environmental Predictions/National Center for Atmospheric Research (NCEP/NCAR) reanalysis data and observed streamflow data, over the study area, were used for the calibration and verification of the downscaling models. The model calibration (1950-1989) and validation (1990-2010) were performed on each calendar month separately and later results were aggregated to produce the time series of prediction. It was found that, the model was able to produce better predictions over the summer and winter months than in autumn and spring. The model tended to over predict the peaks of streamflows particularly after the 1997 drought in Victoria. It was further observed that the NCEP/NCAR reanalysis variables used in the study did not show a clear change corresponding to the drop in streamflow observed after 1997. The problems associated with the method over the recent severe drought have revealed important information to enable improvements for future model work. Downscaling streamflows from the GCMs skips complex hydrologic modelling, saves time and effort in predicting streamflows. Also, the current work in downscaling streamflows from GCM outputs is believed to be the first in Australia. The present research employed downscaling models based on the 12 calendar months enabling a better capture of streamflow characteristics, unlike the models based on seasons used in the past studies.