Better Representation of Climate Change Impacts from Multi-model Ensembles Better Representation of Climate Change Impacts from Multi-model Ensembles Jamal Zaherpour et al. 1 School of Geography Zurich, October 2015
A number of models a range of outputs = uncertainty in results Dealing with uncertainty 2 Three main actions to reduce uncertainty Acquiring higher quality data Improving hydrologic models & using better mathematical techniques Applying effective techniques for better data assimilation Addressing Uncertainty in GHMs Liu, Y. Gupta, H. V.,2007
Multi-Model Combination (MMC) The idea: To combine outputs of several models (GHMs) to get a result better than those of individual models/ensemble mean Combination Approaches Shamseldin et al., 1997 Simple Ensemble Mean Weighted Ensemble Intelligent Combination Advanced
MMC Technique using Intelligent Approaches Using Evolutionary Algorithm (EA) and machine learning to ‘discover’ an optimum way of combining GHMs so they replicate observed data (GRDC) Tools: Symbolic Regression (SR) and Gene Expression Programming (GEP) to combine the multiple models Robs= f(DBH, H08, LPJmL, PCR_GLOBWB, WaterGAP2) Equations produced are validated according to their ‘fit’ with observed data (runoff) 4
MMC Technique using Intelligent Approaches 5 Example Eq: Robs = *WaterGAP *h *h08*LPJmL By analysing the SR/GEP equations we can learn how alternative GHMs perform relative to one another
GHMs, EM and MMC Evaluation Method 6
IPE (Ideal Point Error); Numerical Integrated Metric RMSE: root mean squared error MARE: mean absolute relative error CE: Coefficient of Efficiency i: ith participating model (GHMs) max (x) or min (x): the max or min value of the statistic x among the group of models If model fit to observed is perfect IPE will equal 0 The worse the fit, the further from 0 IPE will be. 7 Dawson et al., 2012 All metrics are equally important
Experimental Results 8
Sources of data Observed Discharch: The Global Runoff Data Centre (GRDC) for more than 50 major independent catchments. Monthly resolution data extending 1971 – Inputs to the MMC: ISI-MIP2.1 historic varsoc runs from 5 GHMs: 9 LPJmL PCR-GLOBWB WaterGAP2 H08 DBH
56 catchments from GRDC reference dataset Area >= 100,000 km 2 Observed record length >= 20 years 10 Catchment selection
GHMs and EM performance in simulating observed runoff GHMs and EM performance in simulating observed runoff 11 WaterGAP2: 38 DBH: 3 lpjml: 1 PCR-GLOBWB: 4 H08: 3 EM is better than the best GHM only in 6 catchments
MMC performance Compared to GHMs and EM MMC performance Compared to GHMs and EM MMC outperforms the best GHM in all catchments with an average of 47% reduction in IPE MMC outperforms the EM in ALL catchments
13 Example Catchment: Niger river, Lokoja gauge Model DBHH08LPJmLPCR_GLOBWBWaterGAP2 EM MMC IPE Rank35421
14 Thank you