Unsupervised Machine Learning to Quantify Climate-Model Uncertainty for Resilient Water Resources Planning

Future climate change projections are characterised by uncertainties associated with Global Climate Models (GCMs) and emission scenario (SSPs). Different GCMs and SSPs represent key climate processes differently, yielding divergent projections rather than a single “best” future. In turn, this propagates into decision uncertainty for long-term water-resources management and planning. Climate model uncertainty analysis therefore provides a structured framework to identify, quantify, decompose, and communicate these uncertainties in water resource modelling. This helps bound plausible futures by emphasizing ranges of outcomes rather than single point estimates. This study develops an integrated framework that leverages unsupervised machine learning to characterize and quantify climate-model uncertainty for long-term water-resources management and planning. The framework integrates ranking, clustering, and scenario-discovery methods. We analyze outputs from 24 climate models from the Coupled Model Intercomparison Project Phase 6 (CMIP6) alongside observed reanalysis from the Princeton dataset. Monthly precipitation and temperature are evaluated across multiple locations within the basin to account for spatial heterogeneity. Model ranking was performed by evaluating each climate model against the observed reanalysis dataset. Performance was assessed using mean bias and percent bias, along with metrics capturing seasonality, spatial patterns, and interannual variability for basin-scale monthly temperature and precipitation. For each GCM, engineered features describing annual and seasonal change were then used for clustering. Unsupervised grouping was followed by classification based on Bayes decision theory. Within each cluster, a representative medoid was identified by minimizing the sum of Euclidean distances to all other members, yielding the most central model in that group. Cluster labels Low, Normal, and High projection were assigned by computing the percent change in simulated mean streamflow from the hydrological simulations for each climate model. Results indicate that the representative medoids are GISS-E2-1-G (Low projection), CanESM5 (Normal projection), and EC-Earth3 (Wet projection). The remaining GCMs are then probabilistically assigned to clusters with reference to these central medoids. The framework is demonstrated for the Blue Nile Basin to support long-term water-resources planning under climate uncertainty. This study extends the application of unsupervised machine learning for characterizing and quantifying climate-model uncertainty, with the objective of resilient water resource planning across multiple, dynamically evolving future possibilities.