Adaptive Event-Driven Regime Modelling of Non-Stationary Rainfall Under Climate Variability

The variability of rainfall under the changing climatic conditions becomes a big challenge to the hydrological prediction, water resource management, and disaster risk reduction because of its intermittent and unstable nature. This paper suggests Adaptive Event-Driven Regime Transition Model (AERTM) in the context of a Regime-Conditioned Stochastic Event Dynamics Modelling (RC-SEDM) model to address the shortcomings of conventional continuous and fixed transition rainfall models. The model models rainfall using discrete stochastic events with duration, intensity and dry spells between them but latent rainfall regimes are automatically determined without any seasonal assumptions. The adaptively conditioned regime transitions are based on the features of the prior rainfall events, which makes it possible to simulate persistence, clustering, and regime transitions in a realistic way. The proposed framework has been tested with the long-term data of the daily rainfalls in Pune, India and compared with the Gaussian Mixture Models, Hidden Markov Models and Seasonal Regime Models. AERTM had the lowest Bayesian Information Criterion of all models which gives good balance between model complexity and explanatory power and lower rainfall reconstruction error with a RMSE of 8.9 mm/day. Further evidence of correct reproduction of extreme rainfall behavior was obtained by upper-tail quantile analysis of extreme behavior at the 95 th and 99 th percentile, as compared to observed extremes. These results indicate that adaptive eventconditioned regime modelling is effective in non-stationary rainfall dynamics in climate variability.