DEVELOPMENT OF A GENERALIZED FRAMEWORK FOR CLIMATEINFORMED NONSTATIONARY EXTREME EVENT MODELLING

Abstract

Globally intensifying hydrological extremes pose a major threat to society and question the reliability of traditional modelling. Especially, the uncertainty in precipitation pattern risks the hydrological extremes that stress water infrastructure and management as well as the country’s economy. There is also a notable change in the hydrological extreme events and is likely to continue in the near future. The stationary assumption in the modelling will either underestimate or overestimate extreme events. Hence in recent times, hydrologists have focused on developing stochastic nonstationary modelling of hydrological extremes by incorporating physical processes as a covariate for designing sustainable and reliable infrastructure. The possible physical covariates are hydro-climatic covariates such as climatic teleconnections, local temperature changes, global temperature changes etc., as they impact on rainfall patterns and the basin covariates such as urbanised area, vegetated area, reservoir index etc., as they impact basin properties. Most of the climate indices are governed by atmospheric processes. Therefore, they may intercorrelate with each other. The previous studies have chosen the best covariate model by modelling the combinations of different potential covariates using information criteria without checking the interdependency between the covariates, which may lead to multicollinearity and complexity in the computation. Hence this study aims to develop a framework for covariate-based timevarying flood frequency analysis by including uncorrelated potential covariates. This framework consists of three steps; firstly, the signature of nonstationarity is detected in the distribution parameter rather than the original time series using the Time Sliding Window approach; Secondly, the identification of the potential covariate based on previous literature; then check interdependency within covariates to confirm uncorrelated potential covariates, and finally, nonstationary models are constructed by considering a linear trend in the distribution parameter based on uncorrelated potential covariates.