Non-linear granger causality approach for non-stationary modelling of extreme precipitation

Interactions and feedback between the atmosphere, oceans, and land lead to natural climatic variations responsible for severe weather events such as Extreme Precipitation (EP). Understanding the lagged/delayed effect of teleconnections would improve the predictive capabilities, especially in a complex, non-linear, and non-stationary hydro-climatic system. The objective of the study is two-fold: (i) selection of lagged climate indices to improve the efficacy of the NS model using Granger Causality and (ii) non-stationary modeling of extreme precipitation. This study proposes a new Generalized Regression Neural Network based Non-linear Granger (NL-GRNN) causality approach. The climate indices with their optimal time-lags having significant Granger causality are selected for modeling EP in a Non-Stationary Generalized Extreme Value (NSGEV) framework. Seventeen climate indices are considered and applied for two regions of different climatic conditions, namely Chennai, India, and San Diego, North America. Results show that NL-GRNN based NSGEV models are found to be computationally efficient and show better performance compared to the linear- and artificial neural network based Granger causality approaches. It is observed that the ENSO-related indices and the northern hemisphere climate variability modes such as AO, PNA, and NAO dominantly influence EP over Chennai and San Diego, respectively. The outcomes of the study will be useful in planning and design of water infrastructure, disaster management and disease outbreak.