Probabilistic simulation of big climate data for robust quantification of changes in compound hazard events

Abstract

Understanding changes in extreme compound hazard events is important for climate mitigation and policy. By definition, such events are rare so robust quantification of their future changes is challenging. An approach is presented, for probabilistic modelling and simulation of climate model data, which is invariant to the event definition since it models the underlying weather variables. The approach is based on the idea of a ‘moving window’ in conjunction with Generalised Additive Models (GAMs) and Bayesian inference. As such, it is robust to the data size and completely parallelizable, while it fully quantifies uncertainty allowing also for comprehensive model checking. Lastly, Gaussian anamorphosis is used to capture dependency across weather variables. The approach results in probabilistic simulations to enable extrapolation beyond the original data range and thus robust quantification of future changes of rare events. We illustrate by application to daily temperature, humidity and precipitation from a regional climate model.

Keywords

• Climate change
• GAMs
• Bayesian smoothing
• Stochastic simulation
• Space–time model