Quantification of Uncertainty in Precipitation Frequency Estimates Considering Climate Change

ABSTRACT ONLY - Rising global temperatures have led to an increase in the frequency and severity of extreme precipitation events in many regions worldwide. Consistent with this trend, many areas in the United States are expected to undergo more frequent extreme precipitation in the 21st century. Intensity Duration Frequency (IDF) curves are standard tools for informing the hydrologic design of stormwater management systems and water-related infrastructure. However, IDF curves developed using historical/observational data may not reflect hazards under a changing climate. Synthetic climate model projections provide information to account for the potential effects of climate change in developing IDF curves. In this study, we explore the uncertainty associated with the development of IDF curves under current and future climate conditions for the state of Maryland using the time-series outputs of the North American Regional Climate Change Assessment Program (NARCCAP). We first apply machine learning (ML) to temporally downscale synthetic time-series outputs of climate model projections
(available at a 3-hour temporal resolution) to durations as short as 15 minutes. We explore the performance of the ML models by assessing performance in predicting large target response quantities, identifying systematic trends in errors, investigating input/output relationships using response functions, and leveraging individual and combined performance metrics. Using the downscaled precipitation time-series, we produce sets of current and mid-21st-century future climate projections in the form of an ensemble of IDF curves. We then explore the uncertainty arising from the temporal downscaling and the uncertainty arising from statistical modeling choices made in the development of IDF curves. We assess the uncertainty on two levels: across model and within model. Across model uncertainty refers to the uncertainty arising from the differences in synthetic precipitation and other meteorological variable time-series resulting from the twelve different NARCCAP climate model projections considered in this study. Within model uncertainty refers to the uncertainty arising from the modeling choices, including temporal downscaling methods, time-series types, distributions, and parameter estimation methods used to develop IDF curves using the synthetic time-series from a single climate model projection. Explicit consideration of uncertainty allows for a more complete understanding of how modeling choices can lead to differing conclusions regarding the frequency of large precipitation events. The findings have implications for risk assessments and designs that depend on IDF curves as input.