Precipitation Frequency Analysis in Regions of Complex Topography

Precipitation-Frequency Estimates, which characterize the probability associated with precipitation for a specific duration exceeding given thresholds, are critical for a variety of design engineering applications, such as floodplain and storm water management and dam safety hazard analyses, among others. In this work, we develop and demonstrate two methods to estimate precipitation-frequency estimates in regions of complex topography, specifically the Tennessee River Valley Watershed. We combine a known objective artificial neural network, the SOM algorithm, with two regional frequency methods, L-moments and Bayesian inference methods, to estimate precipitation frequency estimates across the watershed. These frequency methods vary widely in terms of the level of complexity (and consequently effort) and in the way in which epistemic uncertainty is estimated. We apply the SOM algorithm to geophysical information (i.e., latitude, longitude, elevation) and observed precipitation data (i.e., mean annual precipitation, mean one-day annual maxima) from 179 precipitation stations across the watershed, resulting in 14 different groups (i.e., homogeneous regions). We apply the two regional frequency methods to historical annual maximum precipitation observations from these 14 groups separately. Results suggest that the SOM algorithm is a useful tool for identifying and grouping similarly-behaved point precipitation observations. Furthermore, the frequency results from these analyses indicate that uncertainty estimates from the L-moments analysis are consistently less than the uncertainty estimates from Bayesian inference. We also demonstrate the application of these two frequency methods on a gridded precipitation dataset, the 100-member ensemble data from Newman et al. (2015). With some modifications, the techniques presented here can be applied to other point-based precipitation datasets of interest to the user.