USING NONSTATIONARY DEPTH-FREQUENCY CURVES TO CHARACTERIZE LOCAL PRECIPITATION TRENDS

As climate change advances, the stationarity assumption that governs traditional precipitation analysis is becoming untenable. Studies that incorporate nonstationarity typically use global circulation model (GCM) projections to determine the magnitude and direction of expected precipitation changes. However, the high computational costs and the coarse spatial resolution of GCMs make this method unsuitable for local precipitation analysis. In this study, nonstationarity is represented by a precipitation probability distribution with time-varying parameters. Three generalized extreme value (GEV) distributions are fitted: (1) the shift model, where the GEV location parameter varies linearly with time, (2) the stretch model, where the GEV location and scale parameters both vary linearly with time, and (3) the stationary model, a time-invariant distribution provided for the purpose of comparison. This procedure is applied to 24-h annual maximum precipitation records for ninety years (1900-1989) at five long-term measuring sites in Pennsylvania. Results varied among the five sites, suggesting that localized climate effects can cause precipitation differences at a small spatial scale. No significant nonstationarity was detected in two of the five locations. In three locations, however, increases in GEV location and scale combined to create a substantial, though not always significant, rise in the frequency of extreme precipitation. These trends were extrapolated forward over 30 years (1990-2019) and compared with an observed distribution for that year. The nonstationary models appeared to perform better at sites with stronger precipitation trends, which suggests a simple procedure for selecting sites where nonstationary analysis is most needed.