Evaluation of stochastic weather generators for capturing the statistics of extreme precipitation events in the Catskill Mountain watersheds, New York State

Thursday, 17 December 2015
Poster Hall (Moscone South)
Nachiketa Acharya1, Allan Frei1, Emmet M Owens2 and Jie Chen3, (1)CUNY Hunter College, New York, NY, United States, (2)New York City Department of Environmental Protection, Kingston, NY, United States, (3)Wuhan University, State Key Laboratory of Water Resources & Hydropower Engineering Science, Wuhan, China
Watersheds located in the Catskill Mountains area, part of the eastern plateau climate region of New York, contributes about 90% of New York City's municipal water supply, serving 9 million New Yorkers with about 1.2 billion gallons of clean drinking water each day. The New York City Department of Environmental Protection has an ongoing series of studies to assess the potential impacts of climate change on the availability of high quality water in this water supply system. Recent studies identify increasing trends in total precipitation and in the frequency of extreme precipitation events in this region. The objectives of the present study are: to analyze the proba­bilistic structure of extreme precipitation based on historical observations: and to evaluate the abilities of stochastic weather generators (WG), statistical models that produce synthetic weather time series based on observed statistical properties at a particular location, to simulate the statistical properties of extreme precipitation events over this region. The generalized extreme value distribution (GEV) has been applied to the annual block maxima of precipitation for 60 years (1950 to 2009) observed data in order to estimate the events with return periods of 50, 75, and 100 years. These results were then used to evaluate a total of 13 WGs were : 12 parametric WGs including all combinations of three different orders of Markov chain (MC) models (1st , 2nd and 3rd) and four different probability distributions (exponential, gamma, skewed normal and mixed exponential); and one semi parametric WG based on k-nearest neighbor bootstrapping. Preliminary results suggest that three-parameter (skewed normal and mixed exponential distribution) and semi-parametric (k-nearest neighbor bootstrapping) WGs are more consistent with observations. It is also found that first order MC models perform as well as second or third order MC models.