Date of Award

Spring 1-1-2012

Document Type


Degree Name

Master of Science (MS)

First Advisor

Balaji Rajagopalan

Second Advisor

Edith Zagona

Third Advisor

Andrew Wood


On a seasonal time scale, forecast centers of National Weather Service produce streamflow forecasts via a method called Ensemble Streamflow Prediction (ESP). In conjunction with the physically-based Sacramento Soil Moisture Accounting model (SAC-SMA), ESP uses historical weather sequences for the forecasting period starting from model's current initial conditions, to produce ensemble streamflow. There are two major drawbacks of this method|(i) the ensembles are limited to the length of historical record thereby producing limited variability and (ii) incorporating seasonal climate forecasts such as El Niño Southern Oscillation (ENSO) is done by selecting a subset of historical sequences which further reduces the variability of streamflow forecasts. The need for alleviating these drawbacks motivates the proposed research. To this end, this research effort has two components (i) an improved multi-site stochastic weather generator and (ii) coupling it to the SAC-SMA model for ensemble streamflow forecasting.

We enhanced the traditional K-nearest neighbor semi-parametric stochastic weather generator (SWG). In SWG the daily precipitation state (wet or dry) is modeled as a Markov Chain and the weather vector on a given day is simulated conditioned on the previous day's precipitation state and weather vector and current day's precipitation state. A K-nearest neighbor resampling approach is used to simulate from the conditional probability density function. Our improvements to this stochastic generator include (i) clustering the locations into climatologically homogeneous regions and applying the weather generator separately for each region and jointly to better capture the spatial heterogeneity and, (ii) modifying the resampling approach to incorporate probabilistic seasonal climate forecast. We tested this enhanced weather generator by applying it to daily weather sequences at 66 locations in the San Juan River Basin. The proposed method generates a rich variety of weather sequences capturing the distributional properties at all the locations and the spatial dependence. It also simulates consistent weather sequences conditioned on seasonal climate forecasts.

The multi-site stochastic weather generator was coupled with the SAC-SMA model (WGESP) within NWS's new Community Hydrologic Prediction System (CHPS) to produce ensemble streamflow forecast. Spring season ensemble forecasts at several lead times from Nov through Apr for the period 1981{2010 were made from WG-ESP and the traditional ESP for the San Juan River Basin. We show that the weather generator based ensemble produces a rich variability in the flows including extremes and a higher skill at long lead times. Especially, skill in wet year forecast was found to be higher than dry years.

The flexible and robust framework provides many opportunities to further improve the ESP system in enabling increased skills at longer lead times that will be of immense help to water resources managers.