Multimodel Ensembles of Streamflow Forecasts: Role of Predictor State in Developing Optimal Combination.

ensembles;predictor state

Devineni, Naresh ; Dr.Sankar Arumugam, Committee Chair,Devineni, Naresh ; Dr.Sankar Arumugam ; Committee Chair

University:North Carolina State University

关键词: ensembles;    predictor state;   

Others  :  https://repository.lib.ncsu.edu/bitstream/handle/1840.16/2256/etd.pdf?sequence=1&isAllowed=y

美国|英语

来源: null

【 摘 要 】

Information on season-ahead streamflow forecasts is beneficial for the operation and management of water supply systems. Developing such long-lead (3-12 months) stream flow forecasts typically depend on exogenous climatic conditions particularly sea surface temperature (SSTs) conditions in tropical oceans. Identification of such conditions that influence the moisture transport into water resources regions is important to develop low-dimensional statistical models and to analyze climatic forecasts from General Circulation Models (GCMs). The main purpose of this study is to develop probabilistic streamflow forecasts for the Falls Lake Reservoir, NC, for the summer season that is critical for developing water management strategies so that the City of Raleigh's water demands could be met through water conservation measures. The study develops two low-dimensional statistical models based on SSTs in the tropical Pacific, tropical Atlantic and over the NC Coast. Given that prediction from any model is bound to have unavoidable error/model uncertainty, the study intends to combine the forecasts from individual models to develop an improved multi-model forecast. For this purpose, the study develops an algorithm for combining forecasts from individual forecasts by evaluating the performance of individual forecasts contingent on climatic (predictor) conditions. The methodology is demonstrated through the development of multi-model ensembles of streamflow forecasts for the Falls Lake reservoir by combining probabilistic streamflow forecasts from two low dimensional statistical models. Using Rank Probability Score (RPS) for evaluating each year's streamflow forecasts for the summer months (July-August-September) from the two low dimensional models, the methodology proportionately gives higher representation by drawing increased ensembles for a model that has better predictability under similar predictor conditions. The performance of the multi-model forecasts is compared with the individual model's performance using various performance evaluation measures. By developing multi-model ensembles based onleave-one out cross validation and split sampling, the study shows that evaluating the model's performance based on the predictor state provides a better alternativein developing multi-model ensembles instead of combining the models purely based on their long-term predictability. The method is also extended to combine various GCMs to get improved winter (December-January-February) precipitation forecast for the entire US. Finally the study shows the utility of the multi model precipitation ensembles to develop improved streamflow forecasts for the Falls Lake through statistical downscaling.