Hydroclimatic forecasting in the western United States using paleoclimate reconstructions and data-driven models

This thesis investigated climate variability and their associated hydrologic responses in the western United States. The western United States faces the problem of water scarcity, where the management and mitigation of available water supplies are further complicated by climate variability. Climate variability associated with the phases of oceanic-atmospheric oscillations has been shown to influence streamflow and precipitation, where predictive relationships have led to the possibility of producing long-range forecasts. Based on literature review, four oceanic-atmospheric oscillation indices were identified in having the most prominent influence over the western United States including the El Niño - Southern Oscillation (ENSO), Pacific Decadal Oscillation (PDO), Atlantic Multidecadal Oscillation (AMO), and North Atlantic Oscillation (NAO). However, these hydroclimatic processes are not fully understood and are difficult to describe in physically-based models. A viable alternative to generating forecasts is through data-driven models, which extract relationships in a dataset of oscillation inputs and hydrologic outputs to build a structured forecasting model. One of the limitations to using oceanic-atmospheric oscillations in a data-driven model is a short instrumental record from which the model can train on. Data-driven models often perform better when they are subjected to a larger training dataset. Reconstructions have the potential to extend the period of record by several centuries, which may aid in identifying important hydroclimatological relationships and improving the quality of forecasts. With this motivation, this study focused on increasing the forecast lead time through the use of reconstructions of oceanic-atmospheric oscillations in the western United States. First, reconstructions of oscillations were investigated to increase the forecast lead time of four streamflow gages in the Upper Colorado River Basin (UCRB) by using the KStar and M5P data-driven models. Secondly, an expanded spatial examination was performed over the western United States for 21 streamflow gages to increase the forecast lead time using the KStar model. Thirdly, different combinations of oceanic-atmospheric oscillations were tested for precipitation forecasts for 20 climate divisions throughout the western United States. Finally, a support vector machine (SVM) was used to increase the streamflow forecast lead time for 21 gages in the western United States. In order to accomplish this task, a collection of annual time series, processing techniques, testing procedures, and performance measures were used. Reconstructions were available for oscillation indices, streamflow volumes, and climate division precipitation was developed with a common timeframe available as far back as 1658. The instrumental records used ranged from 1900 to 2007 Noise was removed from the dataset using a 3-year, 5-year, and 10-year moving average filter. A 10-fold cross-validation technique was used as opposed to splitting the dataset into training and testing periods so that the entire dataset could be tested and to better capture the non-stationarity of the dataset. The performance of the models were evaluated through a series of independent measures which include the root mean squared error (RMSE), mean absolute error (MAE), RMSE-standard deviation ratio (RSR), Pearson\u27s correlation coefficient (R), Nash-Sutcliffe coefficient of efficiency (NSE), and linear error in probability space (LEPS) skill score (SK). In addition, all of the models were compared with a multiple linear regression (MLR) model. The results indicated that the lead time for streamflow forecasts in the Upper Colorado River Basin were increased up to 5 years with the KStar model. In addition, 1-year and 2-year lead-time forecasts with the KStar model were achieved for 21 streamflow gages in the western United States. A 1-year precipitation forecast was also made for 20 climate divisions with the KStar model throughout the western United States and found that the forecasts deteriorated when any of the four oscillations were dropped as predictors. Finally, the SVM model produced streamflow forecasts in the western United States using the raw data at the 1-year and 5-year lead time. In addition, the results indicated that the use of all four oceanic-atmospheric oscillation indices (i.e. ENSO, PDO, AMO, and NAO) provided the best forecasts, and dropping any of the indices yielded inferior results. It was also found that noise removal increased the performance of the model, by aiding in the identification of the oscillation phases. The contributions made from this research include an extension of the lead-time for streamflow and precipitation forecasts and a better understanding of the effects of climate variability. This study was the first to use reconstructions in a data-driven forecasting model for streamflow and precipitation. Other studies have incorporated reconstructions for use in determining hydroclimatic behaviors and relationships in comparison to the observed record; however, there have been no previous attempts to use reconstructions with data-driven techniques for forecasting purposes. Overall, this research provided a better understanding of climate variability and their hydrologic responses in the western United States. The forecasting models produced through this research are expected to aid water managers in the long-term planning and management of water resources in the western United States