Operational Short-Range to Long Range (SR2LR) Streamflow Forecasting for Poorly Gauged Basins: The Unexplored Dimension of Variational Ensemble Forecasting, the Spatio-Temporal Structure of Modeling Paradigms, and the Role of Machine Learning Strategies to Improve Hydrological Hypotheses
AuthorValdes-Pineda, Rodrigo M.
KeywordsHydrologic Forecasting System
Real-Time Streamflow Forecasting
Satellite Precipitation Products
Variational Ensemble Forecasting
AdvisorValdés, Juan B.
MetadataShow full item record
PublisherThe University of Arizona.
RightsCopyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction, presentation (such as public display or performance) of protected items is prohibited except with permission of the author.
AbstractThe Prediction in Ungauged Basins (PUB) Initiative was created to advance scientific understanding and estimation of hydrological processes, as well as associated uncertainties, to improve prediction capabilities in basins which are poorly gauged (Sivapalan et al., 2003; Blöschl et al. 2013; Blöschl, 2016; Blöschl et al., 2019). The main objectives defined by the PUB Initiative were to (1) improve the ability of existing hydrologic models to predict with reduced uncertainty, and to (2) develop new and innovative models representing the space–time variability of hydrological processes (Hrachowitz et al., 2013). Despite the International Association of Hydrological Sciences (IAHS) dedicated an entire decade (2003-2012) to advance the problem of Prediction in Ungauged Basins (Hrachowitz et al. 2013), the central goal remains largely a challenge (Kratzert et al., 2020).This dissertation discusses the main steps and decisions required to implement, calibrate, and validate an operational (real-time) Hydrological Forecasting System (HFS), for short-range to long-range (SR2LR) daily streamflow forecasting. The HFS was implemented under an operational context, and experimentally evaluated in the “poorly-gauged” Upper Zambezi River Basin (UZRB) and its “ungauged” sub-basins. The state of art focuses on describing several hydrological modelling strategies (HMS), discussing the way hydrological ensembles have been traditionally performed, and how meteorological and hydrological uncertainty have been quantified. Additionally, a novel Variational Ensemble Forecasting (VEF) approach was applied and evaluated assuming that any combination of multiple inputs, models, and optimal parameters sets, is a practical hydrological ensemble that can be used to reproduce daily streamflow forecasts with reduced total uncertainty. The VEF approach implemented allowed for increasing the number of hydrological ensembles (outputs) from all possible combinations of multiple satellite products (or multiple climate models), hydrologic models, and optimal parameter sets. The performance of VEF was compared and evaluated with classical approaches used to develop hydrological ensembles (input-model-output). To complement the application of the VEF approach, three hydrologic processing hypotheses (HPH) to quantify the hydrological uncertainty propagated from the components of a modelling chain were used: (1) Hydrological Pre-Processing (HPR); (2) Hydrological Processing (HP), and (3) Hydrological Post-Processing (HPP). These HPH’s were evaluated with practical examples in the UZRB and its sub-basins, proving to be a more efficient and systematic way to quantify and reduce the uncertainty propagated from an operational VEF implementation. To inform the development of reliable operational hydrologic forecasting products, the HPR hypotheses are evaluated through the analysis of the spatio-temporal structure of meteorological uncertainty. This analysis allowed determining what factors dominate the propagation of meteorological uncertainty for the operational implementation of distributed hydrologic modeling strategies i.e. the spatial resolution of a meteorological input, the leading time of the forecasts associated to it, and the size of the basin under evaluation. The role of machine learning (ML) approaches for daily streamflow forecasting was also evaluated by coupling VEF with ML techniques (VEF-ML approach). In doing so, several hydrologic learning strategies (inference versus pattern-based) were compared and evaluated to improve the HFS performance through hydrologic post-processing hypotheses (see i.e. Nearing et al., 2020 a, b; Gauch et al., 2020). Lastly, we will discuss how the properties of hydrologic forecasts can be improved by applying Hydrologic Forecast Skill Analysis (FCA) to answer the following question: How to perform a Skill Analysis of Seasonal Hydrologic Streamflow Forecasts? The main contributions and results of this dissertation highlight several opportunities and challenges for future research aimed to advance the main goal and the corresponding objectives of the PUB initiative.
Degree ProgramGraduate College
Degree GrantorUniversity of Arizona
Showing items related by title, author, creator and subject.
Near Real-Time Forecasting of Epidemics Using Data Assimilation with Simple ModelsLega, Joceline; Biegel, Hannah; Morzfeld, Matthias; Watkins, Joe; Zhang, Hao (The University of Arizona., 2020)With the advent of expedient data sharing of epidemiological reports, efforts have been made to forecast ongoing epidemics in real-time. Throughout this work, we develop a real-time forecasting framework that can be applied to a variety of diseases. Specifically this framework pairs data assimilation with a simple mechanistic model in the form of an incidence vs cumulative cases (ICC) curve. We provide results in two different contexts: seasonal influenza and COVID-19. We demonstrate that the flexibility in this framework results in accurate forecasts even when disease dynamics change during the outbreak, such as was the case with COVID-19. We further provide the groundwork for an avenue of improvement of the flu forecasts by using epidemiological data available early on in the flu season to categorize the severity of the ongoing season. The presented framework can be applied to other diseases, locations, or population scales.
Likelihood development for a probabilistic flash flood forecasting modelKeefer, Timothy Orrin; Davis, Donald R. (The University of Arizona., 1993)An empirical method is developed for constructing likelihood functions required in a Bayesian probabilistic flash flood forecasting model using data on objective quantitative precipitation forecasts and their verification. Likelihoods based on categorical and probabilistic forecast information for several forecast periods, seasons, and locations are shown and compared. Data record length, forecast information type and magnitude, grid area, and discretized interval size are shown to affect probabilistic differentiation of amounts of potential rainfall. Use of these likelihoods in Bayes' Theorem to update prior probability distributions of potential rainfall, based on preliminary data, to posterior probability distributions, reflecting the latest forecast information, demonstrates that an abbreviated version of the flash flood forecasting methodology is currently practicable. For this application, likelihoods based on the categorical forecast are indicated. Apart from flash flood forecasting, it is shown that likelihoods can provide detailed insight into the value of information contained in particular forecast products.
Seasonal Prediction of North Atlantic Accumulated Cyclone Energy and Major Hurricane ActivityDavis, Kyle; Zeng, Xubin; Univ Arizona, Dept Hydrol & Atmospher Sci (AMER METEOROLOGICAL SOC, 2019-02)Building upon our previous seasonal hurricane prediction model, here we develop two statistical models to predict the number of major hurricanes (MHs) and accumulated cyclone energy (ACE) in the North Atlantic basin using monthly data from March to May for an early June forecast. The input data include zonal pseudo-wind stress to the 3/2 power, sea surface temperature in the North Atlantic, and, depending on the magnitude of the Atlantic multidecadal oscillation index, the multivariate ENSO index. From 1968 to 2017, these models have a mean absolute error of 0.96 storms for MHs and 30 units for ACE. When tested over an independent period from 1958 to 1967, the models show a 22% improvement for MHs and 16% for ACE over a no-skill metric based on a 5-yr running average. Both the MH and ACE results show consistent improvements over those produced by three other centers using statistical-dynamical hybrid models and a 5-yr running average prediction over the period 2000-17 for MHs (2003-17 for ACE) in a simulated real-time prediction. These improvements vary from 25% to 37% for MHs and from 15% to 37% for ACE. While most forecasting centers called for a slightly above-average hurricane season in May/June 2017, our models predicted in June 2017 a very active season, in much better agreement with observations.