Date of Award
5-2023
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Civil Engineering
Committee Chair/Advisor
Dr. Abdul A Khan
Committee Member
Dr. Vidya Samadi
Committee Member
Dr. Ashok Mishra
Committee Member
Dr. Catherine Wilson
Committee Member
Dr. Biswa Bhattacharya
Abstract
Hydrologic models provide a comprehensive tool to calibrate streamflow response to environmental variables. Various hydrologic modeling approaches, ranging from physically based to conceptual to entirely data-driven models, have been widely used for hydrologic simulation. During the recent years, however, Deep Learning (DL), a new generation of Machine Learning (ML), has transformed hydrologic simulation research to a new direction. DL methods have recently proposed for rainfall-runoff modeling that complement both distributed and conceptual hydrologic models, particularly in a catchment where data to support a process-based model is scared and limited.
This dissertation investigated the applicability of two advanced probabilistic physics-informed DL algorithms, i.e., deep autoregressive network (DeepAR) and temporal fusion transformer (TFT), for daily rainfall-runoff modeling across the continental United States (CONUS).
We benchmarked our proposed models against several physics-based hydrologic approaches such as the Sacramento Soil Moisture Accounting Model (SAC-SMA), Variable Infiltration Capacity (VIC), Framework for Understanding Structural Errors (FUSE), Hydrologiska Byråns Vattenbalansavdelning (HBV), and the mesoscale hydrologic model (mHM). These benchmark models can be distinguished into two different groups. The first group are the models calibrated for each basin individually (e.g., SAC-SMA, VIC, FUSE2, mHM and HBV) while the second group, including our physics-informed approaches, is made up of the models that were regionally calibrated. Models in this group share one parameter set for all basins in the dataset. All the approaches were implemented and tested using Catchment Attributes and Meteorology for Large-sample Studies (CAMELS)'s Maurer datasets.
We developed the TFT and DeepAR with two different configurations i.e., with (physics-informed model) and without (the original model) static attributes. Various catchment static and dynamic physical attributes were incorporated into the pipeline with various spatiotemporal variabilities to simulate how a drainage system responds to rainfall-runoff processes. To demonstrate how the model learned to differentiate between different rainfall–runoff behaviors across different catchments and to identify the dominant process, sensitivity and explainability analysis of modeling outcomes are also performed. Despite recent advancements, deep networks are perceived as being challenging to parameterize; thus, their simulation may propagate error and uncertainty in modeling. To address uncertainty, a quantile likelihood function was incorporated as the TFT loss function. The results suggest that the physics-informed TFT model was superior in predicting high and low flow fluctuations compared to the original TFT and DeepAR models (without static attributes) or even the physics-informed DeepAR. Physics-informed TFT model well recognized which static attributes more contributing to streamflow generation of each specific catchment considering its climate, topography, land cover, soil, and geological conditions. The interpretability and the ability of the physics-informed TFT model to assimilate the multisource of information and parameters make it a strong candidate for regional as well as continental-scale hydrologic simulations. It was noted that both physics-informed TFT and DeepAR were more successful in learning the intermediate flow and high flow regimes rather than the low flow regime. The advantage of the high flow can be attributed to learning a more generalizable mapping between static and dynamic attributes and runoff parameters. It seems both TFT and DeepAR may have enabled the learning of some true processes that are missing from both conceptual and physics-based models, possibly related to deep soil water storage (the layer where soil water is not sensitive to daily evapotranspiration), saturated hydraulic conductivity, and vegetation dynamics.
Recommended Citation
Sadeghi Tabas, Sadegh, "Explainable Physics-informed Deep Learning for Rainfall-runoff Modeling and Uncertainty Assessment across the Continental United States" (2023). All Dissertations. 3269.
https://tigerprints.clemson.edu/all_dissertations/3269
Author ORCID Identifier
https://orcid.org/0000-0001-9157-3397
Included in
Civil and Environmental Engineering Commons, Computational Engineering Commons, Computer Engineering Commons