Date of Award
8-2026
Degree Type
Report
Degree Name
Master of Computer Science (MCS)
Department
Computer Science
Committee Chair(s)
Shah Muhammad Hamdi (Committee Chair)
Committee
Shah Muhammad Hamdi
Committee
Soukaina Filali Boubrahimi
Committee
Yiming Su
Abstract
Machine learning methods applied to multivariate time series data have emerged as powerful tools across a range of scientific domains. This report examines two distinct application areas in which such methods yield actionable predictive insights: hydrological streamflow forecasting and solar flare prediction in space weather.
In the domain of streamflow forecasting, a Two-Graph Spatio-Temporal Graph Neural Network (Two-Graph STGNN) was developed to predict river discharge across a 20-station network in the Upper Colorado River Basin. The architecture separates hydrological and meteorological feature streams into two complementary graph representations and fuses them through a learned attention mechanism. Systematic evaluation across 23 input-output configurations with 49 independent training seeds per configuration yielded a peak Nash-Sutcliffe Efficiency of 0.8486 at the Lees Ferry outlet station and basin-wide superiority over the single-graph baseline in 17 of 23 configurations.
In the domain of solar flare prediction, this report describes the design, implementation, and deployment of FlaPLeT (Flare Prediction by Learning from Time Series), a full-stack web application that operationalizes end-to-end machine learning pipelines on multivariate time series data drawn from the SWAN-SF benchmark dataset. The platform supports dataset upload, preprocessing, class-imbalance correction through synthetic oversampling, graph construction, and model training. FlaPLeT is deployed in production on a dedicated Linux server at Utah State University and is publicly accessible at flaplet.org, fulfilling National Science Foundation requirements for open cyberinfrastructure access.
Together, these two research efforts demonstrate that machine learning pipelines for multivariate time series data can be made scientifically rigorous, computationally reproducible, and practically accessible through thoughtful model design and robust software infrastructure
Recommended Citation
Alagarsamy, Kishore Ragul, "Graph-Based Machine Learning for Multivariate Time Series Prediction in Scientific Domains: Streamflow Forecasting and Solar Flare Prediction" (2026). All Graduate Reports and Creative Projects, Fall 2023 to Present. 174.
https://digitalcommons.usu.edu/gradreports2023/174
Included in
Copyright for this work is retained by the student. If you have any questions regarding the inclusion of this work in the Digital Commons, please email us at .