Enhancing deep state space models for complex applications

This thesis investigates the potential to improve deep state space models towards their use in applications such as the prediction and simulation of tropical cyclone (TC) tracks. Given the significant implications of increasingly extreme and unpredictable TCs due to anthropogenic climate change, there is a pressing need for more accurate and rapid forecasting models that can adapt to changing climate patterns. Current forecasting methods, either computationally intensive dynamical models or oversimplified statistical models reliant on historical data, struggle to meet these demands. In response, this work makes three significant contributions. First, it explores overparameterization in deep state space models, specifically the effects of width scaling. Preliminary findings indicate that the width of transition and emission networks can be increased to improve model performance, with the caveat of potential overfitting risks for small datasets. Second, we propose a new deep state space model that employs hierarchical latent states to enhance model expressiveness. Lastly, we present an application of deep state space models to the task of TC track simulation. While the model was able to generate small-scale movements, it struggled to capture overall trends observed in cyclones, indicating a need for improved deep state space models for climate applications. This thesis illuminates the promise and challenges of applying deep state space models to TC forecasting and simulation. It paves the way for future research aiming to develop a deep learning-based model capable of generating tropical cyclones under projected climate scenarios, a critical tool for risk assessment, hazard prediction, and policy decision-making amidst the growing uncertainties of climate change.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Nagda, Chandni

Contributors dc:contributor

• Banerjee, Arindam

• Machine Learning
• Deep Learning
• State Space Models
• Multivariate Time Series

Statement dc:rights

• Copyright 2023 Chandni Nagda

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/121349

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F121349