Abstract

Time series forecasting plays a crucial role in various fields, ranging from financial markets and predictive maintenance in industrial settings to healthcare analytics. Traditionally, statistical approaches were employed primarily for univariate forecasting tasks, but modern applications often require robust solutions capable of handling complex, multivariate, noisy, and non-stationary data. Current state-of-the-art solutions predominantly utilize transformer-based models, which demand significant computational resources, limiting their practicality, especially in resource-constrained environments. This dissertation introduces novel and efficient approaches to multivariate time series forecasting, leveraging NeuroEvolution-based Neural Architecture Search (NAS) methodologies. Specifically, two robust algorithms---Evolutionary eXploration of Augmenting Memory Models (EXAMM) for offline forecasting and Online NeuroEvolution-based Neural Architecture Search (ONE-NAS) for online forecasting---were developed. EXAMM significantly advances offline forecasting capabilities through innovations including distributed island repopulation, Xavier and Kaiming weight initialization, and Lamarckian weight inheritance, achieving enhanced forecasting accuracy with lightweight recurrent neural networks (RNNs) that are substantially smaller than transformer-based architectures. Real-world validations of EXAMM, including stock market forecasting and predictive maintenance for coal-fired power plants, demonstrated its practicality and financial benefits, including a notable \$7.3 million cost reduction. ONE-NAS is introduced as the first dedicated online NAS method tailored specifically for time series forecasting. It continuously evolves RNN architectures without prior training, dynamically adjusting through mutations and crossover to mitigate data drift. Comparative experiments validate ONE-NAS’s superior performance over traditional statistical methods and other contemporary online forecasting approaches, underscoring its robust adaptability to evolving real-time data. Overall, both neuroevolution-based NAS frameworks EXAMM and ONE-NAS provide scalable, efficient, and high-performing alternatives for multivariate time series forecasting, particularly suitable for deployment in computationally limited settings.

Library of Congress Subject Headings

Multivariate analysis--Data processing; Time-series analysis--Data processing; Computer network architectures; Neural networks (Computer science)

Publication Date

7-2025

Document Type

Dissertation

Student Type

Graduate

Degree Name

Computing and Information Sciences (Ph.D.)

Department, Program, or Center

Computing and Information Sciences Ph.D, Department of

College

Golisano College of Computing and Information Sciences

Advisor

Travis Desell

Advisor/Committee Member

Rui Li

Advisor/Committee Member

Alexander Ororbia

Campus

RIT – Main Campus

Plan Codes

COMPIS-PHD

Download

Share

COinS