Date of Award

9-2021

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Department of Operational Sciences

First Advisor

Andrew Geyer, PhD

Abstract

Clustering weather data is a valuable endeavor in multiple respects. The results can be used in various ways within a larger weather prediction framework or could simply serve as an analytical tool for characterizing climatic differences of a particular region of interest. This research proposes a methodology for clustering geographic locations based on the similarity in shape of their temperature time series over a long time horizon of approximately 11 months. To this end an emerging and powerful class of clustering techniques that leverages deep learning, called deep representation clustering (DRC), are utilized. Moreover, a time series specific DRC algorithm is proposed that addresses a current gap in the field. Finally, deep learning based weather prediction is an increasingly common research topic as a means of obtaining more rapid predictions when compared to traditional numerical weather prediction (NWP). Since there are known physical equations that govern atmospheric behavior, namely the Navier-Stokes equations, the concept of reformulating these laws into a physics based loss function is explored with particular interest in whether a model trained with such a loss function can outperform it’s baseline counterpart.

AFIT Designator

AFIT-ENS-DS-21-S-037

DTIC Accession Number

AD1149667

Share

COinS