Analyzing el nino southern oscillation predictions from long-short-merm-memory models

El Nino Southern Oscillation (ENSO) can have global impacts across the world. Because of its prevalence, scientists run models to forecast its next move. Here, long-short-term-memory models (LSTM) were compared to linear regression models (LR) as first steps to explore the potential benefits of simple deep neural networks for predicting ENSO. Each model’s prediction capabilities were tested with sea surface temperatures (SST), warm water volumes, and zonal winds as predictors, individually and in combinations, utilizing both monthly and daily resolution data, across a total of 11 leads. By utilizing these three variables, we examine different forms of climate variability within the coupled system (SST), the subsurface ocean (warm water volume), and the atmosphere (zonal winds), and we quantify the relative importance of each of these processes for ENSO predictability through two statistical modeling approaches: LSTM and LR. Results show that when using monthly data as predictors, predictions from LSTM were similar to predictions from LR. However, with daily data, LSTM exhibited some advantage over LR in terms of the correlation coefficient, especially with daily resolution SST as a predictor and at longer leads. This can be appealing because once the computationally expensive training of LSTM is complete, the predictions employing the trained model can be relatively cheap to perform thereafter.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Atmospheric Sciences

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2018

Author dc:creator

• Huang, Andrew K.

Contributors dc:contributor

• Sriver, Ryan L.

• ENSO
• LSTM
• LR
• modeling
• deep learning
• forecast

Statement dc:rights

• Copyright 2018 Andrew Huang

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/101011

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/101011