Analyzing El Niño–Southern Oscillation Predictability Using Long‐Short‐Term‐Memory Models

Andrew Huang, Ben Vega‐Westhoff, Ryan L. Sriver

Earth and Space Science · 14 January 2019

In one sentence. Compares LSTM and linear regression for ENSO prediction from 1 to 11-month leads, finding LSTM matches LR on monthly data and gains skill with daily SST at longer leads.

Abstract

Abstract El Niño–Southern Oscillation (ENSO) can have global impacts, affecting daily temperature and precipitation, and extreme weather, such as hurricanes and tornadoes. Because of its importance, scientists strive to understand the processes that govern ENSO and develop models to predict its evolution and changes in variability. Here long‐short‐term‐memory models (LSTMs) were compared to linear regression models (LR) to explore the benefits of simple, deep neural networks in predicting ENSO, in addition to quantifying the relative importance of the sources of ENSO's predictability. The models use central Pacific sea surface temperatures (SST), equatorial Pacific warm water volumes, and western Pacific zonal winds as predictors, individually and in combinations, on monthly and daily resolutions, from 1‐ to 11‐month leads. By using these predictors, many characteristic time scales are encompassed—from days‐to‐weeks in the atmosphere, to months‐to‐seasons in the coupled system, and interseasonal‐to‐interannual in the subsurface ocean. Results show, with monthly input, predictions from LSTM were like predictions from LR. However, with daily SST at longer leads, LSTM exhibited some advantage over LR in terms of the correlation coefficient. This suggests that daily SST may contain some nonlinear element that improves LSTM predictability compared to LR. In addition, this suggests that more information, such as gridded data and additional variables, would likely improve predictability using LSTM, but results would be more difficult to interpret. Overall, LSTM may 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.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.