SIParCS 2018- Karen Stengel

Machine Learning Long Term Weather Forecast

Recorded Talk

In a stochastic regime, running deterministic models based on an initial state will naturally diverge as the initial uncertainty amplifies; two or more initial states become less and less similar as time progresses. In atmospheric phenomena, this uncertainty is unbound so it is almost impossible at any given time to fully map all future states. Because of this, forecasting the weather on the seasonal timescale (greater than 20 days) is inherently inaccurate. Traditionally, an ensemble of forecasts, based on slightly different initial conditions, are run to outline likely future states as a probability of occurrence.

A correlation between abnormally hot sea surface temperatures (SST) in the Pacific and above average temperatures in the Eastern United states 20 - 60 days later was suggested by McKinnon et al. 2015. In this paper the authors predict above average ‘hot’ days using statistical models of the Pacific Extreme Pattern SST from 1982-2015. Here, we aim to replicate these results by taking a Deep Learning approach using Neural Networks. All models were run on NCAR’s supercomputer Cheyenne’s K80 GPUs with the goal of increasing runtime performance. Thanks to this technology, we were able to train our models 6 times faster with a 10% increase in accuracy compared to using the CPUs on the same node. By training the networks on “images” of SST data, we have shown that it is possible to predict whether an anomalously hot day will occur in 1613 stations in the Eastern US 20, 30, 40, and 50 days in advance.

Mentors: Davide Del Vento, Alessandro Fanfarillo, Negin Sobhani, Dave Stepaniak

Slides