Journal Article
Atmospheric Chemistry and Physics, vol. 20, iss. 4, pp. 2303-2317, 2020
Authors
Theodore Weber, Austin Corotan, Brian Hutchinson, Ben Kravitz, Robert Link
Abstract
Abstract. We investigate techniques for using deep neural networks to produce surrogate
models for short-term climate forecasts. A convolutional neural network is
trained on 97 years of monthly precipitation output from the 1pctCO2 run (the
CO2 concentration increases by 1 % per year) simulated by the second-generation Canadian Earth System Model (CanESM2). The neural network clearly outperforms a persistence forecast and
does not show substantially degraded performance even when the forecast length
is extended to 120 months. The model is prone to underpredicting precipitation
in areas characterized by intense precipitation events. Scheduled sampling
(forcing the model to gradually use its own past predictions rather than ground
truth) is essential for avoiding amplification of early forecasting errors.
However, the use of scheduled sampling also necessitates preforecasting
(generating forecasts prior to the first forecast date) to obtain adequate
performance for the first few prediction time steps. We document the training
procedures and hyperparameter optimization process for researchers who wish to
extend the use of neural networks in developing surrogate models.
