Seasonal forecasting

Seasonal forecasting has traditionally been modeled using complex dynamical models, rather than statistical methods, often called general circulation models (GCMs). However, seasonal variations, such as those due to El Niño/Southern Oscillation (ENSO) and polar vortices, are difficult to predict using traditional methods. ML and deep learning can improve our forecasting of multi-year ENSO events    and polar vortices.

Subseasonal Forecasting
Subseasonal forecasting is the task of predicting the climate of a region between 2-8 weeks in advance. Weather and seasonal prediction which focus on forecasting climate 1-7 days and 2+ months in advance respectively have already received significant attention and are considered easier prediction problems than the subseasonal scenario. Improvements in subseasonal prediction will be realized in industries such as water management, agricultural productivity, and emergency planning for extreme weather events.


 * In [7], the authors show that machine learning models can be applied generally to the subseasonal forecasting problem context and they highlight the potential for tailored models to make large improvements over existing methods.
 * In [8], Hwang et al. develop and apply two distinct nonlinear regression models to the western United States, both of which outperform the baseline model significantly.
 * In [10], Weyn et al. design a deep learning ensemble model that competes with, but does not outperform existing methods (i.e. the model in use by the European Centre for Medium-Range Weather Forecasts, ECMWF). The proposed model is much more computationally efficient and this result suggests that the authors' research trajectory holds promise for surpassing the conventional methods.

Data

 * SubseasonalRodeo: A benchmark dataset consisting of 12 features (including temperature, precipitation, humidity, etc.) from 14 different data sources used for training and evaluating subseasonal forecast systems in the contiguous western United States.