Water and fire managers in the western United States (U.S.) rely on subseasonal forecasts—forecasts of temperature and precipitation two to six weeks in advance—to allocate water resources, manage wildfires, and prepare for droughts and other weather extremes [white2017]. While purely physics-based numerical weather prediction dominates the landscape of short-term weather forecasting, such deterministic methods have a limited skillful (i.e., accurate) forecast horizon due to the chaotic nature of weather [lorenz1963deterministic]. Prior to the widespread availability of operational numerical weather prediction, weather forecasters made predictions using their knowledge of past weather patterns and climate (sometimes called the method of analogs) [nebeker1995calculating]. The current availability of ample meteorological records and high-performance computing offers the opportunity to blend physics-based and statistical machine learning (ML) approaches to extend the skillful forecast horizon.
This data and computing opportunity, coupled with the critical operational need, motivated the U.S. Bureau of Reclamation and the National Oceanic and Atmospheric Administration (NOAA) to conduct the Subseasonal Climate Forecast Rodeo [nowak2017sub], a year-long real-time forecasting challenge, in which participants aimed to skillfully predict temperature and precipitation in the western U.S. two to four weeks and four to six weeks in advance. To meet this challenge, we developed an ML-based forecasting system and a SubseasonalRodeo dataset [dataset2018] suitable for training and benchmarking subseasonal forecasts.
ML approaches have been successfully applied to both short-term ( week) weather forecasting [doi:10.1175/WAF-D-17-0188.1, doi:10.1175/WAF-D-17-0006.1, doi:10.1175/MWR-D-17-0307.1] and longer-term climate prediction [doi:10.1175/JCLI-D-15-0648.1, doi:10.1175/JAMC-D-13-0181.1, doi:10.1002/2017GL075674], but mid-term subseasonal outlooks, which depend on both local weather and global climate variables, still lack skillful forecasts [robertson2015].
Our subseasonal ML system is an ensemble of two regression models: a local linear regression model with multitask model selection (MultiLLR) and a weighted local autoregression enhanced with multitask -nearest neighbor features (AutoKNN). The MultiLLR model introduces candidate regressors from each data source in the SubseasonalRodeo dataset and then prunes irrelevant predictors using a multitask backward stepwise criterion designed for the forecasting skill objective. The AutoKNN model extracts features only from the target variable (temperature or precipitation), combining lagged measurements with a skill-specific form of nearest-neighbor modeling. For each of the two Rodeo target variables (temperature and precipitation) and forecast horizons (weeks 3-4 and weeks 5-6), this paper makes the following principal contributions:
We release a new SubseasonalRodeo dataset suitable for training and benchmarking subseasonal forecasts.
We introduce two subseasonal regression approaches tailored to the forecast skill objective, one of which uses only features of the target variable.
We introduce a simple ensembling procedure that provably improves average skill whenever average skill is positive.
We show that each regression method alone outperforms the Rodeo benchmarks, including a debiased version of the operational U.S. Climate Forecasting System (CFSv2), and that our ensemble outperforms the top Rodeo competitor.
We show that, over 2011-2018, an ensemble of our models and debiased CFSv2 improves debiased CFSv2 skill by 37-53% for temperature and 128-154% for precipitation.
2 The Subseasonal Climate Forecast Rodeo
The Subseasonal Climate Forecast Rodeo was a year-long, real-time forecasting competition in which, every two weeks, contestants submitted forecasts for average temperature (C) and total precipitation (mm) at two forecast horizons, 15-28 days ahead (weeks 3-4) and 29-42 days ahead (weeks 5-6). The geographic region of interest was the western contiguous United States, delimited by latitudes 25N to 50N and longitudes 125W to 93W, at a 1 by 1 resolution, for a total of grid points. The initial forecasts were issued on April 18, 2017 and the final on April 3, 2018.
Forecasts were judged on the spatial cosine similarity between predictions and observations adjusted by a long-term average. More precisely, letdenote a date represented by the number of days since January , , and let , , and respectively denote the year, the day of the year, and the month-day combination (e.g., January ) associated with that date. We associate with the two-week period beginning on an observed average temperature or total precipitation and an observed anomaly