Improving Subseasonal Forecasting in the Western U.S. with Machine Learning

Abstract

Water managers in the western United States (U.S.) rely on longterm forecasts of temperature and precipitation to prepare for droughts and other wet weather extremes. To improve the accuracy of these longterm forecasts, the U.S. Bureau of Reclamation and the National Oceanic and Atmospheric Administration (NOAA) launched the Subseasonal Climate Forecast Rodeo, 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. Here we present and evaluate our machine learning approach to the Rodeo and release our SubseasonalRodeo dataset, collected to train and evaluate our forecasting system.Our system is an ensemble of two nonlinear regression models. The first integrates the diverse collection of meteorological measurements and dynamic model forecasts in the SubseasonalRodeo dataset and prunes irrelevant predictors using a customized multitask feature selection procedure. The second uses only historical measurements of the target variable (temperature or precipitation) and introduces multitask nearest neighbor features into a weighted local linear regression. Each model alone is significantly more accurate than the debiased operational U.S. Climate Forecasting System (CFSv2), and our ensemble skill exceeds that of the top Rodeo competitor for each target variable and forecast horizon. Moreover, over 2011-2018, an ensemble of our regression models and debiased CFSv2 improves debiased CFSv2 skill by 40-50% for temperature and 129-169% for precipitation. We hope that both our dataset and our methods will help to advance the state of the art in subseasonal forecasting.

The influence of aggregation and statistical post‐processing on the subseasonal predictability of European temperatures

“…Obviously NGR is a simple method that uses simple predictors and assumes normality even when it is inappropriate. Some studies have demonstrated the usefulness of more advanced predictors and post-processing methods in the subseasonal-to-seasonal range (Rodney et al, 2013;Yoo et al, 2018;Hwang et al, 2019;Kämäräinen et al, 2019;Strazzo et al, 2019). Such extensions are often specific to single sources of predictability or to a fixed time aggregation.…”

“…This led to noticeable increases of skill, but also to a bias for winter cold anomalies at long lead times in regions with a skewed climatological distribution (Figure 8). A correction approach that can better handle such distributions and, for example, the multiannual variability change in Iceland, might be realized with other simple (Ferrone et al, 2017;Vigaud et al, 2017) or more advanced (Yoo et al, 2018;Hwang et al, 2019;Kämäräinen et al, 2019;Strazzo et al, 2019) post-processing methods.…”

Application of Postprocessing to Watershed-Scale Subseasonal Climate Forecasts over the Contiguous United States

“…Statistical postprocessing can be achieved through a wide range of techniques, and indeed, recent interest in climate forecast postprocessing has delved increasing into nonlinear machine learning approaches (e.g., Hwang et al 2018). We used the linear PLSR approach alluded to earlier-a components-based regression method similar to principal component regression (PCR) that combines features of principal component analysis (PCA) and multiple linear regression (Abdi 2010).…”

Machine Learning Models for the Seasonal Forecast of Winter Surface Air Temperature in North America

“…With the development of computer science, machine learning (ML) models have begun to greatly influence many fields (Reichstein et al, 2019) and have been applied to both weather forecasting (e.g., Chattopadhyay et al, 2020;Herman & Schumacher, 2018;Weyn et al, 2019) and climate prediction (e.g., Badr et al, 2014;Chi & Kim, 2017;Cohen et al, 2019;Ham et al, 2019;Hwang et al, 2019;Iglesias et al, 2015;Kämäräinen et al, 2019;Nooteboom et al, 2018;Richman & Leslie, 2012). Some newly developed ML models show great power in improving the monthly and subseasonal forecast skills of dynamic models and may even outperform dynamic models.…”

“…For example, Nooteboom et al (2018) proposed a hybrid ML model based on an autoregressive integrated moving average (ARIMA) model and an artificial neural network (ANN) to forecast monthly ENSO indices, and this model was found to outperform the ensemble mean forecast of the Climate Forecast System version 2 (CFSv2). Hwang et al (2019) constructed two ML models and showed that the ensemble of the two ML models and CFSv2 exhibit significantly better subseasonal temperature and precipitation prediction skills over NA than the CFSv2 alone. Ham et al (2019) applied the transfer learning technique to a convolutional neural network (CNN) and showed that the CNN model outperforms the current state-of-the-art dynamic forecast system in monthly ENSO predictions.…”