Item

Abstract

Improving our ability to forecast the weather and climate is of interest to all sectors of the economy and government agencies. However, accurate forecasts for the subseasonal regime are lacking due to the chaotic nature of the weather. We have recently developed a machine learning based adaptive bias correction (ABC) method combining state of the art dynamical forecasts with observations as well as utilizing regression models. We applied this method to probabilistic precipitation and temperature forecasts with a forecasting lead of 3 to 4 weeks over the continental United States from 1999 to 2016. Using the ranked probability skill for comparison, we can show that the skill of the ABC debiased forecasts is higher than in the traditional debiased forecasts for both space and time averaged skill. Another method to further improve the skill is to combine several different dynamical models into a multimodel ensemble e.g., by weighting all model forecasts equally. However, a better alternative to the latter is to weigh model forecasts based on their skill. We recently developed an approach for adaptive ensembling of numerical weather predictions using a machine learning tool called sequential online learning algorithms. As real-time weather outcomes are observed, the algorithm adjusts the weights of each model member according to its past performance. We evaluated the skill of precipitation and temperature forecasts forecasting at 3 to 4 weeks lead time over the United States from 1999 to 2016 and see that our proposed online learning algorithm can outperform uniform ensembles for probabilistic forecasts.