9B.2 Using Machine Learning to Improve Sub-Seasonal to Seasonal Prediction (S2S)

Wednesday, 15 January 2020: 1:45 PM
Richard Garmong, University of Georgia, Athens, GA; and R. Bolinger and R. S. Schumacher

Sub-Seasonal to Seasonal Climate (S2S) prediction is an especially challenging temporal period to forecast. However, it is a very important period to stakeholders as significant decisions for water resources need to be made. The lack of success in prediction using traditional methods presents an opportunity to use machine learning techniques in order to improve forecasts. A Random Forest Classifier was used to ingest multiple data types including dynamical ensembles, teleconnections, Madden-Julian Oscillation (MJO), El Niño Southern Oscillation, and previous month precipitation. This model was trained on a 1982 to 2010 period for each of the sub-basins of the Upper Colorado River Basin. The forecast was for a 0.5 month lead for precipitation. The model was then tested on a November 2017 to June 2019 period and was found to have considerable skill across all but one sub-basin. When successful and unsuccessful forecasts were compared, the MJO was found to have a statistically significant influence on the forecast skill in multiple basins. These results suggest that a Random Forest Classifier may be a viable option for improving S2S prediction and that the MJO has an effect on model predictability in the Upper Colorado River Basin.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner