Leveraging Interpretable Machine Learning Methods for Subseasonal Precipitation Forecasts in Western United States

Improved skill for longer-lead forecasts of regional precipitation has the potential to inform state and local water managers on decisions related to drought preparedness and response planning for weather and climate extremes. The value of improving forecast skill at lead times of weeks to months over the semi-arid western United States is high, especially in the light of recent multi-year droughts punctuated with major flooding episodes. The skill of most currently available forecast systems degrades beyond the weather time horizon, rendering them suboptimal for decision-making. This study is motivated by this quest for improvements in subseasonal prediction using innovations in forecast methodologies.

Here, we develop experimental forecasts of subseasonal precipitation (15-42 days lead time) leveraging a set of discrete predictors, such as North Pacific circulation regimes, Madden-Julian Oscillation (MJO), etc. We explore the feasibility of training a suite of machine learning (ML) approaches (random forests, gradient-boosted decision trees, and neural networks) to improve forecast performance by providing additional pathways for non-linear process interactions between the identified predictors. The forecast systems’ performance is tested over the hydrologic basins (HUCs) of interest over the 17 western states. In addition, we provide a comparative skill assessment of the developed ML models with the top-performing models identified from the most recent S2S Climate Forecast Rodeo Competition (https://www.usbr.gov/research/challenges/forecastrodeo.html) organized by the US Bureau of Reclamation (USBR) and the U.S. National Integrated Drought Information System (NIDIS). We further discuss how these forecast tools need not be viewed as black boxes and how interpretability methods may be used to identify underlying physical processes helping in building trust for the generated predictions.