Interpretable Machine Learning applied to Seasonal Forecasting of Western US Precipitation

Abstract:

An immediate barrier to deploying machine learning in seasonal forecasting is the limited sample size of observational data required for robust model training. To circumvent this issue, we explore training machine learning on large climate model ensembles (perturbed initial condition experiments) intended to provide multiple physically consistent realizations of the relevant teleconnections. After training various machine learning models on thousands of seasons of climate model simulations, out-of-sample seasonal forecast skill is examined across the historical observational record (1980-2020). Tasked with predicting the probability of wide-spread seasonal precipitation anomaly clusters, machine learning approaches tested include Random Forests, XGBoost, neural networks, and Long short-term memory (LSTM) networks. We demonstrate that an ensemble-based approach leveraging skill from each of these methods is capable of competing with or out-performing several operational dynamical seasonal forecast models. Lastly, we demonstrate the interpretability of this approach in terms of: (1) the general importance and interactions between predictor variables (2) the relative weightings of predictor variables for contributing to each individual forecast outcome.