Integrating Chaotic Dynamics and Machine Learning for Medium-term Drought Prediction

Drought prediction remains challenging due the multiple types, underlying chaotic behavior in meteorological processes and complex spatio-temporal dynamics. However, our existing understanding of chaotic dynamical system behavior and machine learning can provide techniques and insights that can help improve drought forecasting at the watershed scale. In this talk, we present our approach to develop a medium-term (i.e., weeks to months ahead) drought prediction model by leveraging toy dynamical systems modeling, machine learning, and high-resolution climate data. To do so, our initial model is a reservoir computer (RC) that has been trained on synthetic data based on Vano’s chaotic four-compartment Lokta-Volterra system. Our model’s goal is to predict future representations of the functions describing the Vano variables. Game theory and Shapley Additive Explanation Values are used to identify the most influential inputs for prediction. We leverage the explanations to support development of RCs to predict drought for exemplar watersheds in the Western United States based on NCAR/USGS’s CONUS404 high resolution hydroclimate reanalysis dataset. Preliminary findings suggest a graph-based RC outperforms a single RC of equivalent computational complexity. Future work will consider several meteorologic drought indicators, expand the spatial extent of the model, and explore model explainability to improve the approach’s fidelity. This work enables us to use machine learning to improve the accuracy and reliability of medium-term predictions for chaotic systems such as drought and thus help mitigate potential impacts from these events.