Geoelectric Field Forecasting with Machine Learning:A Data-Driven, Ensemble-Based Capability for Hazard Mitigation

The proposed effort is focused on developing a data-driven forecast of the physical drivers of geomagnetically-induced currents. Building on our successful demonstration of geomagnetic activity forecasts of 30 minutes or more using a classification-based approach to machine learning, we are now seeking developing a diversified framework of complementary forecasting models based on classification, regression, deep learning, and reduced-complexity physics. In addition to expanding the scope of modeling paradigms, we are also expanding the scope of our data used by our models to include measurements from solar wind monitoring spacecraft and are further enhancing the value of this new data by using realistic models of uncertainty for key solar wind parameters. Using these uncertainty models, we will develop ensemble forecasts from each of our models and will use these ensembles to provide quantified uncertainties for each forecast. Finally, we will develop a metaforecasting algorithm that intelligently combines the forecasts from individual model forecasts using the tools of Ensemble Learning. This work purposefully embraces the uncertainty and variability that is inherent within both our data sources and our models. Discoveries in the fields of data science and machine learning have revealed that when uncertainty is acknowledged and properly leveraged, forecasts from multiple independent models can be combined into a single forecast with far fewer weaknesses than that of any of its constituent forecasts. By predicting an outcome using multiple diverse base models and combining the results from this ensemble of models, the generalization error of the prediction can be significantly reduced, and its accuracy can be greatly improved, enabling us to achieve unprecedented accuracy with sufficiently long lead times to be of use to operators and decision makers.