Capstone Presentation: Performance of Long Short-Term Memory Neural Networks for Geomagnetic Field Modeling
 

Among the effects of geomagnetic storms are hazardous geomagnetically induced currents (GICs), which impact infrastructure, including power systems. The MAGICIAN team, a collaboration between the University of Alaska, Fairbanks and the University of New Hampshire, has found promising results in addressing the problem of predicting GICs using machine learning techniques, modeling local geomagnetic disturbances and their relationship to conditions in the space environment. Blandin et al. (2022) have trained Long-Short Term Memory neural network models predicting the north-south geomagnetic field recorded at four magnetometer observatories in Alaska with the long-term goal of predicting GICs. An additional task addressed in this project was to sample new model parameters to determine whether the models' Heidke skill score metric could be increased and to observe the effect of the solar wind and interplanetary magnetic field (IMF) input variables driving the models. We tested two methods for sampling new parameters: tree-structured Parzen estimators (TPE) and an algorithm incorporating Hyperband together with Bayesian optimization (BOHB). We used a method for testing combinations of data inputs by appending new data columns in succession depending on the resulting performance. We found model parameters resulting in improved stability during training and parameters resulting in increased Heidke skill score values; the resulting skill scores and Pearson correlations are commensurate with the original models. Successively testing input columns showed results consistent with influence primarily from the solar wind speed and IMF vector.