Feature Engineering for Deep Learning to Forecast Solar Events

Within solar science, machine learning has many applications for creating useful systems that don't require full physical models. A system for reliably forecasting solar flares would allow us to protect sensitive equipment. However, scientists do not yet fully understand the causes of these events, and therefore, a prediction system based on a simple combination of known phenomena is not feasible. A system using machine learning has no such limitation, as it is possible for the model to generate accurate predictions based on a combination of factors that are too complex for a human to see. This project applies a technique known as deep learning, using sophisticated neural nets with more hidden layers than previous models, to extract complex information from the data. In addition to more sophisticated machine learning models, a method known as feature engineering can greatly improve deep learning methods. Increasing the complexity of the model can only help to a certain extent, particularly with limited data, so the application of feature engineering is needed to improve the accuracy of a model.

In feature engineering, the current feature set is analyzed to determine which features are valuable and which can be discarded. We accomplished this using statistical analysis of correlation and difference between means. Additionally, we applied scientific knowledge to generate additional features. In this case, we extracted features from magnetograms that previous research had determined were useful for flare forecasting. The addition of these features will provide the model with explicit insight, and to provide it with information already decided to be important for flare prediction, rather than relying on the model to determine those features. By leveraging feature engineering, the model isn't being developed by only using raw data, but with knowledge from previous research.