Improvement of Solar Flare Prediction by an SVM with an Optimized GA

Abstract We have elaborated a machine learning method to improve existing models for predicting solar flares based on images of solar surfaces. It is pointed out that if a very strong solar flare were to occur, it could have serious damages on modern society with a profound electrical and electronic civilization. However, such intense solar flares are very rare, with their classification extremely unbalanced and therefore difficult to predict. In order to cope with this problem, we have been proposing a method that combines machine learning for a classification of Support Vector Machines (SVM) and a search with Genetic Algorithms (GA). In this report, we further extend this method by using an algorithm, which we name as Case Based automatically Restartable Good combination searching GA (CBRsGcmbGA), so as to improve both the learning function and the input data. It should be noted that even the recent prevailing deep neural networks, for example, could not automatically and evolutionarily search for input data in training. The effectiveness of this approach was demonstrated in the prediction of X-class solar flares, achieving an accuracy score (TSS) of more than 85% with a maximum of 91.2%. Otherwise, a type of GA combined with conventional SVM (CBGALO, Case Based GA integrated with Local Optimizer as SVM) achieved up to 84%. Other approaches in the same environment achieved less than 75%.