Jiajia Liu (Sheffield): Application of Machine Learning Techniques in CME Arrival Time Prediction
Coronal Mass Ejections (CMEs) are one of the most violent eruptions in the Solar system. They can cause severe disturbances of the Earth's magnetosphere and further affect the operation and working of high-tech facilities like spacecraft, are serious health hazards for astronauts, can cause disruption in functioning of modern communication systems (including radio, TV and mobile signals), navigation systems, and affect the working of pipelines and power grids. Fast and accurate prediction of CME arrival time is then vital to minimize the losts CMEs may cost when hitting the Earth. Here, we present a new model for (partial-) halo CME Arrival Time Prediction Using MAchine learning methods (CAT-PUMA). Via detailed analysis of the CME features and solar wind parameters, we build the model taking advantage of 182 previously observed geo-effective (partial-) halo CMEs using the Support Vector Machines (SVMs). The model results show that CAT-PUMA is accurate and fast. In particular, i) predictions using the model on a test set, that is unknown to our model, shows a mean absolute prediction error of less than 6 hours of the CME arrival time. Comparisons with other models reveal that CAT-PUMA has a more accurate prediction for 73% events; ii) the prediction can be carried out within minutes after providing the necessary input parameters of a CME. The engine of the CAT-PUMA, with a very user-friendly User Interface (UI), is provided at https://github.com/PyDL/cat-puma. This cross-platform tool allows the community to perform their own fast and accurate predictions of CME arrival time using the CAT-PUMA model.
Further, we have developed a Convolution Neural Network using a single image of the SOHO LASCO C2 observations as input to make arrival time predictions of CMEs. It has been shown that, the average prediction error is around 12.5 hours, comparable to the average performance of previous studies on this subject. A comparison between our CNN model and traditional methods available at the NASA CME Scoreboard has been performed. It is revealed that, in 63% events, our model gives less prediction errors than the average of all traditional models used.