Forecasting of Kp Index with Deep Learning

The Kp index is one of the most important indicators used to monitor the level of the geomagnetic disturbances driven by the solar wind plasma under an influence of the interplanetary magnetic field. In the past, various neural networks have been used to construct models for forecasting Kp. Most models usually produce less accurate results for severe geomagnetic disturbances than for quiet times. The performance of the Kp prediction still has room for improvement due to differences of the training data and the architectures of neural networks. In general, the performance of the models based on the recurrent neural network is better than other architectures. In this study, we compared the predictions of various models, such as the fully connected long short-term memory (LSTM), the bidirectional LSTM, and the convolutional LSTM. The LSTM is able to learn the context or patterns required to make good predictions. Therefore, the LSTM can better remember the impact of previous solar wind and magnetospheric variations than the simple recurrent neural network. In our model, the correlation coefficient between the prediction and the true observed Kp can reach up to 0.95, and the root-mean-square error has been reduced to 0.38. We will discuss the problem with the model that results from including the previous Kp values as input. In summary, we have developed a model that predicts Kp one hour ahead with high degree of accuracy. To reduce the impact of severe space weather to technological infrastructures and lives on Earth, a Kp model with a high performance in predicting Kp is necessary. Keywords: Kp Index, Geomagnetic Activity, Long Short-Term Memory, Solar Wind