Evaluation of Kp and Dst Predictions Using ACE and DSCOVR Solar Wind Data

Abstract

Models that predict the Kp and Dst indices are evaluated using solar wind data at L1. The models consist of ensembles of neural networks that have been developed using ACE Level 2 data from the period 1998–2015. The use of ensembles is motivated by the difficulty of generating functions that generalize well in regions of the input-output space that are poorly sampled, which typically occurs during stronger events. Since the launch of the DSCOVR spacecraft, providing measurements from about August 2016, new and independent data have become available to test the models. ACE Level 2 data for almost the same period are also available, representing another independent set collected after 2015. We evaluate the models using plots and various statistical measures. We also study the performance of the predictions for lead times up to 3 hr. The results show that the models perform better when using ACE or cleaned DSCOVR data compared to the real-time DSCOVR data and that lead times of L1-Earth travel time plus a maximum of 1 hr are possible. As the models use only solar wind for their inputs, and the temporal dynamics of Kp and Dst are very different, we see significant differences in the error distributions that we believe are related to long-term changes in Dst that are not captured by the Dst prediction model.