Solar Wind Speed Prediction via Graph Attention Network

Abstract

AbstractThe solar wind is a plasma flow formed by the expansion of the high‐temperature corona and propagates in the interplanetary space with speeds between 200 km/s and 900 km/s. Accurate solar wind speed prediction and longer lead time will help mitigate the impact of solar storms on aerospace equipment and the Earth's magnetic field. Recently, most approaches do not explicitly capture the relationships between different solar wind features, and the prediction accuracy of 96‐hr is still not good enough. This paper elaborately designs an end‐to‐end model: Graph‐Temporal‐AR model (GTA) for solar wind speed prediction. First, our framework considers each feature as the node to construct the graph structure and adopts the graph attention module to learn the complex dependencies among features. Second, our approach employs the dilated causal convolution to extend the receptive field and prolong the prediction time. Furthermore, we leverage the autoregressive model to solve the scale insensitive problem of the neural network, making our model more robust. Specifically, we combine the OMNI data measured at Lagrangian Point 1 (L1) with the extreme ultraviolet images observed by the Solar Dynamics Observatory satellite to predict the solar wind speed at L1. Compared with the baseline models, GTA obtains significant performance improvements. Through visualization, we find GTA excavates the relationships between multiply variables without domain prior knowledge, which may help us find other unknown associations in heliophysics data sets. The data and code are available from https://github.com/syrGitHub/GTA.