Forecasting solar wind speed from coronal holes and active regions

Coronal holes (CHs) have long been known as one of the main sources of high-speed solar wind streams, but recent evidence suggests that active regions (ARs) also play a significant role as solar wind sources. In this study, we aim to investigate the impact of both CHs and ARs as source regions of the solar wind. Both structures can be identified in extreme ultra-violet (EUV) solar images several days before they become geoeffective. We exploit this relation to construct a model that forecasts the solar wind speed at L1. First, we accurately detect and track the evolution of CHs and ARs over time by employing a segmentation algorithm on solar images. Next, we extract features from the indicated regions in EUV images and magnetograms, such as area and location of the source regions and the corresponding magnetic field configurations. These features, along with solar wind data from previous solar rotations and the current state of the solar cycle, are assimilated over time in a data-driven model that predicts the hourly solar wind speed at L1 four days in advance. During model training, we particularly focus on preserving the distribution of observed solar wind speeds to overcome a common drawback of data-driven solar wind speed prediction models, namely the underprediction of the peak values of solar storms. By adding a suitable regularization to the loss function, we force our model to follow the physical behavior more closely, which results in a significantly improved accuracy for predicting solar storms. Finally, we use our model to draw conclusions about the physical relevance of CHs and ARs for solar wind speed models. The model's performance is evaluated through cross-validation on 14 years of data and compared to other state-of-the-art models.