Comparison of Empirical and Deep Learning Models for Solar Wind Speed Prediction
In this study, we compare representative empirical models with a deep learning model for predicting solar wind speed at 1 au. The empirical models are the Wang–Sheeley–Arge–ENLIL model, which combines empirical methods with a magnetohydrodynamic model, and the empirical solar wind forecast model, wh...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IOP Publishing
2025-01-01
|
| Series: | The Astrophysical Journal |
| Subjects: | |
| Online Access: | https://doi.org/10.3847/1538-4357/ade0b4 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In this study, we compare representative empirical models with a deep learning model for predicting solar wind speed at 1 au. The empirical models are the Wang–Sheeley–Arge–ENLIL model, which combines empirical methods with a magnetohydrodynamic model, and the empirical solar wind forecast model, which uses the relationship between the fractional coronal hole area and solar wind speed. Our deep learning model predicts solar wind speed over 3 days ahead using extreme-ultraviolet images and up to 5 days of solar wind speed before the prediction date. We evaluate the models over the test period (October–December in each year from 2012 to 2020) in view of solar activity phases and the entire period. To validate the model’s performance, we use two evaluation methods: a statistical approach and an event-based approach. For statistical verification during the entire period, our model outperforms the other empirical models, with a much lower mean absolute error of 51.4 km s ^−1 and rms error of 68.6 km s ^−1 , along with a much higher correlation coefficient of 0.69. For the event-based verification for high-speed solar wind streams, our model has superior performance in most of the six metrics evaluated within a ±1 day time window. In particular, it achieves a high success ratio of 0.82, emphasizing the model’s stable performance and ability to minimize false alarms. These results show that our deep learning model has strong potential for practical application as a reliable tool for fast solar wind forecasting with its high accuracy and stability. |
|---|---|
| ISSN: | 1538-4357 |