Scholarlink

<p>This paper compares Transformer architectures (Bisection Transformer, Bidirectional Encoder Representations from Transformers (BERT), Robustly Optimized BERT Approach (RoBERTa)), Generative Adversarial Networks (GAN), deep learning (Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU)), and Nonlinear AutoRegressive Moving Average with eXogenous (NARMAX) in forecasting the Disturbance storm-time (Dst) index using solar wind parameters. The purpose of this paper is to emphasize the importance of selecting a specific model for the 5- to 100-step-ahead forecasting horizon across various algorithmic approaches in operational space weather forecasting and geomagnetic storm prediction. In these tests, NARMAX has the lowest Root Mean Square Error (RMSE) of 14.723 at a horizon of 5 for short-term forecasting, but this decreases at longer horizons. In contrast, GAN exhibits consistent performance with minimal decline, increasing RMSE by 7.5% from horizons 5 to 100. Meanwhile, RoBERTa outperforms in the medium-term, achieving its best performance between horizons 20 and 50. Statistical analysis shows that the Transformer architecture outperforms deep learning at longer time horizons.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Forecasting horizon Dst index based on solar wind data using deep learning and generative artificial intelligence

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Wihayati,
  • Hindriyanto Dwi Purnomo,
  • Suryasatriya Trihandaru

摘要

This paper compares Transformer architectures (Bisection Transformer, Bidirectional Encoder Representations from Transformers (BERT), Robustly Optimized BERT Approach (RoBERTa)), Generative Adversarial Networks (GAN), deep learning (Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU)), and Nonlinear AutoRegressive Moving Average with eXogenous (NARMAX) in forecasting the Disturbance storm-time (Dst) index using solar wind parameters. The purpose of this paper is to emphasize the importance of selecting a specific model for the 5- to 100-step-ahead forecasting horizon across various algorithmic approaches in operational space weather forecasting and geomagnetic storm prediction. In these tests, NARMAX has the lowest Root Mean Square Error (RMSE) of 14.723 at a horizon of 5 for short-term forecasting, but this decreases at longer horizons. In contrast, GAN exhibits consistent performance with minimal decline, increasing RMSE by 7.5% from horizons 5 to 100. Meanwhile, RoBERTa outperforms in the medium-term, achieving its best performance between horizons 20 and 50. Statistical analysis shows that the Transformer architecture outperforms deep learning at longer time horizons.