Prediction of solar flares is critical for minimizing the impact of space weather on communication and power systems. This study explores the application of machine learning models—Long Short-Term Memory (LSTM), Random Forest (RF), and XGBoost (XGB)—to forecast solar flares using multivariate time series data derived from SHARP and GOES. The experimental setup systematically varies two key temporal parameters: window size (12, 24, and 48 hours) and prediction lag (12, 24, and 48 hours). To address the strong class imbalance in flare data, cost-sensitive learning is incorporated using a sample-based class weighting strategy. Results demonstrate that the LSTM model achieves the best performance when using a short 12-hour window and a 12-hour prediction lag, reaching 89% accuracy, a True Skill Statistic (TSS) of 0.6949, and an F1-score of 0.6873. These findings emphasize the importance of short-term temporal dependencies and class imbalance mitigation for improving solar flare forecasting performance.
Paper
Full text
Solar Flare Prediction Using Multivariate Time Series and Cost-Sensitive Machine Learning
Semantic Scholar · Environmental Science · 2025
Abstract
Prediction of solar flares is critical for minimizing the impact of space weather on communication and power systems. This study explores the application of machine learning models—Long Short-Term Memory (LSTM), Random Forest (RF), and XGBoost (XGB)—to forecast solar flares using multivariate time series data derived from SHARP and GOES. The experimental setup systematically varies two key temporal parameters: window size (12, 24, and 48 hours) and prediction lag (12, 24, and 48 hours). To address the strong class imbalance in flare data, cost-sensitive learning is incorporated using a sample-based class weighting strategy. Results demonstrate that the LSTM model achieves the best performance when using a short 12-hour window and a 12-hour prediction lag, reaching 89% accuracy, a True Skill Statistic (TSS) of 0.6949, and an F1-score of 0.6873. These findings emphasize the importance of short-term temporal dependencies and class imbalance mitigation for improving solar flare forecasting performance.