Forecasting Continuum Intensity for Solar Active Region Emergence Prediction using Transformers

TL;DR

This work investigates Transformer-based architectures for forecasting solar active region emergence from continuum intensity signals. By conducting a factorial ablation on a Conv1D front-end and an Early Detection attention loss, the study demonstrates that a timing-aware EarlyDetect Transformer without temporal smoothing delivers true advance warnings, achieving an overall RMSE of 0.1189 and an average lead time of -4.73 h, outperforming an LSTM baseline. The results reveal a trade-off between predictive timing and statistical smoothness, with high sensitivity introducing variance but yielding operationally valuable early warnings. These findings offer a path toward deployment-ready space weather forecasting systems that prioritize timely detection of precursor signals over smooth but late predictions.

Abstract

Early and accurate prediction of solar active region (AR) emergence is crucial for space weather forecasting. Building on established Long Short-Term Memory (LSTM) based approaches for forecasting the continuum intensity decrease associated with AR emergence, this work expands the modeling with new architectures and targets. We investigate a sliding-window Transformer architecture to forecast continuum intensity evolution up to 12 hours ahead using data from 46 ARs observed by SDO/HMI. We conduct a systematic ablation study to evaluate two key components: (1) the inclusion of a temporal 1D convolutional (Conv1D) front-end and (2) a novel `Early Detection' architecture featuring attention biases and a timing-aware loss function. Our best-performing model, combining the Early Detection architecture without the Conv1D layer, achieved a Root Mean Square Error (RMSE) of 0.1189 (representing a 10.6% improvement over the LSTM baseline) and an average advance warning time of 4.73 hours (timing difference of -4.73h), even under a stricter emergence criterion than previous studies. While the Transformer demonstrates superior aggregate timing and accuracy, we note that this high-sensitivity detection comes with increased variance compared to smoother baseline models. However, this volatility is a necessary trade-off for operational warning systems: the model's ability to detect micro-changes in precursor signals enables significantly earlier detection, outweighing the cost of increased noise. Our results demonstrate that Transformer architectures modified with early detection biases, when used without temporal smoothing layers, provide a high-sensitivity alternative for forecasting AR emergence that prioritizes advance warning over statistical smoothness.

Figures (7)

• Figure 1: Schematic overview of the end-to-end machine learning pipeline for continuum intensity decrease detection during the emergence of ARs. The system extracts central tiles from the tracked SDO/HMI magnetic flux ($\Phi$) map cut-outs, combining them with acoustic power maps to form feature tensor $X$. Sliding windows of length $W = 110$ are extracted with $P = 12$ prediction targets. An encoder-based Transformer predicts continuum intensity evolution $\hat{Y}$. Training utilizes an emergence-aware loss function optimizing MSE, early detection rewarding, and derivative loss.
• Figure 2: The factorial ablation study design. We compare two main architectures (Standard Transformer vs. Early Detection Transformer) and orthogonally evaluate the impact of a temporal 1D convolutional (Conv1D) front-end, resulting in four distinct model configurations: Baseline, Baseline+Conv1D, EarlyDetect, and EarlyDetect+Conv1D.
• Figure 3: Comparison of predicted and actual evolution of the continuum intensity of AR13183 for seven central tiles (indicated in the right bottom corner). Each tile includes five subplots: (1) Main intensity forecast with all models overlaid, (2) Observed derivative, (3) LSTM derivative, (4) EarlyDetect derivative, and (5) Absolute error for all models. The lower-right corner includes two 9x9 continuum maps showing the AR at the start and end of the plotted window. The presented results are denormalized to physical units and reflect variations relative to the quiet Sun, which is considered as background.
• Figure C1: Comparison plot for AR 11698. This figure illustrates the high-variance behavior reported in Table \ref{['tab:ar11698_metrics']}. While the EarlyDetect model achieves a mean timing of -9.75h (early) on this active region, the visual forecast is volatile, characterized by sharp pre-emergence drops in specific tiles (e.g., Tile 49). The LSTM, while smoother, lacks this aggressive early sensitivity.
• Figure C2: Comparison plot for AR 11726. This region represents a consistent success case for the Transformer architecture. The EarlyDetect predictions (red line) are visibly shifted to the left of the observed intensity drop (black line) across the active tiles, confirming the quantitative finding of a -9.40 hour mean advance warning. Note: This AR does have missing values in the data towards the end of the selected timeline, shown by the straight line from 4-24 to 4-25.