Speaker
Description
In this work, we employ an attention-based deep learning approach to predict flare occurrence from multivariate time series of SHARP magnetogram features. The model takes as input active region data over varying time windows and outputs probabilistic predictions for C+-, M+-, or X+-class flare events. To capture the temporal evolution of active regions, the architecture leverages self-attention mechanisms and learnable positional embeddings, enabling it to model dependencies across time even in the presence of missing data.
We investigate how the temporal extent of the input affects forecasting skill by evaluating the model on sequences of different lengths, ranging from short snapshots to full 24-hour histories. This comparative analysis aims to determine whether longer observation windows, which offer a broader view of the magnetic evolution, allow the model to recognise precursor patterns that may not be evident over shorter intervals.
To ensure relevance to operational settings, the model is designed to accommodate missing data through a masking mechanism that guides the attention layers to focus only on valid observations. This allows the system to maintain predictive capability even in the presence of incomplete or irregular time series — a frequent challenge in space weather monitoring.
Our approach aims to support robust and flexible flare classification, and represents a step toward developing real-time forecasting tools that are resilient to data quality variability.
| Do you plan to attend in-person or online? | In-person |
|---|
