Authors: Aatiya Ali (Georgia State University), Viacheslav Sadykov (Georgia State University), Berkay Aydin (Georgia State University), Benedict Mervyn (Georgia State University)

Solar Energetic Particle (SEP) events, specifically solar proton events (SPEs), pose substantial radiation risks to spacecraft systems and human exploration missions. Relativistic electrons propagate faster than protons and typically arrive at Earth earlier, making them a promising precursor for short-term SPE forecasting. We leverage electron flux from ACE/EPAM and SOHO/EPHIN alongside GOES soft X-ray (SXR) data to predict S1-level SPEs (≥10 pfu at ≥10 MeV) at Earth. Here, we focus on the 15, 30, and 60-minute lead times and 24- and 72-hour observation windows. We assess Slim-TSF (a classical statistical-windowing ensemble) against an encoder-only Neural Transformer to evaluate self-attention efficiency in capturing temporal dependencies during particle transport. Slim-TSF proved highly resilient to extended (72-hour) observation windows, maintaining peak performance (TSS >0.85), and emphasizing lower-energy EPAM channels as important predictors. When GOES SXR data are included, Slim-TSF yields an optimal true positive (TP) rate of 91.8% and a true negative (TN) rate of 99.2%. Conversely, the Transformer suffers from performance degradation over longer observation windows, dropping its TP rate from 82.0% to 70.5% when excluding SXR indicators. However, integrating SXR data anchors the Transformer’s attention mechanism, boosting its 72-hour window TSS from ~0.4 to above 0.6, and producing an 82.0% TP rate from 70.5% driven primarily by the EPHIN 1.08 MeV electron channel data. Ultimately, we find Slim-TSF to be more resilient and reliable for operational forecasting compared to the Transformer, which remains overly sensitive to long-window background noise and requires additional flare data to match performance. These results establish the structural trade-offs between a statistical-windowing ensemble and a complex Transformer architecture, emphasizing the considerations essential for developing and delivering reliable, efficient short-term SPE forecasting in support of the Artemis missions and future deep-space exploration.