Evolution-Aware Temporal Modelling of Magnetic Complexity for Multi-Horizon M/X-Class Solar Flare Prediction

Solar flare prediction is a rare-event forecasting problem governed by nonlinear, time-dependent magnetic evolution in solar active regions. This study presents an evolution-aware framework for multi-horizon prediction of M- and X-class flares using Space-weather HMI Active Region Patch magnetic parameters, engineered temporal differences, and GOES-derived flare-history features. Active regions are represented as chronological sequences and evaluated using long short-term memory, bidirectional long short-term memory, decomposition-based recurrent networks, Transformer models, a physics-regularised decomposition model, and probability ensembles. Forecast horizons of 3, 24, and 72 h are assessed using chronological training, validation, and test splits. Probability thresholds are selected from validation data only. To address methodological robustness, the revised evaluation includes static LR/MLP/SVM baselines, feature-group ablations, moving-average kernel sensitivity, learning-curve inspection, dropout sensitivity, bootstrap confidence intervals, blocked temporal cross-validation, and McNemar tests. The validation-selected 3 h family meta-average ensemble achieved a test true skill statistic of 0.753. The tuned bidirectional recurrent model achieved 0.481 at 24 h, while the simple probability-average ensemble achieved 0.594 at 72 h. Static baselines showed that final pre-prediction states are highly competitive, especially at 24 h, but temporal/evolution-aware models improved TSS at 3 h and 72 h. Feature ablations further showed that the dominant predictive signal is horizon-dependent: recent GOES flare-history memory is strongest at short horizons, while SHARP magnetic-state variables remain central at 72 h. The results provide a cautious, leakage-conscious evaluation framework for operational space-weather applications and identify where temporal modelling adds value beyond static active-region classification.