Integrated Data Analysis and Machine Learning-Based Interpretation of Solar Flare Events

Solar flares, driven by magnetic reconnection processes in active solar regions, present significant challenges for space weather forecasting due to their complexity and nonlinearity. This study proposes an integrated data analysis and machine learning framework of solar flare events across multiple GOES intensity classes ($\mathrm{A}, \mathrm{B}, \mathrm{C}, \mathrm{M}, \mathrm{X}$), including nonflaring instances. The dataset combines GOES X-ray flux data with NOAA’s Solar Region Summary (SRS) reports and highresolution vector magnetic features from SDO/HMI-SHARP, enabling spatio-temporal data fusion. Feature engineering involves the extraction of magnetic and statistical parameters relevant to flare genesis (e.g., R_VALUE, TOTUSJH) along with binary classification of flares, and all data is preprocessed using robust normalization. To address severe class imbalance, Synthetic Minority Oversampling Technique (SMOTE) and BorderlineSMOTE oversampling are employed exclusively to the training sets. A strict $\mathbf{7 0 - 1 5 - 1 5}$ temporal split ensures operational validity, with all Long Short-Term Memory (LSTM) models-including Bidirectional LSTM with Attention (BiLSTM+Attention) and Random Forest (RF), XGBoost uses 5 fold cross validation-trained on chronologically ordered data to prevent leakage. The SMOTE-LSTM demonstrates strong predictive performance outperforming BiLSTM+Attention and classical models (Random Forest/XGBoost). With inference times ${\lt}50 \mathrm{~ms}$ per sequence, SMOTE-LSTM demonstrates realtime operational viability. This work provides a reproducible pipeline for flare forecasting, emphasizing temporal validation and open-source implementation.