CatBoost-enhanced convolutional neural network framework with explainable artificial intelligence for smart-grid stability forecasting

Introduction Mobile robots increasingly support inspection and emergency response in smart-grid infrastructure but require accurate, interpretable backend diagnostics. This work is proposing a hybrid model that integrates CatBoost (for tabular features) with a deep 1D-CNN (for spatial feature extraction) and integrates Local Interpretable Model-agnostic Explanations (LIME) to provide transparent, instance-level rationales. Methods We evaluate on a synthetic DSGC-based stability dataset (14 features) and externally on the IEEE PES 2018 fault-clearing corpus. The hybrid concatenates CatBoost output probabilities with a three-layer CNN feature vector, followed by dense layers (ReLU and Sigmoid). Models are trained using the Adam optimizer. Performance is reported via Accuracy, Precision, Recall, F1, confusion matrices, ROC-AUC, and LIME explanations. Results On the generated synthetic data, the hybrid achieved 98.23% accuracy (F1 = 97.56%), outperforming ANN, DNN and CNN baselines. External validation on IEEE PES 2018 yielded F1 = 97.6%. Discussion Combining gradient-boosted trees with deep convolutional features improves discrimination while and it is preserving local explainability. This way it can be supporting both grid operations and stability-aware robotic mission planning. Future work will extend to multiclass/regression settings and compare XAI methods (e.g., SHAP) alongside additional tabular learners (XGBoost/LightGBM).