The increased sophistication of smart grids and integration of renewable energy has only increased the complexity of monitoring electrical faults and making accurate and real-time fault diagnosis essential for network stability. The current prevailing methods are generally typified by low degrees of robustness, high degrees of false alarms, learning of redundant features and low fault impact predictability under operating conditions which are dynamic in nature. To address these issues, this research proposed an Adaptive Hierarchical Attention Fault Classifier Network (AHA-FCNet) and a Hierarchical Fault Impact Prediction Network (HFIP-Net). The first step is to improve the quality of the signal by normalizing and smoothing the signal and the hybrid feature selection is used to remove the irrelevant features in the voltage, current, and power signal. Two-stage classification of fault/non-fault state and a classification of line-to-ground (LG), line-to-line (LL), double line-to-ground (LLG) or three-phase faults (LLL) faults based on attention-guided deep learning are carried out by AHA-FCNet. HFIP-Net also envisions a voltage drop, current variation and power loss through hierarchical regression learning. The experimental results showed that the highest performance was achieved with the highest classification accuracy of 98.30%. For prediction, HFIP-Net achieved MAE of 0.85, RMSE of 1.39, and R² of 99.45. The proposed architecture is a smart and intelligent plan of real-time smart grid fault detection and stability improvement.