A Comparative Analysis of Bi-LSTM and XGBoost for Time-Series Classification in Power System Stability using SMOTE and Focal Loss

Abstract

The rapid deployment of renewable energy sources has raised several issues regarding grid stability. Renewable energy sources differ from traditional energy generation due to their volatile nature and dependence on climatic factors. Therefore, accurate predictions of smart grids' stability are essential for efficient energy management. Current approaches do not incorporate dynamic aspects into consideration. The purpose of this paper is to overcome this limitation through the development of a time series-based framework for smart grid stability detection. A bidirectional long short-term memory model is created to capture both short-term and long-term relationships of the grid frequency and power signal. A hybrid method is proposed combining the Synthetic Minority Oversampling Technique and focal loss to improve results for imbalanced datasets. An extreme gradient boosting model is trained based on flattened temporal sequences and statistical feature descriptions. The experimental findings indicate that the suggested framework demonstrates high predictive performance, with XGBoost achieving the best accuracy, while BiLSTM is effective for capturing temporal patterns and maintaining high stability in classification recall.