Mitigating concept drift challenges in evolving smart grids: An adaptive ensemble LSTM for enhanced load forecasting
Abdul Azeem, Idris Ismail, Syed Sheeraz Mohani, Kamaluddeen Usman Danyaro, Umair Hussain, Shahroz Shabbir, Rahimi Zaman Bin Jusoh
Abstract
This paper tackles the challenge of concept drift (CD), where data patterns evolve over time, hindering the accuracy of traditional forecasting models in smart grids. The study proposes a novel online LSTM ensemble learning framework integrated with dynamic feature engineering and attention-based mechanism for electrical load forecasting. This framework addresses CD by leveraging the strengths of ensemble learning to handle various drift types, while incorporating real-world data complexities. The proposed approach significantly outperforms existing techniques, including Random Forest, Support Vector Machine, and conventional LSTM models, across diverse datasets and CD phases. Compared to baselines, our framework achieves improvements in R² score ranging from 4.29 % before drift to a remarkable 55.61 % during concept expansion to evolution. These results showcase the superior accuracy, reliability, and adaptability of our proposed AE-LSTM framework for energy consumption forecasting in smart grids, paving the way for a more efficient and reliable future.