LSTM Based Short-Term Load Forecasting via Ensemble Empirical Mode Decomposition and Graph Clustering

The increasing penetration of renewable energy and the emergence of complex, non-stationary load patterns have raised new challenges for short-term load forecasting in modern power systems. In this paper, we propose a hybrid forecasting framework that integrates ensemble empirical mode decomposition (EEMD), graph-based frequency clustering, and attention-enhanced long short-term memory (LSTM), called EEMD-GC-AttLSTM. The proposed method decomposes raw load signals into intrinsic mode functions (IMFs) using noise-assisted EEMD, groups frequency-aligned components via Louvain community detection, and employs a multi-channel LSTM with temporal attention to adaptively fuse predictive features. Experiments on a real-world residential dataset demonstrate that our model outperforms the EMD-based and deep learning baselines, achieving up to 19.0% improvement in MAE and 17.6% in RMSE. These results confirm the advantage of combining data-driven decomposition, structural modeling, and adaptive feature fusion for accurate and interpretable load forecasting.