A Novel Hybrid Approach for Load Forecasting: Multi-Head Attention Integrated LCG Model with Statistical Outlier Management
Abstract
Electrical load forecasting (ELF) is essential for effective load dispatching and future power system planning, and it has been widely explored, particularly in developed regions. Among its types, short-term load forecasting (STLF) is increasingly important for optimizing electricity usage. This study introduces a novel STLF model designed to address the non-linearity and time-dependent fluctuations in regional load data. The model integrates Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) with a multi-head attention mechanism. While CNN captures spatial features and patterns, the LSTM-GRU combination handles temporal dependencies. The attention mechanism improves the model's focus on critical data segments and enhances interpretability. The architecture is unique, featuring a parallel configuration of CNN and LSTM-GRU sub-models, each equipped with its own multi-head attention mechanism. The model's effectiveness is validated using historical load data from the Chattogram district and other public datasets. Results show that it outperforms several state-of-the-art methods, setting a new benchmark for regional short-term load prediction.