Multi-step hourly electrical load forecasting for a single household: A hybrid deep learning approach based on CNN, BiGRU, and Attention Mechanism

Accurate electrical load forecasting is crucial for the efficient management of modern power grids. However, the highly variable consumption patterns of single households present a significant challenge, especially for multi-step forecasting, which holds high practical value. This study addresses this challenge by proposing a novel hybrid deep learning model that combines Convolutional Neural Networks, a Bidirectional Gated Recurrent Unit, and an Attention mechanism. The proposed model was trained and tested to forecast the next hour in 6 steps, based on 24 hours of 10-minute interval data. The experimental results demonstrate that the proposed architecture, despite the complexity and granularity of the problem, achieves remarkably low error rates, with a Mean Absolute Error of 0.0393, a Mean Squared Error of 0.0044, and a Root Mean Squared Error of 0.0666. This performance is competitive with results obtained by many state-of-the-art models on single-step forecasting tasks. Consequently, this study makes a unique contribution to the field by presenting an effective and robust solution to the multi-step hourly forecasting problem, a topic less explored in the literature.