Paper Details
Abstract
The transition towards net-zero smart cities required complex and agile energy forecasting models to manage nonlinear energy consumption patterns and incorporate data from various sources in real-time. AI has emerged as a transformative tool in this domain, leveraging Deep Learning techniques to provide highly accurate and adaptive energy predictions. This study proposed two hybrid deep learning architectures combining CNN for short-term feature extraction, LSTM for long-range temporal modeling, with an Attention mechanism for dynamic time-step weighting to enhance predictive performance in both single-step and multi-step energy forecasting. Evaluated on the ENTSO-E dataset, the proposed CNN-LSTM-Att architectures significantly outperformed other baselines, with CNN-LSTM-Att.v2 achieving the best single-step forecasting results with RMSE of 106.96 MW, reducing MAPE to 1.07%, and CNN-LSTM-Att.v1 achieving the best multi-step day-ahead forecasting results with RMSE of 438.11 MW and reducing MAPE to 4.18%. These experimental results highlighted the importance of spatial-temporal feature extraction and attention-driven modeling in delivering robust forecasts for sustainable urban energy management.