Hybrid CNN-LSTM model for urban energy load forecasting with IGA-XAI for smart grids: Peak and off-peak variability insights
Sarowar Morshed Shawon, Shah Nawaz Haider, Arnab Barua, Steve Austin, Ifaz Ahmed Adan, Mohammad Shahadat Hossain, H.T. Zubair
Abstract
Urban areas face time-specific stochastic energy demand pattern, complicating the distribution of energy efficiently, keeping costs in sight. Present approaches seldom consider the intricate and multi-scaling dynamics of urban energy requirements specially during the Peak and Off-peak periods, resulting in inefficiencies and potential overloading of energy systems. To solve this, this study proposes a hybrid 1D-CNN-LSTM model to forecast short to medium-term electrical load across three critical daily interval such as Day Peak, Evening Peak and Off-Peak in urban context. Unlike most state-of-the-earth models, limited to 24-72 hours, the proposed model achieves class leading accuracy in day-ahead peak and off-peak load forecasts while maintaining robust performance for up to 7 days in advance. The proposed 1D-CNN-LSTM hybrid model, enabled by Integrated Gradients Attribution (IGA) based eXplainable AI (XAI) was capable of accurately forecasting the intricacies in the load patterns achieving class leading performance with R² of 0.9602, 0.948 and 0.9464 for Day Peak, Evening Peak and Off-Peak, respectively, for the city of Chattogram. The model was also tested for two other major metropolitan cities (Dhaka and Sylhet), proving its generalization capability. To further test robustness, Gaussian noise (5-20%) was added to the meteorological inputs, and the CNN-LSTM model demonstrated resilience with only a modest 2.5% average accuracy drop across all three peak demand periods. The proposed method enhances urban energy management by enabling more precise load forecasting, supporting smarter grid operations, improved reliability and progress towards sustainable energy systems.