TAL-Net: a temporal attention LSTM framework for urban electricity and emissions forecasting in the U.S.

Urban sustainability depends critically on accurate forecasting of electricity demand and associated CO2 emissions, particularly as cities strive to meet climate targets while ensuring energy resilience. However, existing approaches often model these two aspects in isolation, missing key interdependencies that influence urban energy dynamics. This study presents TAL-Net, a novel Temporal Attention LSTM Network designed for the joint forecasting of electricity demand and CO2 emissions in complex urban energy systems. Leveraging deep learning and domain-informed feature engineering, we evaluate TAL-Net alongside four other state-of-the-art models using high-resolution data from two contrasting U.S. regions - California and Texas - characterized by differing climates, energy portfolios, and urban infrastructure. TAL-Net consistently achieves superior forecasting accuracy, demonstrating lower MAPE, MAE, and RMSE across both regions. Our findings highlight the value of attention mechanisms in capturing temporal patterns and cross-variable dependencies, offering AI-driven insights to guide urban energy planning, carbon mitigation strategies, and grid management in fast-evolving metropolitan contexts.