A Carbon-Aware Artificial Intelligence Framework for Achieving Net-Zero Smart Cities

Abstract

Urban areas generate approximately 70 percent of global greenhouse gas emissions, yet most cities still lack the analytical infrastructure needed to track, predict and actively reduce those emissions in real time. Existing approaches to urban carbon management are hampered by static models that cannot adapt to shifting energy mixes, incomplete cross-sector data linkages and the absence of closed-loop control mechanisms that translate carbon predictions into actionable dispatch decisions. This paper introduces the Carbon-Aware Artificial Intelligence Framework for Smart Cities (CAIF-SC), an end-to-end pipeline that integrates long short-term memory networks and gradient-boosted ensemble learning for high-accuracy greenhouse gas prediction with a reinforcement learning dispatch optimiser that translates those predictions into sector-level carbon reduction actions in real time. CAIF-SC is evaluated on a longitudinal panel dataset spanning ten years (2014–2024) across 36 cities using 20 from India and 16 from the United Kingdom using drawn from the Smart Cities Mission, the UK Net Zero Strategy and multiple satellite and sensor networks. The framework achieves a mean absolute error of 2.1 tCO₂e per annum against monitored ground-truth emissions, reduces simulated city-wide greenhouse gas output by a mean of 27.4 percent relative to business-as-usual baselines and projects a mean net-zero attainment year of 2041 across the study cohort. These results outperform five comparator methods using including static regression, standalone deep learning and expert-calibrated optimisation models using on every reported metric. Beyond performance benchmarks, CAIF-SC provides explainable sector attributions and policy-ready counterfactual scenarios, making it directly actionable for municipal climate officers.