Machine Learning-Based Building Energy Consumption Prediction and Carbon Reduction Potential Assessment in U.S. Metropolitan Areas

Abstract

Building energy consumption accounts for 40% of U.S. energy usage, presenting critical challenges for urban sustainability. This paper presents a machine learning framework integrating energy consumption prediction with carbon reduction assessment across five major metropolitan areas. We analyze 50,000+ buildings from 2019-2023, combining meteorological data, building characteristics, and socioeconomic factors to develop predictive models using LSTM networks, Random Forest algorithms, and Support Vector Machines. Our framework introduces a novel carbon assessment indicator system accounting for regional grid emission factors and building-specific operational patterns. Experimental results demonstrate Random Forest algorithms achieve 8.2-12.7% mean absolute percentage error, representing 15-23% improvement over traditional methods. LSTM networks excel for buildings with complex temporal patterns. Carbon assessment reveals reduction potential of 2.8-7.2 million tons CO₂ equivalent annually, with envelope improvements and HVAC upgrades contributing 70% of total potential at implementation costs of $15-85 per ton CO₂. The framework provides scalable prediction capabilities and actionable insights for urban energy policy, supporting evidence-based interventions toward carbon neutrality goals.