Lifecycle Visualization and Prediction of AI Electricity Footprint for Green Computing Ethics

This paper proposes an integrated spatiotemporal framework that quantifies, visualizes, and forecasts the electricity- and carbon-footprint of AI workloads across their full lifecycle. Power use is decomposed into computing, transmission, and cooling components; GIS-based mapping links these loads to regional supply capacity, losses, and time-varying grid-emission factors. A hybrid time-series + machine-learning model then predicts short- to long-term supplydemand balance. Tests on 12 global data centres (180 days) reveal pronounced regional and diurnal heterogeneity: peak loads reach 47.3 kW per rack in North America, and short-term forecasts achieve a Nash-Sutcliffe efficiency of 0.936. Carbon intensity ranges from 1.40 kg CO per GPUh in Europe to 2.89 kg CO per GPUh in Asia-Pacific. Scenario analysis indicates emission-reduction potentials of 28.4 % through spatial scheduling, 41.8 % via temporal shifting, and 9.7 % via load balancing. The methodology provides a quantitative basis for greener AI resource allocation and policy design.