New model estimates city-level carbon intensity with 9% less error

A new study by Bayer, Schiller, Pham, and Pruckner (arXiv:2607.19851) tackles a critical blind spot in urban decarbonization: carbon intensity factors—the grams of CO2 per kWh of electricity—aren't uniform across a city. Using high-resolution solar radiation data and building load profiles, the team calculated that for an exemplary summer noon, local factors within a single city ranged from 0 g/kWh (zero-carbon solar) to 316 g/kWh (fossil-heavy grid). This spatial-temporal variation means that a smart building control system or EV charging station optimizing solely on a city-wide average could misallocate energy use and miss emissions reductions.

The researchers then built a transferable surrogate model that requires only basic inputs like census grid data and local solar generation, which are typically available to municipalities and grid operators. They compared traditional decision tree methods (e.g., XGBoost) with a state-of-the-art neural network. Both approaches delivered highly accurate estimations, especially in densely built urban areas, and the surrogate model proved transferable to regions with sparse data. A case study of one building showed that neglecting local variations could cause emission estimation errors of up to 9%, underscoring the practical importance of this work for real-time building control, district heating, and EV fleet scheduling.