GHGbench: A Unified Multi-Entity, Multi-Task Benchmark for Carbon Emission Prediction
GHGbench: A Unified Multi-Entity, Multi-Task Benchmark for Carbon Emission Prediction
Authors: Yifan Duan, Siyuan Zheng, Lihuan Li, Chao Xue, Flora D. Salim Date: 2026-05-13 Paper ID: openalex:2605.13743
Summary
GHGbench is a unified benchmark platform for carbon emission prediction that addresses data fragmentation by providing harmonized datasets for both company-level and building-level forecasting. It defines rigorous evaluation tasks, including in-distribution and cross-region transfer, while establishing baselines across multiple architectures ranging from gradient-boosted trees to tabular foundation models. The study reveals critical insights into model performance, notably highlighting the effectiveness of multimodal remote-sensing features in challenging generalization scenarios and the persistent difficulty of cross-city emission prediction.
Key Contributions
- Introduces GHGbench, a comprehensive benchmark featuring 32,000+ company records and 491,591 building records across 26 metropolitan areas.
- Demonstrates that tabular foundation models outperform gradient-boosted trees on cross-city building-emissions forecasting tasks, reaching statistical significance via paired-bootstrap testing.
- Establishes that multimodal remote-sensing embeddings provide measurable performance gains specifically when tabular-only generalization reaches its limits.
- Identifies systemic failure modes in emission prediction, including poor city-to-city transferability and the limitations of sector-factor-based estimation.
Open Questions & Future Work
Key Concepts
- ghgbench: A unified multi-entity, multi-task benchmark for company- and building-level carbon emission prediction.
Archivist Review
The paper provides a significant contribution by consolidating fragmented carbon-emission data into a unified benchmark. I approved GHGbench as both a concept and a dataset, as it provides a structured platform for climate-related time-series research. I also approved a specific open question focused on the observed failure modes in cross-city generalization, as it identifies a concrete, domain-specific research challenge for future predictive models.
Approved Concepts
- GHGbench: It establishes a standardized framework for carbon emission prediction across multiple entities and scales, addressing the fragmentation of existing datasets.
Approved Open Questions
- Cross-city emissions generalization bottleneck: This addresses a key bottleneck for climate-tech scalability where models fail to generalize to new cities or sectors without significant performance loss.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.13743
- paper_id
- 2605.13743
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-forecastingbenchmark
- architectures
-
- datasets
- ghgbench
- concept_slugs
- ghgbench
- dataset_slugs
- ghgbench
- skill
- TimeSeriesSkill
- created_at
- 2026-05-16T06:03:33Z