Large language models have transformed artificial intelligence, but every query they answer draws electricity from a power grid whose carbon intensity can swing dramatically depending on where and when the computation runs. A new open-access study published in Discover Computing on 24 October 2026 presents CarbonBalance, a framework that couples machine-learning forecasts of grid carbon intensity with a scheduling engine capable of steering delay-tolerant AI inference tasks toward cleaner energy windows. In the scheduling scenarios evaluated by the authors, the framework reduced total simulated emissions from 1,503 gCO₂ under a baseline policy to 878 gCO₂ under the selected schedules — a saving of 625 gCO₂, or 41.6% — while critical workloads maintained zero scheduling delay.
The research, led by Sanjit Kumar Dash with Subhankar Poddar, Mohammed Altaf Ahmed and Qutubuddin Mohammed, addresses a gap the authors identify in existing carbon-aware computing work. Prior approaches have largely concentrated on datacenter operations, container scheduling, or training workloads, and forecasting models have typically been developed separately from the scheduling decisions that would consume their outputs. CarbonBalance integrates the two: a hybrid forecasting pipeline predicts carbon intensity, and a multi-objective optimization scheduler uses those predictions, together with energy profiles for different sizes of language models, to decide when and where non-urgent inference tasks should execute.
The forecasting pipeline rests on two complementary modeling families. Prophet, a time-series decomposition method that separates data into trend, seasonality and noise components, is fitted separately for individual countries using their historical annual carbon-intensity observations and generates three-year-ahead forecasts. Alongside it, tree-based ensemble models — LightGBM, Random Forest and Gradient Boosting — are trained globally across all available countries on an engineered multivariate dataset containing temporal, energy, weather and regional features. The final prediction is a weighted ensemble of the three tree models, with each weight calculated from its validation R² value divided by the sum of the three validation R² values, ensuring the weights sum to one.
The data underpinning these models come from two public Kaggle sources: the World Energy Consumption dataset, which supplies the carbon_intensity_elec variable in grams of CO₂ per kilowatt-hour, and a weather dataset providing hourly and daily temperature, humidity, wind speed, solar irradiance and precipitation measurements. After preprocessing — country-name standardization, annual aggregation of weather observations, median imputation of missing values, and exclusion of countries with fewer than five yearly records — the modelling dataset contained 5,165 records spanning 234 country and region entities from 2000 to 2022, with 38 engineered features. The split was chronological: 4,159 samples from 2000 to 2017 for training and 1,006 samples from 2018 to 2022 for testing, with three-fold time-series cross-validation during model development to avoid look-ahead bias.
The performance figures reported for the forecasting stage are striking. The final weighted ensemble achieved a test R² of 0.9969, meaning it explained roughly 99.69% of the variance in carbon intensity in the test set, with a mean absolute error of 6.84 gCO₂/kWh. The individually tuned LightGBM model reached a test R² of 0.9962 and a test MAE of 8.55 gCO₂/kWh, with hyperparameters optimized through 30 Optuna trials using three-fold time-series cross-validation; the best trial recorded a validation RMSE of 18.3208. SHAP interpretability analysis identified the three-year rolling mean of carbon intensity as the most influential feature, with a correlation coefficient of 0.993 with future carbon intensity, reflecting strong autocorrelation in energy systems as infrastructure changes gradually. Regional-deviation features also ranked highly, while weather feature importance varied geographically: wind and solar features dominated in renewable-rich areas, and temperature mattered most in fossil-fuel-dependent regions because of heating and cooling demand.
An important caveat runs through the study: the carbon-intensity target is a dataset-provided historical annual estimate, not a real-time grid measurement. Real-time carbon-intensity services such as Electricity Maps or WattTime were not used as input sources in the current implementation, and the authors flag integration with such services as a direction for future deployment. The scheduling evaluation is likewise simulation-based rather than a live deployment. The per-inference energy consumption values used for the four LLM model classes are predefined representative estimates — workload-energy assumptions rather than hardware-specific measurements from the Kaggle T4 GPU environment where the framework was implemented — and actual hardware profiling is deferred to future validation work.
The scheduling engine itself formulates a multi-objective optimization problem that minimizes a weighted sum of emissions, delay penalties and resource costs, subject to constraints that every workload completes by its deadline and that resource usage stays within capacity at every time step. Adjustable weights α, β and γ set the trade-off between environmental, performance and efficiency objectives according to operational priorities. Critical tasks — those with real-time constraints or no latency tolerance — bypass the optimization engine entirely and run immediately under a default low-latency policy, so service-level guarantees are never traded for carbon savings. The authors note that in a new deployment these weights could be calibrated against historical workload traces, with a systematic sensitivity analysis identified as part of future validation.
The headline 41.6% figure comes from five workload scenarios evaluated across five candidate regions. The baseline policy is defined as immediate execution of each workload at a reference carbon intensity of 400 gCO₂/kWh; savings are the difference between baseline emissions and emissions under the selected carbon-aware schedule, calculated as E = N × e × CI for N inferences consuming e kilowatt-hours each. A worked example illustrates the arithmetic: for a medium workload representing a 7–13B-parameter model with 1,000 inferences at an adjusted energy consumption of 0.000578 kWh per inference, baseline emissions are 231.2 gCO₂ at 400 gCO₂/kWh, while scheduled emissions fall to 135.3 gCO₂ at the selected carbon intensity of 234 gCO₂/kWh — a saving of 95.9 gCO₂, or 41.5% for that workload. Savings scaled with urgency tolerance: low-urgency tasks achieved the largest average reduction at 25.3% while absorbing 8.2-hour delays, medium-urgency tasks realized 18.7% savings with 3.6-hour delays, high-urgency workloads reduced emissions by 7.2% with only 0.8 hours of delay, and critical tasks ran at zero delay with zero reported carbon reduction. The authors explicitly caution that these percentages represent the evaluated scheduling scenarios rather than a universal per-inference saving.
The study also stops short of claiming superiority over competing systems. Although CarbonBalance relates to approaches such as carbon-aware Kubernetes container scheduling and datacenter-level temporal and spatial workload shifting, the current work includes no direct experimental benchmark against an external scheduler; the framework was evaluated using its own scheduling policy and workload configurations, with controlled comparisons against threshold-based carbon dispatching named as an important direction for future evaluation. The broader context the authors sketch is sobering: grid carbon intensity can exceed 800 gCOâ‚‚/kWh in fossil-fuel-dependent regions while falling below 50 gCOâ‚‚/kWh in renewables-rich ones, and AI serving infrastructure has largely been built around single-metric optimization of response time, ignoring the environmental consequences of execution timing and geographic placement.
For the field, CarbonBalance offers a proof of concept that carbon-intensity forecasting and latency-sensitive inference scheduling can be unified in one pipeline, with MLflow tracking experiments for reproducibility and SHAP analysis exposing which features drive each prediction. The authors outline several next steps: incorporating real-time carbon data for live adjustments, extending support to batch training and continuous-learning workloads that have more energy flexibility, adding reinforcement learning for dynamic policy optimization, tightening integration with cloud orchestrators such as Kubernetes for carbon-aware auto-scaling, and enabling user-defined sustainability policies and organizational carbon budgets. Within the limits of its simulation-based evaluation and historical annual data, the framework demonstrates the feasibility of making carbon awareness a first-class constraint in AI workload dispatch — a step toward AI systems that weigh not only how fast they answer, but how cleanly.
Subject of Research: A hybrid machine-learning framework that forecasts grid carbon intensity and schedules AI inference workloads to reduce simulated carbon emissions.
Article Title: Carbonbalance framework for carbon intensity forecasting and carbon aware scheduling of ai workloads
Article References: Dash, S. K., Poddar, S., Ahmed, M. A., & Mohammed, Q. (2026). Carbonbalance framework for carbon intensity forecasting and carbon aware scheduling of ai workloads. Discover Computing, 29(1), Article 802. https://doi.org/10.1007/s10791-026-10708-7
Image Credits: AI Generated
DOI: 10.1007/s10791-026-10708-7
Keywords: carbon intensity, carbon-aware scheduling, large language models, LightGBM, Prophet, machine learning forecasting, AI sustainability, energy consumption, green computing, workload optimization, emissions reduction, simulation study
News Source: Blake Davidson. (October 29, 2026). CarbonBalance Framework Shifts AI Workloads to Cleaner Power, Cutting Simulated Emissions 41.6%. Scienmag.
