CarbonCast SARIMAX
Time-Series Carbon Intensity Forecasting
The Objective
I worked under Professor Abel Souza to reduce the carbon footprint of cloud computing workloads. The strategy was to predict periods of low carbon intensity on the power grid and schedule heavy compute jobs during those windows. I built a Seasonal Auto-Regressive Integrated Moving Average with eXogenous factors (SARIMAX) statistical model to benchmark against heavier machine learning approaches. The goal was to prove that if a lightweight statistical model could predict intensity windows accurately, we could avoid wasting massive amounts of compute power running ML models just to schedule tasks. While the statistical model required significant fine-tuning to hit the required accuracy thresholds, it validated the theoretical approach.
The solar prediction
The wind prediction
Predictions & Validation
I conducted testing regarding the accuracy of the SARIMAX model against the prexsisting CarbonCast ML models for predicting low-intensity windows. The SARIMAX model was more effeceint given certain regions but not for others. Additionally given certain power sources (like solar), the SARIMAX model was a more obvoius choice due to the repetitive and preedictable nature of the power source - [USE final report for souzxa here to add more content]
The training process analysis