Electric load forecasting and computing-power scheduling for zero-carbon data centers

Abstract

The rapid growth of computing demand has turned data centers into large and highly variable electricity consumers, which makes accurate load forecasting a prerequisite for operating them close to a zero-carbon target. This work studies the joint problem of forecasting the electric load of a data center and scheduling its computing power so that flexible jobs are shifted toward periods of low grid carbon intensity and high renewable supply. A two-stage framework is proposed. In the first stage, a forecasting network combines variational mode decomposition with a temporal-attention Transformer encoder and a probabilistic head that outputs both the expected load and its quantiles. In the second stage, a carbon-aware scheduler consumes the forecast and solves a constrained optimization that trades operating cost against carbon emission, assisted by a learned dispatch policy. Experiments on a data-center campus load dataset, the Alibaba cluster trace and a regional carbon-intensity series show that the forecaster reaches a mean absolute percentage error of 3.87% at the 24 h horizon, and that the scheduler lowers daily carbon emission by 18.3% and operating cost by 9.6% while raising renewable utilization from 41% to 63%, with a quality-of-service violation rate below 1.2%. The results indicate that coupling probabilistic forecasting with carbon-aware scheduling is an effective path toward low-carbon data center operation.