An energy-saving strategy for edge data center clusters driven by deep learning and real-time load forecasting

This paper introduces a collaborative energy-saving strategy for edge data center clusters, which utilizes deep learning and real-time load forecasting. The strategy is designed for a scientific computing and analysis center with 104 mixed-density cabinets. The researchers built a multivariate time-series model incorporating various cabinet parameters into a single input space.

The Hierarchical Temporal Transformer (HTT) was developed to model load correlations among different types of cabinets and capture load migration relationships between cabinet groups. Quantile forecasting was also introduced to estimate uncertainty intervals for future loads. The forecasting results led to a collaborative optimization mechanism that included precision air-conditioning pre-control, UPS efficiency optimization, and migratable workload scheduling, with model predictive control used for rolling closed-loop adjustments.

Experimental results demonstrated that the HTT achieved lower mean absolute error (MAE) and mean absolute percentage error (MAPE) values compared to other models, including ARIMA, SVR, LSTM, TCN, Informer, and PatchTST. When combined with the collaborative scheduling, HTT reduced the Average Power Usage Effectiveness (PUE) from 1.340 to 1.246, achieving a 7.0% energy-saving rate and an estimated annual electricity saving of 1,035,064 kWh.

The study also showed a reduction in temperature SLA violation rate to 0.6%. The researchers emphasize that their method offers a balance between prediction accuracy, energy efficiency, and operational reliability for real-time energy-efficiency optimization in edge data center clusters.