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

To address energy-saving optimization in edge data center clusters under high-density cabinet deployment, dynamic workload fluctuations, and increasingly strict energy-efficiency constraints, this paper proposes a collaborative energy-saving strategy driven by deep learning and real-time load forecasting. The proposed strategy is developed for the information equipment room of an electric power scientific computing and analysis center. A multivariate time-series modeling framework is constructed for 104 mixed-density cabinets, integrating cabinet power, temperature and humidity, UPS status, precision air-conditioning status, power distribution parameters, alarm information, and time-event features into a unified input space. Methodologically, a Hierarchical Temporal Transformer (HTT) is designed to model intra-group load correlations among network cabinets, conventional server cabinets, and high-density server cabinets through cabinet-group local attention, while cluster-level global temporal fusion captures load migration relationships among cabinet groups. Quantile forecasting is further introduced to estimate uncertainty intervals of future loads. Based on the forecasting results, a collaborative optimization mechanism consisting of precision air-conditioning pre-control, UPS efficiency optimization, and migratable workload scheduling is developed, and model predictive control is used to realize rolling closed-loop adjustment. Experimental results show that, under a 60 min forecasting horizon, the proposed HTT achieves an MAE of 16.1 kW, an RMSE of 23.9 kW, and a MAPE of 2.02%, outperforming ARIMA, SVR, LSTM, TCN, Informer, and PatchTST. When combined with collaborative scheduling, HTT reduces the Average PUE from 1.340 under fixed-threshold rule control to 1.246, with an energy-saving rate of approximately 7.0% and an estimated annual electricity saving of 1,035,064 kWh, while reducing the temperature SLA violation rate to 0.6%. Ablation experiments, load-scenario experiments, and forecasting-horizon sensitivity analysis demonstrate that the proposed method achieves a favorable balance among prediction accuracy, energy efficiency, and operational reliability, providing a deployable technical route for real-time energy-efficiency optimization in edge data center clusters.