DescriptionPower of large-scale systems such as an HPC or a datacenter is a significant issue. Cooling units consume 30% of the total power. General control policies for cooling units are local and static (manual overall optimization nearly once a week). However, free cooling and IT-load fluctuation may change hourly optimum control variables of the cooling units. In this work, we present a deep neural network (DNN) power simulator that can learn from actual operating logs and can quickly identify the optimum control variables. We demonstrated the power simulator of an actual large-scale system with 4.7-MW-power IT load. Our robust simulator predicted the total power with error of 4.8% without retraining during one year. We achieved optimization by the simulator within 80 seconds that was drastically faster than previous works. The dynamic control optimization each hour showed a 15% power reduction compared to that of conventional policy in the actual system.