Performance Analysis for Leveraging AI to Reduce Energy Consumption of a Data Center: A Safe Deep Reinforcement Learning and Workload-Aware Framework

Abstract

Abstract Electricity demand from data centers is rising quickly with the deployment of artificial intelligence (AI), and cooling remains the largest non-IT load, particularly in hot-arid climates. This paper presents a performance analysis of leveraging AI to reduce the energy consumption of a 10 MW AI/cloud data center located in Abu Dhabi. A component-level energy model is developed that couples server and accelerator power, uninterruptible power supply (UPS) losses, airflow and hot-air recirculation, server-fan penalties, air-cooled chiller efficiency and dry-cooler economization. On this digital twin, a five-layer framework integrates a long short-term memory (LSTM) load forecaster, a double deep Q-network (DDQN) that sets supply-air temperature, chilled-water temperature and airflow under a digital-twin safety shield, forecast-driven server consolidation and thermal-aware shifting of flexible workloads. Trained on one synthetic year and tested on an independent year against a static baseline, an ASHRAE-optimized rule-based strategy and model-based oracle bounds, the LSTM gives the lowest day-ahead error (MAPE 3.04%). The complete framework reduces annual facility energy by 15.3% (13.85 GWh), cooling energy by 40.8% and peak demand by 14.1%, and improves power usage effectiveness (PUE) from 1.421 to 1.291, while holding the 99th-percentile rack inlet temperature below 27°C. It captures 99.9% of the oracle-attainable saving and pays back in 1.7 years under stated assumptions. The results also show that PUE alone can misrepresent AI-driven savings.