AI-Driven Data Center Airflow Management and Cooling System Optimisations

Data centers (DCs) need to regulate airflow and adjust cooling system fan speed to maintain optimal thermal conditions and energy efficiency, but prior methods are challenging due to numerous configurations and data availability. This paper proposes an AI-centric framework for predicting supply airflow and optimizing cooling based on real data. The framework leverages RF, XGB, CNN, LSTM, and ensemble CNN-LSTM machine-learning models to predict supply-airflow patterns and adjust fan speed on the fly. The models were simulated using real data from an HPC DC ENEA, Italy. The study found that CNN-LSTM outperformed other methods with an average MAE of 0.02036 for predicting supply-airflow patterns every 15 min. The model was optimized by sensitivity analysis scenarios using fan speed, achieving an average MAE of 0.02026 when fan speed was dropped by 50%. Operators can dynamically manage airflow and adjust the fan speed by simulating the trained model without significant physical intervention.