Machine Learning–Based Cooling Energy Prediction and Intelligent Control in Data Centers

Data centers are core infrastructure for the digital economy, yet cooling systems still account for about $\mathbf{4 0 - 5 0 \%}$ of total energy use in many facilities, and have become a key bottleneck for green, low-carbon operation. Traditional set-point and rule-based strategies depend heavily on operator experience and often suffer from low prediction accuracy, delayed response to thermal disturbances and insufficient energy optimization. This paper proposes a machine learning-based cooling energy prediction and intelligent control method for an air-cooled internet data center. First, we construct a multi-source feature system that fuses room-level environmental parameters, IT load indicators and outdoor weather conditions. An improved random forest (RF) model with feature re-weighting and asymmetric loss is trained to predict short-term cooling power at the plant and CRAH levels. On this basis, we model server-room thermal dynamics using a long short-term memory (LSTM) network and embed it in a model predictive control (MPC) framework that optimizes chilled-water supply temperature, CRAH fan speed and temperature deadbands under comfort and safety constraints. A year-long dataset from a large internet data center is used for validation. The proposed RF model achieves 94.8% prediction accuracy, improving mean absolute percentage error by 15.2 percentage points over a backpropagation neural network baseline. In field deployment, the intelligent control strategy reduces cooling energy by 18.3% and improves PUE from 1.45 to 1.28 while maintaining room temperature within $\pm 0.5^{\circ} \mathrm{C}$ across low, medium and high IT-load scenarios. The method enhances both efficiency and intelligence of data center cooling systems and provides a practical technical route for green operation aligned with “dual-carbon” goals.