BiGRU neural model predicts data center PUE

The paper authors present a Bidirectional Gated Recurrent Unit (BiGRU) neural network model to predict Power Usage Effectiveness (PUE) for data centers and compare its performance to a standard GRU model.

• BiGRU-based PUE prediction is trained on a simulated Singapore data center dataset from EnergyPlus consisting of 52,560 samples with 117 features; relevant feature subsets are selected via Recursive Feature Elimination with Cross-Validation (RFECV) to tune hyperparameters and train the final model, whose performance is evaluated using MSE, MAE, and R-squared metrics against a GRU baseline.
• The work is accepted at the 2025 International Joint Conference on Neural Networks (IJCNN) in Rome, Italy, focuses on data center energy efficiency and carbon footprint reduction, and uses PUE as the operational efficiency metric to support more sustainable, cost-effective data center energy management amid growth in edge computing and AI workloads.