End-edge-cloud collaborative-driven waste-heat prediction of liquid cooling system for high-performance computing data centers

High-performance computing (HPC) systems are characterized by extremely high energy consumption, making efficient waste-heat management a critical challenge for sustainable data center operations. Traditional modeling approaches often struggle to capture the complexity of dynamic thermal loads, strong temporal dependencies, and real-time requirements of liquid cooling systems. To address these challenges, this study proposes an endedge-cloud collaborative (EECC) architecture for short-term waste-heat prediction, enabling intelligent thermal management in HPC data centers with improved accuracy. Within this architecture, a hybrid long short-term memory (LSTM)-Transformer neural network model is developed, integrating the short-term memory capability of LSTM with the long-range dependency modeling capability of Transformer. Experiments using real operational data from the Frontier supercomputer at Oak Ridge National Laboratory (ORNL) demonstrate that the proposed model outperforms several baseline models. Compared with the second-best Transformer model, the hybrid model reduces the root-mean-square error (RMSE) and mean absolute error (MAE) by approximately 21.3% and 28.4%, respectively, while achieving the highest coefficient of determination (R2) of 0.9767. In probabilistic forecasting, it achieves a prediction-interval coverage probability (PICP) of 0.9652 and a mean predictioninterval width (MPIW) of 0.4904, providing reliable and narrow intervals that support uncertainty-aware operational decisions. The model also demonstrates strong robustness under different random initializations, seasonal variations, and non-stationary operating conditions, confirming its practical applicability in real HPC environments. This study establishes a reliable artificial-intelligence-driven solution for high-precision shortterm waste-heat prediction, advancing green computing practices and enhancing energy efficiency and sustainability in HPC data centers.