A multi-method explainable AI for analyzing hybrid SARIMAX-GRU temperature forecasting in high-temperature data centers

Introduction Accurate temperature prediction is essential for improving energy efficiency and ensuring reliable operation in modern data centers. However, despite their high accuracy, deep learning forecasting models often lack interpretability, limiting their deployment in critical environments. This study extends a previously developed hybrid SARIMAX-GRU forecasting model by introducing an explainable artificial intelligence (XAI) framework to analyze and validate its decision-making process under different operating conditions. Methods The proposed framework employs four XAI techniques, including SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), Integrated Gradients, and Permutation Importance, to analyze the contribution of temporal residual features in temperature prediction. The analysis considers nine scenarios covering low, moderate, and high temperatures (24 °C, 27 °C, and 30 °C) with workload transitions between 50%, 75%, and 100%. Pearson and Spearman correlation analyses are used to evaluate the agreement between explanation methods in terms of importance scores and feature rankings, respectively. Results The results show that the hybrid SARIMAX-GRU model mainly depends on recent residual features, particularly Residual_t1 to Residual_t4, while older residuals have limited influence. The dominant features vary according to workload transitions and sensor locations, reflecting changes in thermal dynamics. SHAP, Integrated Gradients, and Permutation Importance demonstrate highly consistent explanations, with Pearson correlations ranging from 0.93 to 1.00, while LIME provides more variable but complementary interpretations. Spearman correlation analysis further confirms strong agreement in feature rankings among SHAP, Integrated Gradients, and Permutation Importance, whereas rank agreement involving LIME varies across scenarios. Discussion The XAI analysis confirms that the model captures meaningful short-term thermal dependencies rather than arbitrary patterns. The variations in feature importance across different operating conditions highlight the model's ability to adapt to changing thermal behaviors, while the agreement among multiple XAI techniques improves confidence in the generated explanations. Conclusion This study demonstrates that integrating XAI with hybrid temperature forecasting enhances model transparency and provides valuable insights into data center thermal prediction. The proposed framework supports more trustworthy intelligent cooling and energy optimization decisions by identifying the key factors influencing model predictions.