Tri-hybrid transformer framework for multi-target performance prediction of solid desiccant evaporative cooling systems

Le résumé fourni par la source

Solid desiccant evaporative cooling system (DECS) offers a sustainable alternative for building air conditioning; however, its coupled heat and mass transfer processes make accurate performance prediction challenging. Although data-driven models have increasingly been applied to this problem, existing approaches often suffer from limited prediction scope, reliance on single-paradigm architectures, insufficient evaluation of generalization capability, and limited interpretability. To address these challenges, this study proposes a Tri-Hybrid Transformer framework. The Tri-Hybrid Transformer integrates XGBoost leaf encodings, MLP-derived latent features, and raw inputs, which are concatenated and subsequently refined through a transformer-based attention module to capture higher-order feature interactions. The framework is validated using experimental data from an instrumented DECS operating under process air conditions of 30–40°C and 12–20g/kg, with regeneration temperatures ranging from 60 to 90°C. Five-fold cross-validation yields R2 values of 0.983 for Ts,out, 0.956 for Ws,out, 0.988 for CC, and 0.992 for COP. The proposed model consistently outperforms benchmark methods, including XGBoost, SVR, Random Forest, MLP, and Tabular Transformer. SHAP analysis further confirms physically consistent relationships between the input variables and the predicted outputs. These results demonstrate the potential of hybrid transformer-based models for accurate and interpretable prediction of DECS performance, enabling data-driven optimization of sustainable cooling technologies.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.