Abstract: This research aims to develop an advanced deep learning-based ensemble algorithm, utilizing environmental temperature and solar radiation as feature factors, to conduct hourly temperature field predictions for steel-concrete composite decks (SCCDs). The proposed model comprises feature parameter lag selection, two non-stationary time series decomposition methods (Empirical mode decomposition (EMD) and time-varying filtering empirical mode de-composition (TVFEMD)), and a stacking ensemble prediction model. To validate the proposed model, five machine learning (ML) models (random forest (RF), support vector regression (SVR), multilayer perceptron (MLP), gradient boosting regression (GBR) and extreme gradient boosting (XGBoost)) were tested as base learners and evaluations were conducted within independent, mixed, and ensemble frameworks. Finally, predictions are made based on engineering cases. The results indicate that consideration of lag variables and modal decomposition can significantly improve the prediction performance of learners, and the stacking framework, which combines multiple learners, achieves superior prediction results. The proposed method demonstrates a high degree of predictive robustness and can be applied to statistical analysis of the temperature field in SCCDs. Incorporating time lag features helps account for the delayed heat dissipation phenomenon in concrete, while decomposition techniques assist in feature extraction.