<p>Strip flatness defects consistently pose significant quality challenges in the cold tandem rolling process. It is of great significance to predict flatness of products according to the process conditions for flatness control. In this paper, a novel deep ensemble cascade flatness prediction model based on Deep Ensemble Forest (DEF) was proposed, which can accurately predict strip flatness under various conditions. To parallelly learn the training data information from different perspectives and obtain excellent generalization ability for multi-tasks, the proposed DEF model includes deep parallel kernel which combines two ensemble ideas of Boosting and Bagging. Furthermore, according to the distribution characteristics of industrial data collected in the rolling process, the proposed model provides specific processing methods to prevent noise caused by information differences in simple splicing of industrial stack data. Additionally, to obtain superior model prediction performance in DEF construction, the adaptive weighting decision coefficient (RSRE) is defined to make dynamic decisions at the output layer. The performances of the proposed model and some state-of-the-art models are compared by coefficient of determination (R²). The results indicate that the proposed model achieves superior prediction accuracy and demonstrates strong adaptability compared to the tested baselines. The R² values for the five flatness coefficients are 0.973, 0.982, 0.969, 0.948, and 0.987, respectively.</p>