Flatness is a critical quality parameter for cold-rolled strips, significantly affecting downstream processing and product performance. Traditionally, data barriers between hot- and cold-rolling production lines have limited prediction accuracy. To address this issue, we developed a cross-process industrial big data platform for seamlessly integrating data from hot- and cold-rolling lines. Based on this platform, a flatness prediction model was developed utilizing a Bayesian Optimized Light Gradient Boosting Machine (BO-LightGBM). The results demonstrate that incorporating hot-rolled data significantly improves prediction accuracy compared to models using only cold-rolled data. Specifically, when hot-rolled data is included, the BO-LightGBM model achieves a Mean Absolute Error (MAE) of 0.7545 and a Root Mean Square Error (RMSE) of 1.0077, reducing errors by 14.9% and 10.6%, respectively, compared to models without hot-rolled data. The BO-LightGBM model outperforms conventional methods, including Backpropagation (BP) and Fully Connected Neural Networks (FCNN). Compared to the BP model, MAE and RMSE decrease by 38.6% and 36.9%, respectively, and by 40.8% and 38.3% compared to the FCNN model. Additionally, the BO-LightGBM model achieves an R2 value of 0.9680, improving by 5.3% over the BP model and 5.7% over the FCNN model. Shapley Additive Explanations (SHAP) analysis provides insights into critical factors affecting flatness in both rolling stages. This study advances flatness prediction, providing a solid foundation for optimizing rolling processes.