The strip flatness is a crucial index determining product quality in the strip tandem cold rolling (SCR). Data-driven methods (DDM) for flatness prediction are able to capture the uncertainties and nonlinearities in SCR compared with traditional analytical method and numerical simulation methods. However, they only depend on the industrial data during SCR, which restricts the adaptability and precision, especially for the strip head-end part, when the incoming state of hot rolled strip fluctuates violently. To break the restrictions, a novel paradigm of flatness prediction and optimization for SCR is proposed in this study based on a cloud-edge collaboration platform. The platform is designed to reasonably allocate computing tasks to achieve real-time optimization of flatness, and to provide an intact dataset by collecting the industrial data with heterogeneous interfaces and data types from SCR and strip tandem hot rolling (SHR) production line; an artificial neural network algorithm optimized by an artificial bee colony algorithm (ABC-ANN) is proposed to predict the flatness; and feedforward feedback coordinated regulation (FFCR) method is established to optimize the flatness control. The results of a cause study on a 1420 mm SCR production line indicate that the new proposed mode is feasible and outperforms the traditional modes only involving the data in SCR, and it can accurately predict and optimize the flatness including the strip head-end part.