To address the challenge of multi-objective synchronous prediction in cold rolling processes, this study proposes a hybrid CNN-BiLSTM-Attention model integrating multi-stand spatiotemporal features. By constructing a structured input matrix (5 stands \(\times \) 18 features), we employ a dual-channel convolutional network to extract spatial correlations of multi-stand process parameters, coupled with a bidirectional long short-term memory (BiLSTM) network to model temporal dependencies. A feature-level attention mechanism dynamically quantifies the contribution weights of chemical compositions and process parameters. Experimental validation with 12,805 sets of industrial data demonstrates that compared to baseline models like Back Propagation Neural Network and Recurrent Neural Network, the proposed model achieves a normalized mean absolute error of 0.0210 for rolling force and deformation resistance prediction, representing an 11.00–31.91% reduction compared to baseline models, with a coefficient of determination (R2) reaching 0.987. Attention weight analysis reveals the nonlinear dominance of incoming material carbon content, rolling speed, and front tension. To validate error suppression through inter-stand coupling modeling, an isolated prediction group was established. Results show the model reduces average rolling force error by 24.53% and deformation resistance error by 11.46% via dynamic stand correlation modeling. For industrial implementation, a verification platform was developed on a 1420 mm cold tandem rolling mill. Through multi-process data fusion and periodic pre-training mechanisms, the model achieves synchronous prediction of rolling force and deformation resistance across five stands with prediction time <0.2s, meeting real-time control requirements. This study pioneers cross-process synchronous prediction of multi-stand parameters in tandem cold rolling, effectively resolving accuracy degradation caused by hot-cold rolling data isolation and error accumulation in conventional methods.
