Abstract: The rolling force prediction model is the core of the cold rolling set control system. For a long time, the traditional rolling force theoretical model has exhibited low prediction accuracy and a strong dependence on empirical parameters due to the complex influencing factors in the cold rolling process, such as multivariable interactions, strong coupling, nonlinearity, and time-dependence. These limitations have hindered its ability to meet the production requirements of high-precision cold-rolled ultra-thin-gauge strips. The setting of the rolling force primarily depends on the calculation accuracy of deformation resistance and the friction coefficient. By analyzing the classical Bland-Ford-Hill cold rolling force theoretical model, inverse calculation formulas for deformation resistance and the friction coefficient were established, thereby obtaining their true values. Subsequently, a least squares support vector machine (LSSVM) model optimized by the differential evolution (DE) algorithm (DE-LSSVM) was constructed. By inputting the true values of deformation resistance and the friction coefficient into the DE-LSSVM for training, corrections were made to these parameters, thus optimizing the rolling force theoretical prediction model. Experimental results demonstrate that, compared with the traditional rolling force theoretical model, the deviation of the rolling force prediction model based on optimization of deformation resistance and friction coefficient is controlled within 5%.