Abstract: The structure of a radial basis function (RBF) neural network based on the k-means clustering algorithm was optimised by employing the simulated annealing algorithm for improving the prediction accuracy. The NUMISHEET 02 fender was considered as the object of research and six equivalent drawbead forces were used as input variables. Based on Spearman correlation analysis and Latin hypercube sampling, the data which had smaller correlation coefficient values were chosen as training samples for the simulated annealing-radial basis function (SA-RBF) neural network. The numerical simulations of training samples were performed by employing the Dynaform software package. The evaluation functions of forming quality were established based on the wrinkling defects and crack defects. The nonlinear relationship between the equivalent drawbead force and the associated objective function was established by incorporating a SA-RBF neural network. NSGA-Ⅱ algorithm was employed to achieve the Pareto frontier and the best equivalent drawbead forces were determined by applying grey correlation analysis theory. Finally, the numerical simulation of fender forming was performed based on the optimised drawbead forces. The resultant forming limit diagram (FLD) indicates decreased wrinkles in the optimised forming part and greater uniformity in the plastic deformation, thereby leading to improvement in the quality of fender forming.