Aiming at the problems of complexity and low prediction accuracy of fluorine-free mold fluxes viscosity prediction in crystallizer, an extreme learning machine model based on improved whale optimization algorithm was proposed and used for fluorine-free mold fluxes viscosity prediction. Firstly, the fluorine-free mold fluxes composition data set was constructed, and the correlation analysis of the composition and viscosity value in the slag was carried out. Then, the population of whale optimization algorithm was initialized by using the improved Tent chaotic mapping and the inverse learning strategy, and the convergence factor of nonlinear convergence and adaptive t-distribution variation strategy were integrated to improve the algorithm's optimization ability of hyper-parameters in the extreme learning machine. Finally, the viscosity value prediction comparison experiments were conducted on the fluorine-free mold fluxes dataset to verify the effectiveness of the improved algorithm. The results indicate that compared to models such as BPNN (back propagation neural network) and ELM (extreme learning machine), the average absolute percentage error is reduced by 29.50% on average, and the optimization accuracy, prediction accuracy and stability are greatly improved.