For the characteristics of strong coupling and non-linearity of parameters in traditional rough rolling width prediction model, a new rough rolling width prediction model based on improved butterfly algorithm optimized extreme learning machine (IBOA-ELM) was proposed from the data-driven perspective. Firstly, the random weight and bias of the extreme learning machine (ELM) were optimized by butterfly optimization algorithm (BOA) to improve the prediction accuracy of ELM model. Then, for the problems that the butterfly optimization algorithm was easy to fall into local optimum and the convergence was poor, the butterfly  optimization algorithm was improved by introducing  the strategies of Fuch chaotic mapping, non-linearity inertia weights, refraction reverse learning and so on to further improve the accuracy of the width prediction model. Finally, the model was simulated and tested by the hot rolling production site data of a steel mill. The results show that the data-driven IBOA-ELM model has obvious advantages in prediction accuracy, and the hit rate of predicting the rough rolling width within ±8 mm is 93%, which is significantly better than the comparison models, and can be used for predicting the rough rolling width of hot rolled strips with strong applicability.