The Beddoes-Leishman (B-L) semi-empirical model is a classical method for predicting the dynamic stall of wind turbine airfoils. However, the model performs poorly for rough-surface airfoils due to the premature entry of turbulence and stall condition under the roughness effect. In this study, an improved dynamic stall model is developed for rough-surface airfoils. Two corrections are involved in the modifications. First, the calculation of flow separation points is corrected considering the difference between the normal and tangential separation points, and then an attenuation coefficient is introduced for the effective angle of attack to account for the rough-surface effect. Subsequently, the proposed model is extensively verified by experimental data of S-series airfoils with rough surfaces. Results show that the proposed model can well predict the dynamic stall characteristics of rough-surface airfoils with a higher accuracy than that of the original B-L model under various mean angle of attacks, oscillating amplitudes and reduced frequencies. At last, the model is utilized to predict the operating loads of a wind turbine using the Blade-Element Momentum theory. Significant difference between the dynamic loads predicted by smooth and rough models is observed. This indicates that the roughness effect on the dynamic stall model as well as the prediction of the operating loads on wind turbines cannot be ignored.