Abstract: An improved prediction model using Levenberg-Marquardt algorithm was carried out to make up for some defects such as slow convergence speed and local minimum point in traditional algorithm of BP neural network. Then the new prediction model was applied to indicate the relationship between various technical parameters and K/S values in dyeing with reactive dyestuff. Meanwhile, the convergence time and predictive precision of both prediction models were compared. The results revealed that LM-BP network model achieved the required precision after about 7 steps training while more than 1500 steps were needed in traditional BP network model. The correlation coefficient between predictive and experimental K/S values was improved from 0.995 to 0.999 accordingly. It was considered that LM-BP method was superior in the prediction by antitheses. The predictive time was shorter and the predictive veracity was higher in LM-BP method.