Abstract: Five coking coals and 44 groups of blended coals were studied, and the coking experiments with coal cup were completed using a 40 kg small coke oven. According to the yields of heavy component, dense medium component and loose medium component (Y_HC, Y_DMC and Y_LMC) obtained by all-component separation as well as the FT-IR parameters of I₃ and I₄ which reflect hydrogen bond association, aliphatic chain length and branched degree, the prediction model for coke quality was established with the BP neural network. Then, the characteristics of the model were discussed and the coking mechanism by the new model was analyzed. The results show that using new defined coal structure parameters to predict coke quality has some advantages. The predicted and measured values of coke formation rate (CR), micro-strength (MSI), reactivity of particulate coke (PRI) and post-reaction strength (PSR) are in good agreement, and the fitting correlation coefficient of y versus x reaches 0.986, 0.982, 0.956 and 0.926, respectively. The prediction results of CR, MSI and PRI by the model are good with the mean variation of nine samples being 0.53%, 1.58% and 1.28%, respectively. However, the prediction result of (PSR) is poor with the mean variation being 12.22%. The results can provide a good foundation for the establishment of a new method for coal blending.