Abstract: To improve the economic benefit of delayed coking unit (DCU), it’s necessary to establish a precise yield prediction model for various feedstocks and operation conditions. A lumping-BP neural network hybrid model in a cascade form was established for a DCU with capacity of 1.4 Mt/a to predict the liquid yield of the unit, based on the mathematical model of the reactor under the assumption of dynamic balance from 11 lumping dynamic model and the BP neural network input of the mechanism model calculation results and the historical data of key sites. In the case study, the coking diesel yield was predicted by the hybrid model, and compared with the results of mechanism model, empirical model. The results demonstrate that among these three methods, the prediction accuracy of the hybrid model is the best. The impact of the material properties and operating conditions fluctuation on the hybrid model results is small, the root mean square error, mean absolute error and the average relative error is 0.751 percentage point, 0.524 percentage point, 2.01%, respectively.