Abstract: NiMo，CoMo and NiMoW catalysts were evaluated with different diesel feedstocks in a high-throughput reactor under the conditions of a temperature of 300-360 ℃，a pressure of 4.4-7.4 MPa，a LHSV of 0.75-12 h^-1 and a volume ratio of hydrogen to oil of 200-800. The Neural network technology based on Keras was used to establish diesel ultra-deep hydrorefining models suitable for three different catalysts，and the prediction of sulfur，nitrogen，monocylic aromatics and polycyclic aromatics content in diesel products was realized. The results show that these models have good prediction performance and generalization ability. The average relative error of the prediction of sulfur and nitrogen content in the product is less than 10%，and the average relative error of the prediction of monocylic and polycyclic aromatics content is less than 3% and 6%，respectively. These models can be used to optimize the process conditions of the three catalysts simultaneously，and the range of process conditions for different catalysts can be determined under the premise that the sulfur and polycyclic aromatics content of diesel products can satisfy the requirements of China Ⅵ Standard.

Key words: hydrofining, Keras neural network, data driven model, process optimization, sulfur, nitrogen, monocylic aromatics, polycyclic aromatics