Accurate prediction of rate of penetration （ROP） in deep geological drilling process can help to improve drilling efficiency and reduce drilling costs， which can provide key technical support for safety and efficient deep geological drilling construction. In this paper， a dynamic prediction method of ROP in deep geological drilling process based on regional multi-well data optimization and model pre-training is proposed. First， the deep geological drilling data warehouse is designed by selecting lithology identification software， drilling process intelligent monitoring cloud platform， and geological cloud system as data sources. Secondly， the regional multi-well data optimization technique is used to select the matching data with the target well in the data warehouse， and the ROP model is pre-trained on this basis. Finally， the ROP prediction model is dynamically updated through combining the actual drilling data of deep geological drilling process， and introducing techniques such as the wavelet filtering， extreme learning machine， and incremental learning strategy. The experimental comparison results verify that the proposed method has strong ROP prediction performance and generalization capability.