A generalization method of intelligent ROP prediction under mechanism constraints
Accurate prediction of the rate of penetration(ROP)is supportive to rational resource allocation before well drilling,and has important practical significance in making proper drilling plan and increasing drilling efficiency at lower cost.Intelligent ROP prediction has become a hot spot in research.However,conventional intelligent models show poor transferability between wells.In this paper,comprehensive logging data are preprocessed by denoising and completion techniques.Then,constraints are constructed depending on drilling knowledge.By introducing the domain adversarial neural network(DANN),the transfer mechanism of ROP models between wells is established.Finally,based on sliding window,incremental updating and real-time logging data,a method for real-time update of the ROP model with downhole conditions is formed.The following results are obtained.First,both data layer constraint and network layer constrain can improve the accuracy and stability of the intelligent model,and the BP model under two mechanism constraints yields much more accurate results than ordinary BP model.Second,the DANN-based ROP model can effectively transfer the data knowledge from the offset well(source domain)to the test well(target domain).Third,the double-sliding-window data updating mechanism based on the incremental learning algorithm allows the model to adapt to the subsurface drilling conditions in a real-time manner,with further enhanced predication accuracy and generalization ability.Fourth,mechanism constraints,transfer learning,and real-time updates play a combined role in strengthening the generalization performance of the model.The average relative error of ROP prediction for new wells is reduced to 20.2％.It is concluded that the proposed ROP model yields better transferability and higher accuracy than traditional models,providing a new approach and direction for intelligent prediction of ROP.
万方数据
维普
