When acquiring underground reservoirs,oil and gas information,borehole logs may get lost or distorted in some hole sections due to instrument failures or wellbore collapse.Compared with traditional log reconstruction methods based on empirical models or multiple regression analysis,machine learning can better reconstruct borehole logs and accurately characterize the complex nonlinear relationships between borehole logs.Nevertheless,it still faces challenges such as poor generalizability,high trial-and-error costs,and low levels of automation.This paper applies the automatic control technology to multi-model selection and hyperparameter optimization by taking data processing and characterization engineering,model saving and prediction,model interpretation and fairness checking as the technical process.And by means of data pre-processing and visualized post-processing techniques,a method for the establishment of log reconstruction process based on automatic machine learning is developed and then validated in practical applications.The following results are obtained.First,among automatic machine learning,the tree-based Bayesian optimization search strikes a balance between prediction performance and computation efficiency.Second,multi-model selection is better than single model,and interpretability analysis and fairness checking can guide model selection and ensure model generalizability.Third,the incorporation of the non-numerical information of geological stratification and cuttings logging is conducive to improve prediction accuracy further.Fourth,the treatment of missing values and the selection of normalized methods have certain influence on model performance.In conclusion,compared with traditional machine learning methods,the automatic machine learning can better exploit the potential of multi-model selection and parameter optimization,and can automatically search the model applicable to the research objectives.What's more,the automatic machine learning improves accuracy and efficiency while reducing manual intervention and trial-and-error costs,which makes machine learning methods more applicable to various prediction tasks in the field of petroleum geological exploration.
维普
万方数据