| In the field of petroleum exploration and development,the inversion and interpretation of geophysical data(such as well-log data and seismic data)is the main approach for the determination of reservoirs properties and the foundation for the evaluation of development potential.Recently,with the rise of big-data and artificial intelligence techniques,using data-driven process instead of hypothesis-driven process and machine learning instead of theoretical derivation provide alternative ways for the interpretation of geophysical data.Meanwhile,how to use machine learning techniques to assist the interpretation of geophysical data effectively has also become a hotspot of current researches.Among various machine learning algorithms,the random forest algorithm has many advantages such as strong learning ability,high algorithm robustness and high dimensional adaptive,which has been successfully applied in many engineering fields and achieved excellent results.However,at least for now,the application of random forest in geophysical interpretation is just getting started,which still needs further study and systematic summary.In this thesis,we take three geophysical interpretation tasks(including logging lithology discrimination,logging formation property estimation,and seismic channel-sand body identification)as examples.By considering the characteristics of each task,we propose serval improved random forest algorithms such as the prototype similarity based random forest,linear random forest,and pruning random forest to improve the applications of random forest in geophysical interpretation,which provide specific optimization solutions for further practices.Specifically:(1)To deal with the complex and variable sample distribution patterns in logging lithology discrimination task,by combining the standard random forest classification algorithm with the mean-shift algorithm,we introduce the prototype similarity based random forest approach for better applications.Based on the logging response prototypes of different lithologies extracted by the Mean-Shift algorithm,the introduced approach will construct a prototype similarity space and use the random forest classification algorithm to learn the specific lithology logging response patterns effectively.The applications on 9 blocks in different oilfields show that compared with 5 other machine learning classification algorithms,the proposed prototype similarity random forest approach improves the accuracy of lithology discrimination significantly,which is able to perform more effective logging lithology discrimination.(2)By combining the standard random forest regression algorithm with the M5 linear decision tree algorithm,we introduce the linear random forest algorithm for more effective logging regression modeling.The introduced algorithm replaces the local constant model in the standard random forest with a local linear model.The randomized M5 linear decision tree algorithm is used to fit a series of piecewise linear models,and then bagging ensemble is used to enhance the stability,robustness and smoothness of the final model.The applications of 23 logging regression modeling tasks on 6 blocks show that compared with 8 other machine learning regression algorithms,the introduced linear random forest regression algorithm can reduce the prediction error significantly.Further experiments indicate that the superiority of the proposed algorithm is derived from its nature of smaller sample requirement,stronger algorithm robustness,and more reasonable model formula,which makes it very suitable for logging regression modeling and prediction problems.(3)The seismic channel sand-body identification task can be transformed into a machine learning classification problem.In order to reduce the influence of data error,high input feature dimension,and irrelevant or redundant input information,by combining the random forest classification algorithm with the post-prune operation of decision trees we introduced the pruning random forest algorithm.For the construction of each base decision tree in the proposed algorithm,the whole samples will be partitioned into the training set and the pruning set randomly.A randomized CART model with complete growth will be learned based on the training set at first,and then the postpruning based on the pruning set will be performed to improve the accuracy and robustness of trees.The application and experiments in the West Bohai Sea show that the introduced algorithm improves the accuracy,robustness and feature selection ability of models,while also simplifies the parameter optimization process in practice.Compared with 5 other machine learning classification algorithms,the introduced pruning random forest algorithm has obvious advantages for the machine learning assisted seismic channel sand-body identification. |
