Study On Prediction And Application Of Channel Sand Body In A Construction Area In Songliao Basin

Facing increasingly complex types of oil and gas reservoirs,reservoir prediction is especially targeted at hidden types such as channel sand bodies.If you want to obtain information about the physical properties and lithology of the reservoir,you must make sure that the attributes and the parameters of the reservoir are different.Linear correlation.At present,machine learning has been widely used and studied.Different reservoirs in different industrial areas are also suitable for different algorithms,and each algorithm has advantages and disadvantages.This paper focuses on the random forest regression algorithm,analyzes and compares support vector machines and BP neural network,and discusses the applicability in the prediction of channel sand body thickness.Based on the research on the principle of seismic attributes and its geological significance,firstly perform spectral decomposition on the seismic data volume to obtain the seismic data volume of each frequency division,and then extract several seismic attributes of the frequency data volume,mainly of the amplitude class;secondly In this paper,we use the function of random forest to rank the importance of features to optimize seismic attributes,which are respectively for the four lithologically sensitive log curves of natural gamma(GR),acoustic time difference(AC),and shallow and deep resistance(RLLD & RLLS).Four sets of corresponding sensitive seismic attribute bodies are selected;again,the values of various logging curves are used as learning targets,and different seismic attributes are used as inputs to build a machine learning sample set,and a log data volume regression model is established.In this part of the study,root mean square error(RMSE)is used as an evaluation index,and error analysis is performed on models constructed by three algorithms: support vector machine,random forest,and BP neural network.The RMSE evaluation results of the three models are respectively for the 4.86,4.38,and 5.22,a random forest regression algorithm with the smallest error value was selected for subsequent research,and then four well-logging 3D seismic volumes were mapped.Finally,the same method is used to evaluate the importance of the parameters.According to the importance of the parameters of the four different types of logging data as different weights,the seismic attribute fusion is performed on the four-logging attribute 3D seismic volumes to obtain the predicted 3D data volume of the sand body.Based on the horizon calibration,the sandstone threshold of the predicted data volume is determined based on the reference drilling results,so as to obtain the sand body distribution of the reservoir section in the entire work area.According to the analysis of the predicted value error,the absolute error between the predicted value and actual value of the sand body thickness of several main target layers is less than 1 m,and the coincidence rate is more than 85%.Layer prediction results,especially for areas with low well control.The full-text processing method is simpler and more convenient,faster,less human intervention,and more objective.