Abstract:

Aiming at the problems of underwater distributary channel sand body, such as thin single layer thickness, serious superimposition and crossing, strong lateral heterogeneity, and poor consistency of well seismic relationship, this study presents a prediction method of CV-XGBoost reservoir thickness based on cross validation. Firstly, correlation analysis and multicollinearity evaluation are used to remove redundant attributes, then the model training and parameter set optimization are carried out, and finally thickness prediction is carried out with verification set. The results show that: 1) For reservoir prediction with few samples, it is necessary to do cross validation to improve the accuracy of reservoir prediction; 2) XGBoost uses a regular term with the second order partial derivative to control the convergence progress of the model, and its prediction accuracy is better than that of conventional LASSO (least absolute shrinkage and selection operator) regression, GBDT (gradient boosting decision tree) and SVM (support vector machine) methods; 3) The reservoir prediction results with low verification set ratio can be used to understand the macro distribution of sand body, and the higher verification set ratio is helpful to improve the accuracy of sand body description; 4) The average amplitude, average energy, arc length and dominant frequency in this study area are the attributes that contribute greatly to thickness prediction.

Key words: cross validation, XGBoost, attribute optimization, sand body thickness prediction