Formation Grain Size Profile Prediction Model Considering the Longitudinal Continuity of Reservoir Using Artificial Intelligence Tools

Abstract

Grain size characteristics (d50, UC) of formation sands are crucial parameters in a sand control design. UC and d50 are commonly derived from sieve or laser particle size analysis (LPSA) techniques on a limited number of core samples in the process of drilling, which cannot represent the variations of grain sizes in the formation by the limited number of core samples. Moreover, staged and hierarchic design of sand control usually needs the whole longitudinal distribution profile of grain size. The grain size characteristics of the reservoir are formed in the process of a long history and have a good correlation with the formation environment of the sediments. Sand control design can only use test well data, because of lacking actual producing position cores. The vertical and horizontal anisotropy and heterogeneity of reservoirs bring difficulties and greater risks to the design of sand control schemes. Therefore, it is very important to find a simple and effective reservoir granularity prediction method. The existing prediction models by artificial intelligence method use single point logging data as eigenvalues to predict d50 and UC without considering the longitudinal continuity of data. This paper presents an efficient solution to predict grain size profile based on conventional logging curves by using four machine learning method (ANN, Random forest, XGBoost, SVM). In order to make full use of the geological continuity of the reservoir, the longitudinal continuous points according to the spatial correlation are adopted as the machine learning feature parameters from the perspective of geological analysis and the data-driven grain size profile prediction model are established by using the logging curve trend and background information, which further improves the prediction accuracy of the model and provides basic data for sand control. The ANN model of five point mapping has the best prediction effect in predicting d50 with a highest correlation coefficient 0.819 and a lowest error MAE 9.59. The XGBoost model of five point mapping has the best prediction effect in predicting UC with a highest correlation coefficient 0.402 and a lowest error RMSE 1.15. This method has been successfully used in offshore oil field in sand control optimization.