Abstract:

Lithology identification is universally regarded as a prerequisite for some fundamental geological work such as stratigraphic correlation and analysis of sedimentary system distribution. The tight sandstone reservoirs are generally composed of multiple thin sandstone-mud alternating beds, featuring a variety of lithological types, most of which are similar in log responses, posing great difficulty for lithology identification. Given the powerful analytical performance of machine learning techniques shown on pattern recognition, XGBoost, a stable and highly efficient model is adopted to predict lithology of the tight sandstone reservoirs. However, this model needs many empirical parameters to complete its prediction, and its computing efficiency will be reduced when it deals with more independent variables. The CRBM and PSO models are, therefore, introduced to improve the prediction performance. CRBM is functional to extract features from original data so that fewer but more significant independent variables can be created by the CRBM in modeling. While PSO is effective to determine the optimal values via iterative computation for those empirical parameters used by XGBoost. Based on the data derived from some cored wells of Chang 4+5 Members in western Jiyuan oilfield, two experiments are designed to validate the prediction performance of the hybrid model proposed, CRBM-PSO-XGBoost. Furthermore, two optimized models formed by PNN and SVM are selected to make a comparison, as a way to consolidate the validation effect. The experiment results manifest that the prediction accuracy comes to peak by the proposed model, up to over 90%. In all, the hybrid model proposed in the study, capable of identifying lithology of tight sandstone reservoirs, is of wider application to lithology prediction compared with previous models.

Key words: XGBoost model, CRBM model, PSO model, lithology identification, tight sandstone reservoir, Jiyuan oilfield, Ordos Basin