Machine learning method for lacustrine shale oil reservoirs: Improving movable fluid porosity prediction

Movable fluid porosity (Φ mf ) is a key parameter to evaluate the seepage capacity and movable oil volume of shale reservoirs. Accurate measurement of the Φ mf relies on expensive and time-consuming experimental methods, and little work has been done on intelligent prediction. In this study, the Upper Cretaceous Qingshankou Formation shale in the Changling Depression of the Songliao Basin is selected as the research object. The 635 data points are divided into three datasets: training set, validation set, and testing set, allocated in a ratio of 0.7:0.15:0.15. Using five logging parameters including Gamma ray log (GR), Deep lateral resistivity log (RLLD), Acoustic log (AC), Density log (DEN), and Neutron log (CNL), four machine learning models are selected to predict Φ mf , including Decision tree (DT), Random forest (RF), Gradient boosting decision tree (GBDT), and Artificial neural network (ANN). Multiple evaluation indicators are adopted to compare the accuracy and applicability of different algorithms. The results show the performance of different algorithms in predicting Φ mf , from highest to lowest, as follows: GBDT > RF > DT > ANN. The sensitivity analysis based on the Shapley Additive exPlanations (SHAP) indicates that AC has the greatest positive impact on predicting the Φ mf . The applicability analysis shows that compared with the traditional Multiple regression analysis (MRA), the use of machine learning algorithms can effectively improve the prediction accuracy, with a maximum increase of 24 %. This study holds that the GBDT can predict the Φ mf of shale reservoirs efficiently and accurately, providing valuable insights for the global evaluation and development of lacustrine shale oil resources. • Machine learning method is used to predict the movable fluid porosity at first. • Multiple machine learning methods are compared. • The influence of logging parameters on model prediction results has been clarified. • The Gradient boosting decision tree model has been validated to have high accuracy and applicability.