Hybrid Machine Learning Model Based on GWO and PSO Optimization for Prediction of Oilwell Cement Compressive Strength under Acidic Corrosion

Abstract

Summary It is difficult to solve the problem that the cement sheath of oil and gas wells is corroded by acid gas, and the change in compressive strength (CS) of the cement sheath after corrosion is the key to affecting the sealing capacity of the cement sheath. In this study, we used four traditional machine learning (ML) algorithms—artificial neural network (ANN), support vector machine regression (SVR), extreme learning machine (ELM), and random forest (RF)—to establish a model for predicting the CS of corroded cement stone. We used Shapley additive exPlanations (SHAP) to explain the influence process of the input characteristics of the model on the output results, and explored the influence mechanism of various factors on the CS. The results show that SVR and RF are two of the four models with better prediction ability. Particle swarm optimization (PSO) and gray wolf optimization (GWO) algorithms are used to optimize SVR and RF models. After optimization, the prediction accuracy determination coefficient (R2) of the SVR and RF models was higher than 0.90, the R2 of the optimal model PSO-RF was 0.9275, and the root mean square error (RMSE) was 2.6516.