Permeability Predictions for Tight Sandstone Reservoir Using Explainable Machine Learning and Particle Swarm Optimization

Jing-Jing Liu,Jian-Chao Liu,Hongqing Song

doi:10.1155/2022/2263329

Abstract

High-precision permeability prediction is of great significance to tight sandstone reservoirs. However, while considerable progress has recently been made in the machine learning based prediction of reservoir permeability, the generalization of this approach is limited by weak interpretability. Hence, an interpretable XGBoost model is proposed herein based on particle swarm optimization to predict the permeability of tight sandstone reservoirs with higher accuracy and robust interpretability. The porosity and permeability of 202 core plugs and 6 logging curves (namely, the gamma-ray (GR) curve, the acoustic curve (AC), the spontaneous potential (SP) curve, the caliper (CAL) curve, the deep lateral resistivity (RILD) curve, and eight lateral resistivity (RFOC) curve) are extracted along with three derived variables (i.e., the shale content, the AC slope, and the GR slope) as data sets. Based on the data preprocessing, global and local interpretations are performed according to the Shapley additive explanations (SHAP) analysis, and the redundant features in the data set are screened to identify the porosity, AC, CAL, and GR slope as the four most important features. The particle swarm optimization algorithm is then used to optimize the hyperparameters of the XGBoost model. The prediction results of the PSO-XGBoost model indicate a superior performance compared with that of the benchmark XGBoost model. In addition, the reliable application of the interpretable PSO-XGBoost model in the prediction of tight sandstone reservoir permeability is examined by comparing the results with those of two traditional mathematical regression models, five machine learning models, and three deep learning models. Thus, the interpretable PSO-XGBoost model is shown to have more advantages in permeability prediction along with the lowest root mean square error, thereby confirming the effectiveness and practicability of this method.

Highlights

Permeability is an important parameter in tight sandstone reservoir evaluation and oil and gas field development and is the basis for establishing geological models, accurately estimating oil and gas reserves, and determining reasonable development plans [1,2,3,4]
To improve the accuracy and interpretability of the permeability prediction of tight sandstone reservoirs, an improved XGBoost model based on the particle swarm optimization (PSO) algorithm and attributable interpretation was proposed
The following conclusions were drawn: (1) The Shapley additive explanations (SHAP) can explain the importance of input features from a global perspective, mine the key features that affect the permeability prediction, and reduce the dimensionality of the samples according to the key features, but can clarify the quantitative contribution of each feature towards the permeability prediction based on the evaluation results of each individual sample

Summary

Introduction

Permeability is an important parameter in tight sandstone reservoir evaluation and oil and gas field development and is the basis for establishing geological models, accurately estimating oil and gas reserves, and determining reasonable development plans [1,2,3,4]. The main methods for the high-precision prediction of permeability are (i) mathematical regression methods such as porosity and permeability regression based on petrophysical data or the response of logging curves [7,8,9,10]; (ii) theoretical modeling based on high-pressure mercury intrusion data, e.g., Winland, Purcell, Swanson, and Katz-Thompson; [11,12,13] and (iii) prediction based on special logging series, e.g., nuclear magnetic resonance logging and dipole sonic logging [14,15,16] These methods have certain differences in terms of data requirements and parameter selection and are challenged by two main problems: (1) the accuracy of the mathematical regression method is poor. Only a limited number of core plugs can be collected, and these

Methods

Results

Conclusion