Effects of Tuning Hyperparameters in Random Forest Regression on Reservoir's Porosity Prediction. Case Study: Volve Oil Field, North Sea

Abstract

ABSTRACT Ensemble learning is a recent development in machine learning. Random forest regression (RFR) is one such widely utilized ensemble learning algorithm. However, the current literature lacks studies that primarily focus on the effects of hyperparameter tuning in RFR when predicting reservoir properties of hydrocarbon reservoirs. Thus, in this study we investigated the effects of three commonly used hyperparameters; namely, n_estimators, max_features and min_samples_leaf to predict porosity of Volve oil field in North Sea. Four parameters; depth, gamma ray logs, neutron porosity logs and resistivity logs were used as inputs, while calculated porosity was used as target outputs to develop the RFR models. The RFR models were developed through: (i) tuning each hyperparameter individually, (ii) tuning hyperparameters by coupling them into three groups and, (iii) tuning all three hyperparameters at once. Results showed that the highest performing model had an R2 value of 0.8517 with n_estimators of 100, max_features of 0.5 and min_samples_leaf of 1. Furthermore, it was observed that tuning max_features had a higher impact on improving the performance of the RFR model when predicting porosity of Volve oil field in North Sea. INTRODUCTION World is gradually moving towards an era of artificial intelligence (AI) where every major sector possibly be supplemented by man-like machines. Machine learning (ML) is a branch of AI which has the ability to predict or forecast outputs, decrease computational time and extract features from complex and high-dimensional data sets (Zhan and Kitchin, 2021). These properties of ML are favourable in industries where big data is utilized in several stages of the production line. Oil and gas industry is a perfect match for ML since it uses huge number of data to analyse and interpret both in upstream and downstream sectors. There are several ML algorithms which have been utilized in the literature for reservoir characterization including prediction of petrophysical properties. Artificial neural network (ANN) is one such ML algorithm. ANN's originated from mimicking biological neural systems (Wang et al., 2019). They are widely used in predicting petrophysical properties of hydrocarbon reservoirs (Al Khalifah et al., 2020, Urang et al., 2020, Okon et al., 2021). Supportive vector machine is another commonly used ML algorithm in the literature to estimate reservoir properties (Jamalian et al., 2018, Zhong & Carr, 2019, Wu et al., 2020). Deep learning is an ML algorithm which can be defined as an advanced version of ANN's. This algorithm is also being used in reservoir characterization (Chen et al., 2020, Zhang et al., 2021, Arigbe et al., 2019). 2019).