Development of data-driven models for estimating mud and filtrate alkalinity using machine learning applications

Abstract

Energy is essential for the sustainable development and economic growth of modern societies. As the population and industrialization grow, energy demand increases daily. Oil and natural gas are among the most important energy sources to meet this demand. Extracting these resources efficiently depends on the success of drilling operations. Drilling mud, used in these operations, performs key functions such as cooling the drill bit, stabilizing the wellbore, transporting rock cuttings to the surface, and controlling well pressure. Traditional methods for measuring mud alkalinity (PM) and filtrate alkalinity (MF and PF) are slow and prone to human error. These issues can reduce the efficiency of drilling operations. This thesis aims to speed up these processes and provide reliable results using data-driven machine learning models. The study focuses on how machine learning can reduce operational risks and costs in drilling operations. These models not only save time but also provide consistent and accurate results compared to traditional testing methods. As part of the study, experimental work was first contucted to identify the potential error margins encountered during alkalinity measurements. When transitioning to the model-based approach, 1600 drilling reports, covering different geological and operational conditions were used. The data includes drilling parameters such as mud properties, fluid loss, chemical composition, and pressure conditions. In the preprocessing stage, missing and inconsistent data were cleaned, and irrelevant data points were removed. Python libraries (e.g., Tabula, PyMuPDF) were used to structure the dataset accurately. Various machine learning algorithms were applied, including Linear Regression, Ridge Regression, Lasso Regression, ElasticNet, Random Forest, Gradient Boosting, XGBoost, CatBoost, LightGBM, Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP). The best hyperparameters for each model were found using GridSearchCV. The dataset was split into 80% for training and 20% for testing. Models were evaluated using metrics such as R² (Determination Coefficient), MSE (Mean Squared Error), MAE (Mean Absolute Error), and MAPE (Mean Absolute Percentage Error). Best models for each alkalinity measurements were determined by using TOPSIS (The Technique for Order of Preference by Similarity to Ideal Solution) which is the multi-criteria decision-making approach of tested models in this study. Firstly, the relationship between alkalinity measurements (PM, PF, and MF) was analyzed in preliminary part. It was observed that MF values were the most accurately predicted compared to other alkalinity measurements. The average accuracy was R² = 0.88 for training and R² = 0.705 for testing. In the first stage, the focus shifted to predicting MF using a more detailed machine-learning approach. XGB (Extreme Gradient Boosting) algorithm performed best, with an accuracy of R² = 0.996 for training and R² = 0.858 for testing. Using SHAP analysis, less important parameters were removed to improve model performance. Next, the relationships between MF, PF, and PM values were studied. RF (Random Forest) algorithm gave the best results for relationship between MF and PF with R² = 0.882 for training and R² = 0.864 for testing. For predicting PM, the Extreme Gradient Boosting (XGB) algorithm performed best for relationship between MF and PF, with an accuracy of R² = 0.809 for training and R² = 0.802 for testing. Using these models, MF values were predicted with using other drilling parameters, and these predictions were then used to estimate PF and PM values. This approach simplified the process, reduced model complexity, and improved speed and reliability. The results showed that XGB and RF models provided the highest accuracy for predicting MF, PF, and PM. These algorithms, which are tree-based techniques, were found to be the most suitable for this task. A user-friendly Streamlit application was also developed as part of this thesis. This app allows users to upload data, train models, and visualize prediction results easily. It provides a fast and efficient tool for engineers and technical teams in the field. This thesis demonstrates that machine learning models, particularly Extreme Gradient Boosting (XGB) and Random Forest (RF), are reliable and effective tools for predicting drilling mud and filtrate alkalinity (PM, PF, and MF). These data-driven models outperform traditional testing methods by being faster, more reliable, and more accurate. The study supports the digital transformation of the energy industry, showing the effectiveness of machine learning in optimizing drilling operations. The results are applicable to larger datasets and various operational conditions, making the models more flexible and generalizable. Additionally, a user-friendly Streamlit application was developed to enable users to easily upload data, train models, and visualize prediction results. This research also highlights the growing role of machine learning in the oil and gas industry, particularly as drilling parameters are increasingly measured using automation systems. By integrating alkalinity measurement into the automation process, this study provides significant advancements in efficiency, cost reduction, and accuracy in drilling operations.