Well Log Analysis Projects
This repository contains two projects focused on well log analysis and oil production forecasting using various data analysis and machine learning techniques.
Project 1: Salt Creek Field Well Log Analysis
Overview
- Dataset: 'Salt-Creek.DAT' containing 403 rows and 10 columns of well log data
- Variables: GR, log10(LLD), log10(MSFL), DT, RHOB, NPHI, PEF, POR, Kg, ln(Kg)
- Objective: Analyze well logs using statistical analysis, classification, and regression
Methodology
-
Data Preprocessing
- Removed outliers using the Interquartile Range (IQR) method
- Reduced data from 403 to 286 rows
-
Exploratory Data Analysis
- Visualized correlations using Seaborn's pairplot
- Observed Gaussian or log-normal distributions for all variables
-
Dimensional Reduction
- Applied Principal Component Analysis (PCA)
- Selected 4 principal components explaining ~90% of variance
-
Classification
- Used K-means clustering with 6 clusters
- Visualized clusters along first two principal components
-
Regression
- Implemented a Decision Tree Regressor to predict permeability
- Used 70% training, 30% testing split
- Evaluated using MAE, MSE, RMSE, and R-squared metrics
Results
- Successfully created a model for classifying data and predicting permeability
- Suggested future improvements: hyperparameter tuning, ensemble methods, and refined clustering
Project 2: Volve Field Time Series Analysis
Overview
- Dataset: Volve Field time series data from 16 wells (2008-2016)
- Focus: Daily Oil Production Rate (m^3/day)
- Selected wells: P-F-12 and P-F-14 (highest non-zero production days)
Methodology
-
Data Exploration
- Visualized time series plots for P-F-12 and P-F-14
- Removed "shut-in" time (zero production) for analysis
-
Persistence Forecast
- Applied to P-F-12 well
- Limitations: 1-step horizon, struggles with large changes
-
LSTM (Long Short-Term Memory) Model
- Transformed time series into supervised learning problem
- Made data stationary through differencing
- Scaled data from -1 to 1
- Used 300 epochs and 4 neurons for training
-
Model Validation
- Applied LSTM model to P-F-14 well
- Attempted to increase forecast horizon
Results
- LSTM outperformed Persistence Forecast (RMSE: 12.7 vs 27)
- Validation on P-F-14 showed unrealistic sudden production increase
- Identified areas for improvement in increasing forecast horizon
Future Work
- Refine LSTM model to improve long-term forecasting
- Explore other machine learning techniques for well log analysis
- Investigate methods to increase forecast horizon reliability
Technologies Used
- Python
- Pandas
- Seaborn
- Scikit-learn
- Keras (for LSTM implementation)
This README provides an overview of two well log analysis projects. For detailed information, please refer to the individual project reports and code files within this repository.
