final-year-project

Pore Pressure Prediction in Well 15/9-F-12 Using Machine Learning

Overview

This repository contains the dataset documentation, preprocessing workflows, and machine learning models for predicting pore pressure (PP) in well 15/9-F-12 (Volve Field, Norwegian North Sea).

The project evaluates four machine learning algorithms across an independently reconstructed well-log dataset comprising 1,811 observations spanning the depth interval 3127.71 m to 3403.55 m.

Modeling Framework

The study benchmarks four regression models representing different algorithmic families:

1. Multiple Linear Regression (MLR): Linear baseline
2. Random Forest Regression (RF): Bagged decision tree ensemble
3. CatBoost Regression: Gradient boosted decision trees
4. Multilayer Perceptron (MLP): Artificial neural network

Performance evaluation is conducted using four standard metrics:

• Mean Absolute Error (MAE)
• Mean Squared Error (MSE)
• Root Mean Squared Error (RMSE)
• Coefficient of Determination (R-squared)

Data Partitioning Strategy

To account for vertical spatial autocorrelation in well logs, the data is partitioned into contiguous depth blocks rather than random shuffling:

• Training Set (70%): Shallower interval (~1,268 observations)
• Validation Set (15%): Middle interval (~272 observations)
• Testing Set (15%): Deepest interval (~271 observations)

All data preprocessing parameters (such as feature scaling) are fitted strictly on the training set and applied forward to validation and test sets to eliminate data leakage.

Repository Contents

• README.md: Project summary and framework outline
• DATA_SUMMARY.md: Predictor dataset verification, feature mappings, and split partitions

Further implementation scripts and modeling code will be added as workflows progress.