🛢️ Oil & Gas Pipeline Incident Analysis (CER Dataset)

---

📌 Project Title

Geospatial and Meteorological Analysis of Substance Release Incidents in Canadian Pipelines

---

📖 Overview

This project analyzes pipeline incident data provided by the Canada Energy Regulator (CER). The dataset contains 723 recorded incidents (2008–2020) involving substance releases in oil and gas pipelines and related facilities.

The main objective is to understand which geographical and environmental factors are associated with the probability of a substance release incident.

---

🎯 Research Question

What geographical and meteorological factors are associated with pipeline incidents involving substance release?

The response variable is:

• SubstanceRelease (Yes/No)

The predictors include:

• Geographical variables (Province, Latitude, Longitude, population-related features)
• Operational variables (Company, Status, Release Type)
• Incident characteristics (Significant, Substance type)
• Temporal variable (Year)

---

📂 Dataset Description

The dataset includes 723 observations from CER records:

• 📍 Location data (Latitude, Longitude, Province, nearest populated center)
• 🏢 Company responsible for the pipeline
• ⚠️ Incident attributes (Release type, substance type, significance)
• 📅 Year of incident
• 📊 Operational status (Closed, Submitted, Initially Submitted)
• 🧾 Free-text fields describing cause and outcome of incidents

---

🧪 Methodology

1. Exploratory Data Analysis (EDA)

Initial data exploration using tidyverse and ggplot2:

• Yearly trend of substance releases
• Province-wise incident distribution
• Company-wise incident frequency
• Impact of operational status and release type
• Substance-level breakdown

Key insight:

• Certain provinces show higher likelihood of substance release events.
• Company and location are important predictive variables.

---

2. Feature Engineering

We created additional variables:

• Count of reasons (n.why.It.Happened)
• Count of incident descriptions (n.What.Happened)
• Geographic clustering using K-means on Latitude & Longitude

Optimal clustering was selected using within-cluster variation analysis.

---

3. Modeling Approach

We applied logistic regression models (GLM with logit link) to predict:

Probability of SubstanceRelease = Yes

Train/Test Split

• 70% training set
• 30% testing set

---

4. Models Evaluated

Multiple Modles:

• Province-only model
• Year-only model
• Location (Latitude, Longitude)
• Significant incident flag
• Release type
• Status
• K-means cluster (geospatial grouping)
• Combined feature models

---

5. Final Model

The best performing model included:

• Latitude
• Longitude
• Year
• Province
• K-means cluster (geospatial grouping)

---

📊 Results

• Model performance ranged between 50% – 70% accuracy
• Final selected model achieved approximately:

✅ ~70% prediction accuracy

---

🧠 Key Findings

• 📍 Geography is the strongest predictor of substance release incidents
• 🗺️ Provinces and spatial clustering significantly influence risk
• 🏢 Company-level effects exist but are harder to generalize
• 📅 Temporal trends show variation in incident frequency over years
• ⚠️ Some variables (e.g., Release Type, Status) have limited predictive power

---

⚠️ Limitations

• High-cardinality categorical variables (e.g., Company, Substance) were excluded from final models due to:
  • Class sparsity
  • Training/testing inconsistency
• Some text variables required simplification (counts instead of NLP modeling)
• Accuracy ceiling limited due to data imbalance and complexity

---

📁 Repository Contents

• project3.R → Main R script
• project3.Rmd → R Markdown analysis report
• project3.pdf → Final compiled report
• projectData.csv → CER incident dataset

---

📌 Conclusion

This study shows that geographical and spatial factors play a key role in predicting pipeline substance release incidents. While operational and categorical variables provide additional context, location-based modeling offers the strongest predictive signal.

---

📊 Tools & Libraries

• R (glm, kmeans)
• tidyverse
• ggplot2
• ggthemes

---