This project focuses on analyzing and predicting emergency incidents reported by the London Fire Brigade (LFB) in the Borough of Bexley (2019–2022). Using advanced data mining, clustering, and machine learning techniques, the project aims to uncover trends and improve emergency resource planning through predictive modeling.
Three key business problems were addressed:
- Where and when fire incidents are most likely to occur.
- Which fire station is most likely to respond to a new incident.
- Can incident types (Fire, False Alarm, Special Service) be predicted from early data?
-
Business Problem 1 – Descriptive:
Identify hotspot wards and peak time ranges to optimize emergency response readiness and resource planning. -
Business Problem 2 – Predictive:
Predict which fire station ground will respond to a new incident based on time, location, and other incident features. -
Business Problem 3 – Predictive + Descriptive:
Predict the type of incident and analyze the factors contributing to fire, false alarms, and special services for improved categorization and decision-making.
- 🚒 Supports faster and smarter emergency response by forecasting station dispatch responsibilities.
- 🕔 Allows optimized deployment during peak hours (e.g., 5–9 PM).
- 📊 Enables data-driven public safety campaigns and ward-level emergency planning.
- 💡 Helps reduce false alarms and improve equipment/resource utilization through intelligent prediction.
The project used a full data mining pipeline:
-
Descriptive Modeling
- Clustering (K-Means) with PCA for trend discovery
- Spatial and temporal hotspot analysis
-
Predictive Modeling
- Classifiers: Random Forest, XGBoost, Logistic Regression
- Balanced training using SMOTE
- Hyperparameter tuning with GridSearchCV
- Evaluation using Stratified K-Fold CV, accuracy, and confusion matrices
Top Results:
- Incident Type Prediction (Random Forest): 96.3% accuracy
- Fire Station Prediction (RF/XGBoost): 100% balanced accuracy
- Languages: Python
- Libraries: pandas, scikit-learn, seaborn, matplotlib, XGBoost
- Techniques: K-Means, PCA, SMOTE, Random Forest, GridSearchCV
- Platform: Jupyter Notebook
- Dataset: London Fire Brigade (2019–2022), Bexley Borough incidents
- Source: London Fire Brigade (LFB)
- Scope: 9,000+ records from Bexley
- Features:
- Temporal:
DateOfCall,HourOfCall - Spatial:
WardName,Latitude,Longitude - Operational:
PumpCount,PumpHours,Notional Cost - Labels:
IncidentGroup,IncidentStationGround
- Temporal:
https://drive.google.com/file/d/1c7ombS0zbwJSKBm3sVD8Jwi834xoYb__/view?usp=sharing
✔ Metadata alignment and data type fixing
✔ Missing value handling (geo-coordinates, times)
✔ Outlier treatment (IQR capping)
✔ Class imbalance correction (SMOTE)
✔ Label encoding and standard scaling
✔ Feature selection using PCA loadings
✔ Descriptive clustering (K=3 and K=6)
✔ Predictive modeling with cross-validation
- 🔥 Fire incidents peak between 5 PM–9 PM in Thamesmead East, Slade Green, Belvedere
- 🧠 PCA revealed
PumpCountandNotional Costas key features - 🚓 Predictive modeling accurately classifies incident types and station response zones
- 🗺️ K-Means clusters revealed high-resource incident groups
I undertook this project to apply end-to-end data science methodologies in a real-world public safety domain, combining descriptive analytics with predictive power. The challenge of optimizing emergency services with data is not only technically demanding but also socially impactful.
- 📌 Integrate real-time streaming data for live predictions
- 🧠 Add NLP analysis from incident reports
- 📡 Deploy models into an operational dashboard for LFB
- 🌍 Expand project to other boroughs beyond Bexley
- 📊 Use GIS for visualizing and tracking response coverage
To run or explore this project:
git clone https://github.com/your-username/fire-incident-analysis.git