Case Study: Public Transport Demand Forecasting¶

Case Study

Executive Summary¶

A metropolitan transport authority implemented an AI-powered demand forecasting system to predict passenger volumes across the network. The system enabled dynamic service adjustments, improving service reliability by 12%, reducing overcrowding incidents by 35%, and optimizing staff rostering.

The Challenge¶

Situation¶

• 1.2 million daily passenger trips
• 450 bus routes, 15 train lines, 3 light rail corridors
• Service planning based on historical averages
• Significant demand variability (events, weather, holidays)
• Limited ability to respond to demand changes

Problems¶

• Fixed schedules didn't match actual demand patterns
• Overcrowding during peaks affecting customer experience
• Empty services during off-peak wasting resources
• Event-driven demand spikes poorly managed
• Difficulty planning for new routes or changes

Business Impact¶

• Customer satisfaction declining (4.1 → 3.6/5 over 3 years)
• Complaints about crowding increased 40%
• $8M annual cost from inefficient resource allocation
• Staff rostering challenges
• Political pressure on service quality

The Solution¶

AI Approach¶

Model Type: Time series forecasting with external factors Architecture: Ensemble (Prophet + LSTM + Gradient Boosting) Integration: Operations planning system

System Design¶

flowchart LR
    subgraph IN["<strong>Data Inputs</strong>"]
        I1[Historical Ridership]
        I2[Real-time Counts]
        I3[External Factors]
    end

    subgraph FE["<strong>Feature Engineering</strong>"]
        F1[Time Features]
        F2[Calendar Events]
        F3[Weather & Special Factors]
    end

    subgraph MOD["<strong>Forecasting Models</strong>"]
        M1[Prophet - Trend]
        M2[LSTM - Sequential]
        M3[XGBoost - Boosting]
    end

    subgraph OUT["<strong>Operations Integration</strong>"]
        O1[Service Planning]
        O2[Rostering System]
        O3[Real-time Display]
    end

    IN --> FE --> MOD --> ENS[Ensemble Forecast]
    ENS --> OUT

    style IN fill:#e3f2fd,stroke:#1976d2,stroke-width:2px
    style FE fill:#fff3e0,stroke:#f57c00,stroke-width:2px
    style MOD fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    style OUT fill:#e8f5e9,stroke:#388e3c,stroke-width:2px
    style ENS fill:#e0f2f1,stroke:#00796b,stroke-width:2px

Forecasting Horizons¶

External Factors Integrated¶

Calendar: - Day of week, time of day - Public holidays - School terms - Major events (sports, concerts)

Weather: - Temperature - Rainfall - Severe weather warnings - Heat/cold waves

Special Factors: - Construction/disruptions - Fare changes - COVID/pandemic effects - Economic indicators

Implementation¶

Timeline¶

Team¶

Data Preparation¶

Data Sources: - Automatic Fare Collection (tap on/off) - Automatic Passenger Counters (buses) - Train car weight sensors - Weather bureau API - Events calendar (venues, sports) - School term calendar - Historical service changes

Data Volume: - 5 years of ridership data - ~2 billion journey records - 15-minute aggregation granularity - Stop/station level detail

Feature Engineering: - 156 features per stop-time combination - Lagged values (same time yesterday, last week, last year) - Rolling statistics (7-day, 28-day averages) - Fourier features for cyclical patterns - Holiday and event embeddings

Results¶

Forecast Accuracy¶

Operational Impact¶

Customer Experience¶

Financial Impact¶

Challenges and Lessons Learned¶

Challenge 1: Data Quality¶

Issue: Fare collection data had gaps and errors Solution: - Data quality pipeline with automated cleaning - Imputation for missing values - Confidence scoring for forecasts Lesson: Invest in data quality monitoring

Challenge 2: Event Handling¶

Issue: One-off events (new stadium, major concert) hard to predict Solution: - Similar event matching algorithm - Manual adjustment capability for planners - Rapid learning from actual outcomes Lesson: Combine AI with human expertise for unusual events

Challenge 3: COVID Disruption¶

Issue: Pandemic completely changed demand patterns Solution: - Regime detection for structural breaks - Weighted training with recent data - Scenario-based forecasting Lesson: Build for adaptability, not just accuracy

Challenge 4: Real-time Integration¶

Issue: Difficult to update forecasts fast enough for operations Solution: - Pre-computed forecasts adjusted with real-time signals - Edge computing for rapid updates - Graceful degradation when systems lag Lesson: Architect for speed and resilience

Challenge 5: Planner Trust¶

Issue: Planners didn't trust AI forecasts initially Solution: - Side-by-side comparison with traditional methods - Explainability showing key factors - Allowing overrides with feedback loop Lesson: Trust is earned through transparency and performance

Governance and Compliance¶

Governance Structure¶

• Executive sponsor: Chief Operating Officer
• Steering committee: Monthly review
• Technical governance: Data and analytics team
• Risk tier: Tier 1 (Low) - No individual decisions

Data Considerations¶

• No personal data used (aggregated counts only)
• Fare collection data de-identified
• Weather and event data from public sources
• Privacy impact assessment: Not required (aggregate data)

Model Monitoring¶

• Daily accuracy tracking by route/line
• Alert on accuracy degradation >5%
• Weekly performance review
• Monthly model refresh evaluation

Technical Details¶

Model Architecture¶

Prophet (Trend & Seasonality): - Daily, weekly, yearly seasonality - Holiday effects - Trend with changepoint detection - Base 40% of ensemble

LSTM (Sequence Learning): - 2-layer bidirectional LSTM - 168-hour (1 week) input sequence - Multi-step output (up to 7 days) - Base 35% of ensemble

XGBoost (Feature-Rich): - 156 engineered features - Weather and event integration - Handles non-linear relationships - Base 25% of ensemble

Ensemble: - Weighted average based on recent performance - Weights adjusted dynamically per route/time - Confidence intervals from all models

Infrastructure¶

• Training: Cloud HPC (weekly retrain)
• Serving: Kubernetes cluster
• Feature store: Real-time feature computation
• Cache: Pre-computed forecasts
• API: RESTful + streaming

Performance¶

• Throughput: 50,000 forecasts/minute
• Latency: <100ms per forecast
• Availability: 99.95%
• Storage: 2TB feature store

Recommendations for Similar Projects¶

Do¶

• Invest heavily in data quality
• Build for multiple forecast horizons
• Integrate external factors (weather, events)
• Plan for regime changes (COVID taught us)
• Combine AI with human expertise
• Start with high-volume, stable routes

Don't¶

• Expect 100% automation of planning decisions
• Ignore data quality issues
• Over-fit to historical patterns
• Deploy without planner buy-in
• Neglect model monitoring
• Assume patterns are static

Cost-Benefit Summary¶

Costs (First Year)¶

Ongoing Costs (Annual)¶

Benefits (Annual)¶

ROI: 252% | Payback: 5 months¶

Contact¶

For more information about this case study, contact the AI Toolkit team.

Related documents: AI Use Case Template | How to AI Monitoring