STIP：智能公共交通分析平台的技术架构与机器学习应用

STIP (Sofia Transport Intelligence Platform) is a production-grade full-stack real-time public transport analysis platform for Sofia, Bulgaria. It leverages .NET and Python machine learning (ML) technologies to achieve digital transformation of public transport. Key functions include real-time vehicle tracking, delay prediction, reliability scoring, and service alerts. This platform not only demonstrates technical feasibility but also provides practical value for passengers, operators, and city governance.

With accelerating urbanization, the efficiency and reliability of public transport systems have become critical to urban residents' quality of life. Sofia's STIP project addresses this need by offering a complete intelligent transport analysis solution, enabling the digital upgrade of public transport systems beyond a mere technical demonstration.

STIP adopts a microservices architecture with clear layers:

• Data Source Layer: GTFS Static Feed (lines, stations, schedules) and GTFS-RT (vehicle positions, trip updates, service alerts).
• Backend: ASP.NET Core Ingestion API, SignalR Hub (real-time communication), .NET Worker Service (data aggregation, ML retraining).
• Storage: PostgreSQL+PostGIS (spatial data, historical delays) and Redis (real-time data cache).
• ML Layer: Python FastAPI service with XGBoost as the core prediction model.
• Frontend: React19+TypeScript, Leaflet (map rendering), Recharts (visualization).

Core functions include:

1. Real-time Vehicle Tracking: Polls GTFS-RT data every 15s, pushes to frontend via SignalR, and displays on Leaflet maps.
2. Delay Prediction: Uses XGBoost model with features like time (hour, week, holiday), line type, station type, historical average delay, and real-time status.
3. Reliability Score: Formula: (On-time rate ×100) - (Average delay minutes ×5), providing a quantifiable metric for service quality.
4. Site Analysis: Congestion heatmaps, predicted arrival times, transfer hub identification, and site-level reliability scores (powered by PostGIS spatial computing).

Engineering highlights:

• Elastic Design: Polly for retry/fuse (GTFS data acquisition), optional data sources (system degrades gracefully), service isolation (microservices).
• Dev Experience: Docker Compose (one-click startup), detailed .env.example, layered project structure (Domain/Application/Infrastructure/API), Swagger API docs.
• Data Architecture: PostGIS for spatial data storage, spatial indexes (accelerate region queries), partition tables (manage historical delays).

Application value:

• Passengers: Optimal route planning (real-time + prediction), wait time estimation, service alerts.
• Operators: Line optimization (via reliability scores), resource allocation (peak analysis), data-driven decision support.
• City Governance: Transparent traffic data (enhance public trust), reduced private car use (environmental benefits), lower operational costs (economic efficiency).

Future directions:

• Multi-city Adaptation: Configure GTFS data sources for other cities (with API key guides).
• Function Expansion: Integrate weather data (improve prediction), add passenger flow prediction, develop mobile apps, and include more data sources (social media, traffic cameras).
• Tech Evolution: Explore deep learning models (LSTM, Transformer), real-time stream processing (Apache Kafka/Flink), and edge computing (onboard local inference).