MetroCast AI is an end-to-end, AI-powered weather forecasting system. This project combines data collection, deep learning model training, a high-performance backend service, and a modern web interface to provide accurate 7-day temperature forecasts.
The project consists of three main components:
- HYS (Weather Forecasting System) - Training: Python-based module where deep learning models are trained.
- Backend (Rust) - API: Low-latency prediction service developed using Rust and Axum.
- Frontend (Web) - Dashboard: User-friendly visualization interface built with React, Vite, and Tailwind CSS.
At the heart of the system lies the ExcelFormer architecture, a specialized Transformer variant optimized for tabular and time-series data.
- Model: ExcelFormer (Advanced Attention-based Transformer)
- Input Window: 168 hours (7 days) of historical weather data.
- Output Horizon: 168 hours (7 days) of future temperature predictions.
- Data Source: Open-Meteo Archive API.
- Technologies: PyTorch, Pandas, NumPy, Weights & Biases (WandB).
- Chronological Data Splitting: Data is split into training, validation, and test sets according to time order to prevent leakage.
- Early Stopping: Monitors validation loss to prevent overfitting.
- ONNX Export: The best-trained model is converted to ONNX format for efficient high-speed execution in the backend.
- WandB Integration: Real-time tracking of training metrics (Loss, MAE, RMSE).
The prediction serving layer is developed in Rust, focusing on performance and safety.
- Framework: Axum (High-performance web framework).
- Inference Engine: ORT (ONNX Runtime) - Enables ultra-fast model execution on CPU/GPU.
- Documentation: Integrated Swagger UI via Utoipa (
/swagger-ui). - Data Processing: Handles input normalization and inverse transformation of model outputs back to Celsius.
POST /predict: Receives historical weather data and returns a 7-day forecast.GET /health: System health check.GET /swagger-ui: Access to OpenAPI documentation.
A modern interface where users can examine forecasts and monitor system performance.
- Dashboard: Sleek card designs for current and weekly weather data.
- Weather Charts: Interactive temperature variation charts powered by Recharts.
- Intelligence Console:
- Visualizes training history with "Loss" and "Accuracy" (MAE) graphs.
- Direct integration of WandB telemetry into the UI.
- Detailed view of model architecture and hyperparameters.
- Modern UI: Premium dark-mode design built with Tailwind CSS and Shadcn UI components.
- Multi-language Support: Turkish and English language options.
The system is architected for scalability and reliability using Amazon Web Services (AWS).
- Compute (EC2): The Rust-based backend is containerized and deployed on AWS EC2 instances, ensuring high availability and low-latency API response times.
- Storage (S3): Historical weather data and processed CSV files are stored in Amazon S3 buckets, acting as a data lake for the training pipeline.
- Security: Traffic is managed through AWS Security Groups, with port 3000 exposed for the Axum API.
- Python 3.10+
- Rust (Cargo) 1.75+
- Node.js 18+ & npm
- Docker (Optional)
cd backend
# Make sure best_model.onnx is in models/ or backend/models/
cargo run --releasecd web
npm install
npm run dev# Install dependencies
pip install -r requirements.txt
# Fetch fresh data
python data/fetch_data.py
# Run training
python src/training/train.py| Domain | Technologies |
|---|---|
| Languages | Python, Rust, TypeScript |
| AI/ML | PyTorch, ONNX, Scikit-learn |
| API | Axum, Utoipa, Serde |
| Frontend | React, Vite, Tailwind CSS, Recharts |
| Inference/Ops | ORT, AWS S3, Docker, WandB |
This project is licensed under the MIT License.
Developed by: Enes Demir