🚀 TimeForge: AI/ML Models for Satellite Clock & Orbit Error Prediction

Live Demo

---

📘 Project Overview

TimeForge is an advanced AI/ML solution developed for Smart India Hackathon 2025 that predicts satellite clock and orbit errors using 7 days of historical data to generate 24-hour forecasts with confidence intervals.

Detail	Information
Problem Statement ID	SIH25176
Theme	Space Technology
Category	Software
Team ID	97062
Team Name	TimeForge

---

🧠 Technical Solution

🔹 Core Approach

• Input: 7 days of satellite data (15-minute intervals)
• Output: 24-hour forecasts with uncertainty quantification
• Key Innovation: Hybrid Physics + Machine Learning ensemble approach

🔹 Model Architecture

• Multi-Model Ensemble: Gaussian Process, SARIMA + NN, Mini Transformer, Bayesian NN
• Physics-Informed Features: Orbital elements, eclipse detection, geopotential effects
• Optimization: Shapiro–Wilk residual normalization with aggressive regularization

---

🏗️ System Architecture

⚙️ Backend Components

• Data Processing: Feature extraction and preprocessing pipeline
• Model Training & Inference: Multi-model ensemble with automatic retraining
• AI Agent: Interactive Q&A system using LangChain / LangGraph
• API Layer: FastAPI for data exchange and model serving

💻 Frontend Components

• Interactive Dashboard: Real-time visualization of predictions
• Analysis Reports: Automated insights and performance metrics

---

🧰 Technology Stack

Layer	Technologies
Backend & AI	Python 3.8+, PyTorch, FastAPI, LangChain, Pandas, NumPy, Scikit-learn, SciPy
Frontend	React.js (JavaScript), Tailwind CSS, Vite
Infrastructure	PostgreSQL, Docker, AWS Cloud

---

📂 Project Structure

TimeForge/
├── backend/
│   ├── agent/               # AI Q&A system (LangChain)
│   ├── main.py              # FastAPI entry point
│   └── requirements.txt
│
├── frontend/
│   ├── src/
│   │   ├── components/      # React UI components
│   │   ├── hooks/           # Custom React hooks
│   │   └── App.jsx          # Main app component
│   └── package.json
│
├── Model/
│   ├── data_processing/     # Feature extraction & preprocessing
│   ├── gaussianization/     # Residual optimization scripts
│   ├── modelside.py         # Core model implementations
│   └── requirements.txt
│
└── documentation/           # Project documentation & reports

---

📊 Key Features

🎯 Accuracy & Performance

• R² Score: 0.677 (ensemble performance)
• Residual Normality: Shapiro–Wilk p = 0.439
• Uncertainty Quantification: 95% prediction intervals

🧩 Technical Advantages

• Small-Data Ready: Optimized for ~100 points per satellite
• Real-Time Capable: 15-minute forecast intervals
• Multi-Constellation Support: GPS, GLONASS, Galileo

---

🧪 Shapiro Test Results

Visual representation of residual distribution and normality.

GEO Satellite

MEO Satellite

MEO2 Satellite

---

⚙️ Installation & Setup

🧩 Backend Setup

cd backend
pip install -r requirements.txt
uvicorn server::app --reload --port 8000

🖥️ Frontend Setup

cd frontend
npm install
npm run start

🧠 Model Training

cd Model
pip install -r requirements.txt
python gp_clock_pipeline_modelside.py

---

🚀 Usage Guide

1. Satellite Selection: Choose constellation and specific satellite
2. Forecast Configuration: Set prediction horizon (1–24 hours)
3. Real-Time Monitoring: View predictions with confidence intervals
4. Analysis: Access error metrics and performance statistics

---

📈 Performance Validation

Metric	Result
Validation Samples	142
R² Score	0.677
Shapiro–Wilk p-value	0.439
Error Range (Normalized)	0.0003 – 0.0006
Temporal Resolution	15-minute intervals

---

🌍 Benefits & Impact

🛰️ Technological

• Enables high-precision navigation for autonomous systems
• Advances GNSS error correction methodologies

💰 Economic

• Reduces dependency on costly correction systems
• Optimizes resource usage in logistics and transportation

🎓 Educational

• Hands-on experience with real satellite data and AI/ML
• Promotes space-tech innovation among young engineers

---

⚠️ Challenges & Mitigations

Challenge	Mitigation
Limited Data	Physics-informed features + data augmentation
Satellite Variability	Constellation-specific modeling
Forecast Drift	Ensemble methods + regular retraining
Model Stability	Aggressive regularization + deep ensembles

---

🔭 Future Enhancements

• Integration with additional satellite constellations
• Enhanced real-time data streaming capabilities
• Federated learning for distributed model training
• Mobile application for on-the-go analysis

---

📚 References

1. Zhang, L., et al. (2024). Deep Learning CNN–GRU Method for GNSS Deformation Monitoring Prediction.

2. ResearchGate (2023). Fast and Reliable Forecasting for Satellite Clock Bias Correction with Transformer Deep Learning

3. ISRO (2023). GNSS Signal Monitoring & Performance Evaluation Facility.

4. NASA / ION PLANS (2025). Statistical Analysis of GNSS Multipath Errors.

---

👥 Team – TimeForge

Smart India Hackathon 2025 | Space Technology Track
Problem Statement: SIH25176
Team ID: 97062 | Team Name: TimeForge

---

_{© 2025 TimeForge Team | Designed for Smart India Hackathon}