Time Series Prediction
Deep learning approaches for astronomical time series classification and prediction
Project Overview
This project explores advanced machine learning techniques for time series analysis, with a focus on astronomical data. The primary dataset consisted of photometric measurements from exoplanet WASP-126 b, presenting unique challenges due to noise, irregular sampling, and transit detection requirements.
The project compared traditional methods (Dynamic Time Warping with k-NN) against modern deep learning architectures (LSTM and Transformers) for both classification and prediction tasks.
Technical Approach
- Dynamic Time Warping (DTW): Implemented DTW distance metric for time series classification, handling temporal distortions and phase shifts in the data.
- LSTM Networks: Developed recurrent neural network models to capture long-term dependencies in the photometric sequences.
- Transformer Architecture: Implemented attention-based models for superior sequence modeling, achieving the best prediction performance.
Key Results
RMSE Improvement
Transformer models achieved 92% better precision than baseline methods
Multi-step Forecasting
Accurate predictions up to 24 time steps ahead
Transit Detection
High accuracy in identifying exoplanet transit events
Model Comparison
| Model | MSE | Parameters |
|---|---|---|
| LSTM | 0.00000161 | 125K |
| Transformer | 0.00000160 | 310K |
Technologies Used
PyTorch
LSTM
Transformer
DTW
NumPy
Matplotlib