This project involves a comparative analysis of Liquid Time-Constant (LTC) neural networks and Long Short-Term Memory (LSTM) networks. Both models were implemented and trained on the Beijing PM2.5 air quality dataset, aiming to evaluate their effectiveness in time-series forecasting tasks, specifically air pollution (PM2.5) predictions.
- Dataset: Beijing PM2.5 Dataset
- Time Span: January 1, 2010, to December 31, 2014
- Features Included: PM2.5 concentration, temperature, humidity, wind speed, and additional meteorological factors.
- Purpose of Choosing Dataset:
- Complex temporal dynamics with hourly granularity.
- Real-world noisy data making it suitable for evaluating robustness and predictive stability.
| Metric | LTC | LSTM |
|---|---|---|
| MSE | 675.55 | 584.20 |
| MAE | 14.32 | 13.00 |
| R² | 0.9289 | 0.9385 |
- Interpretation:
- LSTM slightly outperformed LTC based on accuracy metrics.
- LTC showed quicker convergence and smoother predictions, indicating robustness and better performance in noisy or irregular data environments.
- Detailed Hyperparameter Optimization: To further enhance LTC’s predictive performance.
- Robustness Testing: Compare models explicitly under varied noise conditions.
- Computational Efficiency Analysis: Measure training/inference speed and memory efficiency.
- Hybrid Approaches: Investigate hybrid architectures combining LTC and LSTM strengths.
This comparative analysis demonstrates clear trade-offs between LTC and LSTM, emphasizing considerations beyond accuracy alone, such as robustness and computational efficiency.