Presented at ICICIS 2023, this paper analyses how hyperparameter optimisation improves convolutional and CNN-LSTM models for predicting the remaining useful life (RUL) of lithium-ion batteries.
Methodology
- Leveraged Bayesian optimisation to tune filter sizes, learning rates, and sequence lengths for both architectures.
- Evaluated on NASA battery cycling datasets, comparing mean absolute error and prognostic horizon metrics.
- The CNN-LSTM variant captured long-term degradation trends, while the pure CNN excelled at faster inference.
Results
- Optimised models outperformed manually tuned baselines across accuracy and stability.
- Hybrid CNN-LSTM achieved the best balance between prediction fidelity and computation.
- Sensitivity analysis highlighted which hyperparameters most strongly influence RUL performance.