Hi CH,
If I understood the setup correctly, the objective is to predict battery SOH at 200 hours using an LSTM trained on 100-hour sequences (720 timesteps), where each sequence is labeled with a single SOH value at the 100-hour mark, using the stateful predict block: https://www.mathworks.com/help/deeplearning/ref/statefulpredict.html but the model is not learning the degradation dynamics associated with the SOH values.
- You can try using shorter sequences with intermediate SOH labels could help capture degradation dynamics better than a single endpoint label.
- Additionally, if the 'Stateful Predict' block resets states correctly between independent sequences; improper state retention might corrupt predictions.
- If data allows, you can also train with longer sequences (e.g., 150+ hours) to expose the model (with intermediate labels) to extended degradation patterns. You can also experiment with different architectures, layers/units to check if overfitting masks the real issue.
Hope this helps!
