Help with Battery Capacity Prediction using LSTM

Hello everyone,

I'm trying to predict the capacity of a battery using a sequence-to-one regression model, but my current results aren't good, and I'm looking for some advice.

My Data:

• I have 1000 data sequences.
• Each sequence represents a single battery charfe and discharge cycle.
• The input features for each sequence are Voltage, Current, Temperature, and Time.
• The shape of each input sequence is [4 features x 3700 time steps].
• The output is the battery capacity, which is a single double value for each sequence.

My Model Architecture:

I'm using an LSTM network with the following layers:

Matlab

layers = [

sequenceInputLayer(4, 'MinLength', 3700, 'Name', 'input', 'Normalization', 'zscore')

lstmLayer(50, 'OutputMode', 'last', 'Name', 'lstm')

fullyConnectedLayer(40, 'Name', 'fc1')

reluLayer('Name', 'relu_fc')

fullyConnectedLayer(1, 'Name', 'fc_output')

regressionLayer('Name', 'output')

];

My Problem:

The model isn't providing useful results. I'm not sure what I might be doing wrong with my approach.

My questions are:

• Is this a suitable architecture for this type of problem?
• Are there any common issues I should check, such as data preprocessing (e.g., normalization) or hyperparameter tuning (e.g., number of hidden units)?
• Could the issue be in how I'm handling the Time feature? Should it be a separate feature, or is there a better way to represent it?

Any suggestions on how to improve my model or overall approach would be greatly appreciated. Thank you!

Answer 1

[Shishir Reddy]

Hi @hussein al-yasseri

Your architecture is a good starting point for sequence-to-one regression. Using an LSTM with OutputMode='last' is appropriate since you're mapping the full sequence to a single capacity value.

Here are the common issues that should be checked -

• Batch Size: Try using mini-batches of size 16 or 32.
• Learning Rate: Consider experimenting with a lower learning rate (e.g., 0.001).
• Sequence Padding: Make sure all sequences are the same length or properly padded.
• Train/Validation Split: Ensure a good mix of battery capacities in both training and validation sets.
• Overfitting: Check for overfitting, use early stopping or dropout layers if needed.

For more information regarding training deep neural networks, kindly refer the following documentation - https://www.mathworks.com/help/deeplearning/train-deep-neural-networks.html

I hope this helps.