How can I improve the predictions of uncertainty of my LSTM model ?

Hello everyone,

I have some issues predicting the uncertainties of my linear Discrete State-Space with my LSTM model. On Simulink, I have created a DSS with two inputs and two outputs, then I added the signal of a Randon Number box (my supposed uncertainty) on the first input. The outputs are stable with and without the RN box.

My LSTM model is composed by those layers :

Networklayers = [sequenceInputLayer(featureDimension) ...   % featureDimension = 4 (2 inputs and 2 outputs)
                 lstmLayer(numHiddenUnits) ...              % numHiddenUnits   = 200
                 lstmLayer(numHiddenUnits) ...              % numHiddenUnits   = 200
                 dropoutLayer(0.2) ... 
                 batchNormalizationLayer ...
                 fullyConnectedLayer(numResponses) ...      % numResponses     = 1 (the uncertainty)          
                 regressionLayer];

I'm training my model with the 2 inputs and 2 outputs as a matrix of data and the uncertainty as the Target.

Problem is, my uncertainty prediction doesn't match the Simulink's uncertainty but seems to follow the inputs and outputs... In a previous test, I managed to make a prediction of the uncertainty only with a very high amplitude, wich wouldn't be consistent at all...

(I normalized both my matrix data and the Target, I have 80,000 values for my Dataset but only with 7 different pseudo-periods and I always finish my trainings above a RMSE of 0.1 on a plateau)

Is my task possible with a LSTN network ?

Thanks for your time

Answer 1

[Aditya Srikar]

It's not clear from your explanation what your inputs, outputs, and targets are, but it seems that predicting the uncertainty of your linear DSS with an LSTM model is not working as expected. It's possible that the LSTM model is not well-suited for this type of prediction, or there may be issues with your training data or network architecture that are affecting the performance. You may want to try different approaches, such as using a different type of model or adjusting the training data to improve the prediction. It may also be helpful to examine the training data and output to identify potential problems or areas for improvement.