1
$\begingroup$

I have an assignment of modelling electricity consumtion based on time series defined in two different time frames, say one has a resolution of 5 seconds and the other one is 1 hour.

I want to use LSTM here. Since the training input for a LSTM network has a shape like [samples, time steps, features], one cannot fit data from different time frames (it's possible, but putting them together on the features axis probably won't be efficient)

What's the standard approach for such problems? One idea coming to my mind is to train two LSTM networks separately, and then concatenate their output, feed it into dense layer and train it (while having the LSTM layers frozen).

Any hints and insights highly appreciated.

$\endgroup$
1
$\begingroup$

To deal with your problem, it is important to know how LSTM works exactly: when you have two linked variables with a very different pace, the LSTM parameters for forgetting or memorizing will probably have to do a big gap that could alter the predictions. Like any other physic ML problem, your variables need to be at a "common sense" scale, and adapted to your prediction's scale objective. For example, if your prediction is at day scale, you could take a hourly rate for both variables (taking the average of the 5 seconds' one). If your prediction is at hour scale, you could take a 15 minutes rate for both variables (taking the average of the 5 seconds' one and duplicating the value or the trend value ~regression of the other one). By the way, if you know how the hourly variable behaves approximately between hours, it could be interesting to simulate the inner values. Remember that LSTM is quite sensitive to noise: the less noise the better. LSTM has also memory limits: the learning time steps should be reasonable (about 100-500, 1000 max).

$\endgroup$

This site is temporarily in read only mode and not accepting new answers.

Not the answer you're looking for? Browse other questions tagged .