I am working on a case where I want to do a multivariate and multi-step time series forecasting.

I have hourly data that measures temperature at approximately 500 different devices. (the devices have different locations). The measurements spans from two years of measurement on some of the device to about two months of measurements for the ones with the lesser data.

What I have done so far is to create one LSTM-model based on data from one of the devices. This data uses 72 hours of lagged temperature and time of day as input and predicts the next 24 hours. But I would like to utilize all of the data so that I can make a general model that is trained on data from all of these devices. But if I merge all of these datasets together and group them by id and sort by date I will have overlapping locations in my training and validation set.

Is there a way to combine datasets that would work as input for a LSTM-model with 3-dimensional input?

I have thought of moving away from "timestep" modeling the data so that the each row of data to the input-layer will have features that represents the historical data. I.e. a more traditionally 2-D, model.

I presume I would loose the benefits of RNN by going this way, since I then would have a one-step state-full model?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.