Network architecture for time series data

Deep Learning
1

We have a time series dataset (spatiotemporal, but not an image/video). The dataset is in 3D, where each (x,y,t) coordinate has a numeric value (such as the sea temperature at that location and at that specific point in time). So we can think of it as a matrix with a temporal component. The dataset is similar to this but with just one channel:
Dataset

We need to predict/forecast the future (next few time steps) values for the whole region (i.e., all x,y coordinates in the dataset). I considered using a ConvLSTM or CNN-LSTM, but most online posts seem to be applied to video frame prediction. Since it’s not a video, I only have one channel for each time instance.

Can you all suggest an architecture that would suit my purpose well, and is there any built-in into fast.ai?

1 Like
2

Hi @microlifecc

I have never done anything like this, so take my words with a grain of salt. But I think you can pretty well use everything you find on video. Video is 3D and also has x,y and t. x and y are basically the pixel positions, and I would think you can pretty much exchange pixel position in a video / image, with a “pixel position” of some grid at sea.

The difference in channels shouldn’t be a big issue, you can build a custom Conv LSTM where the conv layer just uses one single input channel instead of 3.

I am pretty sure there isn’t an architecture within fastai that exactly does what you need, so you will probably have to hand-build a model.