I have a time series (e.g., temperature data) from 1st January 2003 to 31st December 2017 with a one-second sampling rate, which indicates there are about $24 \times 3,600 \times 365 \times15= 473,040,000$ observations. I want to use the Pytorch to consume the time series and develop a neural network to perform forecasts. At first, we use the tensor with $473,040,000$ rows and one column to represent the data. I wonder whether there is a better presentation if your guys met similar data.

  • $\begingroup$ Do you really need one second sampling rate for temperature forecast? $\endgroup$
    – DKDK
    Jan 10 at 1:10