The original version is a little redundant since it should put the huge tensor into the model as the input. However, if the time interval is small, like 5 or 10 mins, the input may
be too huge for memory and lacking of efficiency during training. Therefore, I wrote a version called LSTNet_multi_inputs
which deconstructs the input as (1) short-term time series, like (t-3,
t-2, t-1, t) and (2) long-term skip time series, like (t-2xskip, t-skip,
t). The result is as good as the original one, but much faster.