I'm trying to forecast a time series based on several influencing factors. For example, forecasting how many shoppers per hour there will be in my shop tomorrow will depend on three things:
- Yesterday's weather and shoppers/hour
- Tomorrow's weather
- How many tourists there are in town
The first two are easy - I've used a sequence to sequence RNN using LSTM cells and achieved fairly good results. I've used 24 hours on both the encoder and decoder at 1 hour intervals. Before you suggest it - yes I'm planning on trying GRU :)
Now, the tricky part is how to deal with the tourists. The tourists come on massive cruise ships, which arrive at various times. They stay for up to a week but tend to come to the shop more toward the beginning and end of their stay. To deal with this I can produce a time series of the capacity of all the cruise ships currently docked going back a full week. This will change over the week as the ships come and go.
I could increase the length of the encoder network and include this cruise ship time series, but that would waste resources by looking at past weather where it's really not required.
What I'm thinking is to have multiple LSTM encoders that somehow have their states combined before being fed to the decoder. I've done some googling around this and can't find anything about having multiple encoders on different time scales.
My question is: has anyone seen any network that can elegantly handle multiple time series over different time scales?