Deep learning methods with seasonal data I have been building an LSTM model in Python which will predict the number of passengers arriving at a station in the next 15 mins.
My dataset is arrivals at the station every 15 mins across a 50 day period. A plot of the data shows a clear seasonal pattern i.e. most entries to the station occur during the evening rush hour.
My question is this: prior to training an LSTM model, do I need to clean up the data - by this I mean take differences to make it stationary? I have normalised the data as all the examples I have seen say it's best to do this, but I am unclear about whether the trend/seasonality should be removed. 
 A: I‘m not familiar with LSTM, but I suggest keeping the trend in the data. Otherwise, the model will not be able to learn this trend. 
Also, if LSTM can manage it, you should add the time of the day as one of your features. 
A: No, you do not need to make it stationary - there is no such theorem. You can try it, and maybe it will give you better results, since stationary series is easier to forecast, but it is no rule. 
As a rule, if there is a trend, then the series should be differenced (LSTM often fits trends badly, though it is no rule), but seasonality is easily predicted using LSTM given sufficient steps back in time (if you split it into subsequences, which you have to do almost always due to performance issues, these subsequences have to be of sufficient length to cover any seasonality you want to account for). So, difference if there is a trend, leave it be if there is a seasonal pattern.
Stationarity is important only in relation to ARIMA-type models. LSTM in theory can discover non-linear relationships, and has no assumption about error distribution, and, thus, can go around non-stationarity. 
