expanding to RGCN

[jzwart]

I have written some code to incorporate RGCN, but there have been some additions to LSTM that needs to be incorporated when predicting with RGCN (e.g. mc dropout)

[jzwart]

RGCN training takes much longer - ~20x's longer / epoch. But predicting is much faster for RGCN than LSTM

[jzwart]

Having some issues with getting RGCN to converge on accurate predictions in the training phase - training RMSE hanging out around 6-7 C for temperature. Currently using 20 hidden units, 30 epochs pre-train, 250 epochs fine tune, 4 connected stream segments. Most likely a bug somewhere in the code and need to revisit later