Long-short term memory (LSTM) networks demonstrated their capabilities in rainfall-runoff modelling or streamflow prediction. Using metereological input data, most existing work has concentrated around training basin-specific networks. There exist approaches which, in addition to the meteorological forcings, also incorporate static attributes of basins. Further, initial work was done to allow for uncertainty quantification by training Bayesian LSTMs. In this paper, the mentioned approaches are combined. Basins from the novel CARAVAN dataset are used to train an LSTM-based encoder-decoder network. Static basin attributes, which are also provided with CARAVAN, are added to the embedding produced by the encoder and then fed through a plain feed forward neural network. Uncertainty quantification is achieved by Monte Carlo Dropout (MCD). Achieving roughly similar loss and accuracy statics, the prediction results from this architecture are broadly in line with existing work on deep learning in the field of hydrology.