Flood prediction in ungauged catchments is usually conducted by hydrological models that are parameterized based on nearby and similar gauged catchments. As an alternative to this process-based modelling, deep learning (DL) models have demonstrated their ability for prediction in ungauged catchments (PUB) with high efficiency. Catchment characteristics, the number of gauged catchments, and their level of hydroclimatic heterogeneity in the training dataset used for model regionalization can directly affect the model's performance. Here, we study the generalization ability of a DL model to these factors by applying an Encoder-Decoder Long Short-Term Memory neural network for a 6-hour lead-time runoff prediction in 35 mountainous catchments in China. By varying the available number of catchments and model settings with different training datasets, namely local, regional, and PUB models, we evaluated the generalization ability of our model. We found that both quantity (i.e. number of gauged catchments available) and heterogeneity of the training dataset used for the DL model are important for improving model performance in the PUB context, due to a data synergy effect. The assessment of the sensitivity to catchment characteristics showed that the model performance is mainly correlated to the local hydro-climatic conditions; the more arid the region, the more likely it is to have a poor model performance for prediction in ungauged catchments. The results suggest that the regional ED-LSTM model is a promising method to predict streamflow from rainfall inputs in PUB, and outline the need for preparing a representative training dataset.