It is vital for adequate management, and operation of water distribution systems (WDS) to have reliable short-term water demand forecasts. Conventional time-series models present limitations when dealing with non-linear changes in water demand. Thus, it is proposed to employ deep learning algorithms to offer a more reliable forecast. Three models are used, two 1-dimensional convolutional neural networks (1D-CNN), (a simple CNN, and a dilated causal CNN), and a recurrent neural network, particularly a long short-term memory (LSTM) based model. The performance of the models is tested on seven real-life water distribution systems in Italy with different uses and number of users. Also, a comparison with benchmark algorithms based on time-window techniques and pattern-based models is made. Additionally, the use of meteorological variables such as rainfall occurrence, temperature, and relative humidity is intended to test whether there is a positive effect on the forecast. Furthermore, a global model is built taking several years of data for training to test whether this bigger model increases generalization and improves accuracy in comparison to the individual cases. In addition, transfer learning is employed to predict individual cases and a WDS in the Netherlands. Lastly, a bigger global model is built and trained with 14 years of data to improve the performance of transfer learning on the Dutch WDS. To begin with, it was seen that 1D-CNNs outperformed the LSTM-based model, and the benchmark algorithms using data of the water demand, and a binary index indicating whether it is a weekday or a weekend day for six of the seven case studies. For the remaining case study, the results indicated that there is less than 1% in error between the best benchmark model and the proposed 1D-CNN algorithms. Moreover, the addition of meteorological variables showed to improve the calibration performance of the models but worsened the predictions on unseen data. It was observed that a simple 1D-CNN overfits when adding these extra variables due to its lack of regularization. Also, the global model showed to improve in accuracy compared to the individual models. The use of transfer learning (TL) did not indicate to improve the performance of one of the case studies, nonetheless, TL showed that by only using 75% of the data for training, the model offers a good generalization on the case with sudden changes in demands by the rapid increase of users due to seasonal touristic activities. For the Dutch WDS, TL performed similarly to the individual model, there errors ranged between 15% and 16% using different quantities of data for training. In addition, when having no data for training, the pre-trained model displayed showed lower than using 25% of data for training for the Italian cases. Lastly, the bigger global model performed in the same way as the smaller global model on the Dutch WDS. Also, when having no data for training, the model performed better.