ABSTRACT This work outlines the performance of three variants of deep neural networks for leak detection in water distribution networks, namely – autoencoders (AEs), variational autoencoders (VAEs), and long short-term memory autoencoders (LSTM-AEs). The multivariate pressure signals reconstructed from these models are analysed for leakage identification. The leak onset time is estimated using a fast approximation sliding window technique, which computes statistical discrepancies in prediction errors. The performance of all three variants is validated using the widely studied L-Town benchmark network. Furthermore, their feasibility for real-world application is studied by applying them to a real-world case study representing the data availability and network design often found in smaller- and medium-sized utilities in Norway. The results for the benchmark network showed that AE and LSTM-AE showed comparable detection performance for abrupt leaks with VAE performing the least. For incipient leaks, the LSTM-AE showed better detection performance with few false-positives. For the real-world dataset, the performance was significantly lower due to the quantity and quality of data available, and the contradiction of inherent requirements of data-driven models. In addition, the analysis revealed that the positioning of pressure sensors in the network is critical for the leak detection performance of these models.