Accurate prediction of water consumption and demand is essential for effective resource management and promoting sustainable development, particularly in the face of challenges posed by natural resource scarcity. This paper presents a comparative study between Deep Learning and conventional Machine Learning algorithms for the prediction task of urban water demand. A novel Deep Learning approach combining Autoencoder with Long Short-Term Memory network called AE-LSTM is applied to predict hourly and daily water demand using historical consumption data from households within the Valencia metropolitan area, Spain. The model’s performance is evaluated against well-known Machine Learning techniques like Support Vector Machines for Regression (SVR) and Random Forest (RF) using various error metrics, including RMSE, MAE, and POCID. Experimental results demonstrate that the hybrid AE-LSTM model outperforms other techniques for both daily and hourly prediction horizons, establishing it as a reliable method for accurate short-term water demand prediction task. The study concludes that integrating innovative Deep Learning models into water management systems can offer a valuable solution, significantly enhancing effectiveness in addressing the growing global demand for water resources.