This study investigates the application of machine learning tools to predict sludge decantability in wastewater treatment. Machine learning covers all the methods used to design models from experimental data. Focusing on the activated sludge treatment process, we use five years of monitoring data from a large Wastewater Treatment Plant (WWTP). The aim is to develop models using automatic water quality monitoring data to estimate the quantities characterising the sludge's settling properties (SV30, Sludge Volume Index (SVI) and Filamentous Index (FI)). Indeed, measuring these properties presents a number of operational difficulties and financial constraints. Through the application of different pre-treatment methods, we develop two dynamic models based on Recurrent Neural Networks (RNN) that could provide estimates for a 100 days horizon with an uncertainty level similar to that of real measurements.