In Bois Rouge sugar mill, the sugar cane crystallisation process is performed in three stages. During the third stage: C crystallisation, the mother liquor exhaustion is the main function, because it is the ultimate step of transforming saccharose in liquid form into crystals. In order to build a predictive scheme to control the C crystallisation, we decided to identify a multi-step ahead predictor of the mother liquor purity. The identification is performed using databases collected in situ. The present work deals with the four different steps of classical identification, from the databases to the predictor performances. In an industrial context, the measurements are often noisy or incomplete, so we have developed a neural network model well adapted to this type of data. Three different predictors have been identified and their performances show that an adaptive form of the predictor is the most appropriate for performing this task.