A disruption prediction system, based on neural networks, is presented in this paper. The system is presently running in real-time on the JET tokamak. A Multi Layer Perceptron predictor module has been trained on nine plasma diagnostic signals extracted from 86 disruptive pulses, selected from four years of JET experiments in the pulse range 47830-57346 (from 1999 to 2002). The disruption class of the disruptive pulses is available. In particular, the selected pulses belong to four classes (Density Limit/high Radiated Power, Internal Transport Barrier, Mode Lock, and H-Mode/L-mode). A Self Organizing Map has been used to select the samples of the pulses to train the Multi Layer Perceptron predictor module, and to determine its target, increasing the prediction capability of the system. The prediction performance has been tested over 86 disruptive and 102 non-disruptive pulses. The test has been performed presenting to the network all the samples of each pulse sampled every 20ms. The missed alarm rate and the false alarm rate of the predictor, up to 100ms prior to the disruption time, are 23% and 1% respectively. Recent plasma configurations might present features different from those observed in the experiments used in the training set. This 'novelty' can lead to incorrect behavior of the predictor. To improve the robustness and reliability of the system, a Novelty Detection module has been integrated in the prediction system, increasing the system performance, and resulting in a missed alarm rate reduced to 7%, and a false alarm rate reduced to 0%.