Adaptive learning for disruption prediction in non-stationary conditions