The calibration of a voltammetric sensor consisting of an array of individually modified electrodes is described. Linear calibration methods do not yield good results because of the inherent non-linear nature of the data. Neural networks can in principle model such dependencies, but their success is crucially dependent on the representation of the data. In this paper, neural networks and Principal Component Regression using several different data representations are compared. It is concluded that neural networks using unsealed first-derivative voltammograms yield the best results.