In this paper a tensor-based approach is developed for calibration of binary self-paced brain-computer interface (BCI) systems. In order to form the feature tensor, electrocorticograms, recorded during behavioral experiments in freely moving animals (rats), were mapped to the spatial-temporal-frequency space using the continuous wavelet transformation. An N-way partial least squares (NPLS) method is applied for tensor factorization and the prediction of a movement intention depending on neuronal activity. To cope with the huge feature tensor dimension, an iterative NPLS (INPLS) algorithm is proposed. Computational experiments demonstrated the good accuracy and robustness of INPLS. The algorithm does not depend on any prior neurophysiological knowledge and allows fully automatic system calibration and extraction of the BCI-related features. Based on the analysis of time intervals preceding the BCI events, the calibration procedure constructs a predictive model of control. The BCI system was validated by experiments in freely moving animals under conditions close to those in a natural environment.