This paper contributes to traffic state prediction at a network scale. The prediction methodology relies on pattern recognition methods with the adaptation of a k-nearest-neighbors technique. The originality of the contribution is its ability to predict network-wide traffic states as time-dependent maps snapshots, hence keeping all the information and the spatial correlations between links at each time step. This non parametric process is trained and tested on the Brisbane network equipped with 79 bluetooth detectors. Bluetooth data are preprocessed and prepared for the prediction step. The results highlight the performances of the network global approach in terms of prediction error and computational complexity, making it suitable for an application at a large scale.