Digitalization paves the road for power networks to evolve towards smart grids, where new intelligent devices are numerous and communication networks are closely interrelated with the power system. Smart grids face new threats due to these networks, and cyber-attacks drive them to fail. One of the most common attacks is False Data Injection (FDI), which only needs a vulnerable device to fake data sent; therefore, it can control and make power networks react to an event that is not really happening, subsequently leading to a malfunction. In this study, a decision tree is trained with variations of the same FDI attacks' dataset to study how each variation affects the accuracy results to two different labels: location and status. To determine the accuracy sensitivity of the decision tree, eight variations for the training processes were used: (1) a complete and (2) reduced features dataset, (3) a real or (4) balanced dataset, (5) division criteria and (6) weight, and (7,8) two labels with different number of distinct values. The comparison aims to determine how the different number of distinct values in labels affects to the accuracy of the detection algorithm.