Bad data detection and identification is one of the most important problems to be solved in real-time power system monitoring. During system operation, the decision-making process is based on analyses that use a database which is assumed to be reliable. Bad data can affect the results of these analyses and as a consequence the decisions taken may not be valid anymore. This may cause serious problems to system operation. This work presents a new method for debugging data in real- time power system monitoring. Data projection tecniques based on Kohonen´s self-organizing maps are employed to show that normalized innovations, obtained from a forecasting-aided state estimator, present excellent discrimination capability when compared to other variables such as raw measurements and normalized residuals. In the proposed method the problem of bad data identification is viewed as a pattern recognition problem, in which normalized innovations are use as input variables to a constructive artificial neural network that is responsible for identifying bad data. The method is able to distinguish between gross measurement and topological errors. Which can include branch or bus misconfigurations. The proposed method is tested for many different operating conditions involving different types of error. Tests are performed using data from the IEEE 24-bus and IEEE 118-bus systems. The performance of the method is evaluated and aspects such as computational efficiency, generalization capability and real-time implementation, among others, are also discussed.