Using Reduction Techniques to Obtain a Minimal Consistent Data Set for Training Artificial Neural Networks for Fault Identification Problems

This work evaluates methods to reduce the size of training data sets for artificial neural networks (ANN) learning, focusing on power systems problems such as fault identification in transmission lines. Usually, many simulations must be performed to obtain a training data set that is representative enough. Furthermore, the training data set is usually composed of elements generated by moving windows. So, each simulation can generate several elements, thus leading to a large data set that could make the ANN training and testing time-consuming. Therefore, using techniques to reduce the size of this training data set while keeping it representative is desirable. Three techniques that could be used for this purpose are presented and the steps to their use are detailed. Time-domain simulations considering random values for all parameters related to transmission line faults are performed to achieve statistical robustness for the results. The results show that the k-means-based technique presented a good performance while keeping mean errors and standard deviations slightly greater than the original condition for the fault classification problem. These promising results suggest the study of other techniques and cases to better understanding and improvement of the procedures.