Evaluation of Machine Learning Algorithms for Multiclass Classification of Voice Calls from Power Systems Operations

The Brazilian electric system is operated by a nonprofit institution called the National Electric System Operator (ONS) which coordinates and controls the operation of energy generation and transmission. For this, the main form of operation occurs through verbal communication between the operators of the system. All telephone calls made between operators are recorded and stored on a server. In this sense, from the transcription and labeling of the audios, it is possible to use Natural Language Processing (NLP) and Machine Learning (ML) techniques to classify verbal communication in the operation. The Python programming language will be used together with a scikit-learn library to perform a comparison of multiclass classifiers, Complement Naive Bayes, Linear Support Vector Classification, Stochastic Gradient Descent Classifier, K-Nearest Neighbors Classifier, Multi-Layer Perceptron Classifier and Random Forest Classifier by tunning the hyperparameters and applying evaluation metrics for classification models. As result the model with best performance is the Multi-Layer Perceptron, reaching up to 85% of accuracy and F1 weighted. This work contributes to generation of products that assist the operation of this sector, it can produce performance indicators and even automatization of tasks and actions.