Feature Engineering-Empowered MLOps for Smart Grid Failure Prediction Software

Unscheduled failures present significant challenges to the reliability of the smart energy distribution systems. This research project reports the complete development and design project for a failure prediction software, aiming at comprehensively analyzing data on unscheduled power failures in four medium-voltage feeders. First, the data of planned and unplanned interruptions, as well as the weather information, are pre-processed. Then, the correlation of the weekly failure with failure characteristics, different scenarios of maintenance activities, and weather characteristics of the relevant week is investigated. Next, the same steps are conducted to check the existence of mutual information between the stated items and weekly failures, followed by a causality analysis on these features. Subsequently, the best fit hyperparameters are derived for predictive modeling of the weekly failures, using grid search in the form of classification. Then the best model with the best performance is selected, based on the different subsets of the extracted features identified in the feature engineering process. Finally, a user interface software is designed to visually present the results of the analyses and predictions. This study not only provides a feature engineering-empowered MLOps software to gain insights into the factors affecting the failures, but also develops tools contribute to improving preventive maintenance strategies and enhancing the reliability of smart grids.