Machine Learning for Self-Calibration Parameters of Data-Driven Models: Case Study of an Integrated Maintenance Digital Platform

Abstract

A well-established problem always in the process of methodological improvement is the coordination of generator maintenance scheduling in Power Systems. Usually, the sources of changes and improvements are the wide range of conditions and singularities of Power Systems, such as structure, needs, technologies, resources, information, etc. The risk-oriented approach is one of the most accepted criteria to address this well-defined problem. However, there are hidden challenges of data-calibration-modeling and technology-integration when applying this approach. One of the most controversial is the calibration of model parameters and variables in the evaluated scenario, as they depend on historical data. In this paper, to address this controversial hidden challenge, we propose to use machine learning for on-line calibration by introducing smart layers based on comprehensive human diagrams that filter and analyze historical data collected through monitoring systems and make all variables and parameters of the risk model ready for use. In particular, we focus the attention on one of the model components, the load modeling. This proposed organic connection ensures a feasible and practical solution to be used in a real system because it addresses a potential challenge in technology integration.