Auditing the Data Curation Practices of Power Systems Datasets

Machine learning (ML) methods are being increasingly applied to power system applications to address stochasticity, observability, and tractability challenges of modern grids modeling, analysis, and control. As ML applications in power systems expand, the field needs rigorous, systematic, and reflexive standards for dataset development. Power systems data, however, is heterogeneous and tailored for specific tasks and difficult to reuse because it is often either proprietary and restricted for public-access or anonymized and non-reproducible. In this paper, we perform a domain-specific audit to assess the current state of dataset development for ML applications in the power system field through the application of a documentation evaluation framework. As a result, we provide (1) a deep dive of datasets and data curation in power systems revealing existing strengths and weaknesses, (2) priorities and recommendations for how ML work in this domain can improve dataset practices, and (3) a power system dataset, fully documented to the standards of excellence we used for evaluation, as an ideal benchmark case and resource demonstrating how datasets in this cyber-physical domain should be documented.