Handling concept drift in data-oriented power grid operations

Abstract

Data-oriented business transformation, also known as “digitalization”, can improve business tasks by providing better insights into the subject through the data. In digitalizing the power grid, more accurate state recognition from the measurement data is expected to promote a low-cost and stable power supply. Acquiring measurement data from the power grid, clustering, and anomaly detection to recognize the current state could lead to better decision-making for power grid operations. While measurement data serves as the starting point, the interpretation of data trends changes due to the influence of the surrounding environment and aging in the real world. This change in data trends, known as concept drift, poses a challenge to efficient data-oriented power grid operations with accurate state recognition using data clustering models. This is because the data clustering model, especially for complex systems like a power grid, is also built data-oriented, and data trends affect the model. To address this combined challenge of concept drift and its impact on the data clustering model, we propose Re-DBSCAN, a stream data clustering model capable of handling uncertain distributions, to detect concept drift and sequentially update its model for data streams from the power grid. The evaluation uses the WECC179 power grid model to simulate power oscillations and their trend changes with the basic concept drift types: abrupt, incremental, and gradual. Compared to other stream data clustering methods that lack a concept drift detection mechanism, the proposed Re-DBSCAN showed less degradation in purity, indicating higher clustering accuracy. The results suggest that by handling concept drift by detecting data trend changes and sequentially adapting the clustering model, Re-DBSCAN can more accurately cluster measurement data containing concept drift based on its trend changes.