Data-Driven Dynamic Models of Active Distribution Networks Using Unsupervised Learning Techniques on Field Measurements

The development of realistic dynamic models for active distribution networks (ADNs) is a challenging problem faced by both the transmission system operators and academia due to the random and complex nature of loads and distributed generation. Real field measurements are often not available and when they are, they usually contain various dynamics, irrelevant data, and outliers. This article proposes a three-stage methodology to effectively build a set of dynamic models for an ADN based on field measurements. In the first stage, an unsupervised learning method identifies and then discards all the irrelevant data for the model parameter estimation. In the next stage, the remaining data are clustered into groups with similar dynamics while in the third stage, a nonlinear dynamic model is developed for each of the derived groups. This papers concludes that, in spite of the large number of measurements representing a wide range of grid configurations, the general dynamic characteristics of an ADN can be accurately captured and modeled using a limited number of clusters. The proposed method is validated using real data from six substations in Southern Germany.