Construction of an Outlier-Immune Data-Driven Power Flow Model for Model-Absent Distribution Systems

For many actual distribution systems, an accurate system model might not be available, so the operator has to fit an approximate power flow model over a set of field measurements. To completely shield the adverse impact of outliers that are unavoidable in a training dataset, this letter proposes a novel outlier-immune method to construct a data-driven linear power flow (DD-LPF) model that exhibits a much better out-of-sample accuracy than those constructed by common approaches. Moreover, a continuous relaxation-rounding algorithm is proposed to further accelerate the training process. The computational time of constructing this proposed DD-LPF model is satisfactory enough, which underlines its potential applicability for field applications.