Fast Computation Methods for Constructing Data-Driven Linear Power Flow Model

Abstract

To balance the construction computational efficiency and linearization accuracy of data-driven linear power flow (DD-LPF) models, this paper provides two fast computation methods to resolve two technical issues. First, for the issue of slow computation of the unknown LPF model parameters resulting from the large scale of historical data, this paper suggests employing a distributed stochastic gradient descent (DSGD)-based method to solve the problem in a distributed parallel mode, rather than in a batch way. Second, for online applications, to improve the linearization accuracy and to maintain high solution efficiency, this paper suggests using a recursive least squares (RLS) method to dynamically update the DD-LPF models. Case studies confirm that the DSGD-based method and the RLS method can all improve the computational efficiency with 2- to 50-fold less computing time and keeping a satisfactory linearization accuracy.