Online Data-Driven Equivalent Model Derivation Based on Distribution Network Signatures Using a Machine Learning Approach

Parameters of distribution network (DN) equivalent models, developed via the measurement-based modelling approach are highly dependent on various factors, such as the loading conditions and the load mix of the DN. Therefore, their effectiveness for the analysis of various scenarios is rather limited. To address this issue, a new methodology is introduced in this paper to derive representative sets of parameters for DN equivalent models online. The proposed method uses DN signatures as features to train random forest regression algorithms that predict online sets of equivalent model parameters. DN signatures include harmonic current spectra and pre-disturbance network operating conditions. Comparisons with a conventional generalization technique and four other machine learning algorithms are carried out in terms of simulated responses from benchmark DNs to evaluate the efficacy of the proposed approach. The applicability of the proposed method is also demonstrated by means of experiments on a laboratory setup.