Synchrophasor Based ZIP Parameters Tracking Using ML With Adaptive Window and Data Anomalies

Abstract

Modeling of load-voltage dependency can enhance the accuracy of system analysis (e.g., voltage stability and distance to collapse). ZIP load model is one of the standard mathematical models used to represent load-voltage dependency. Synchronized measurements from Phasor Measurement Units (PMUs) can be utilized for the online estimation of ZIP parameters using voltage and power measurements. Estimated ZIP coefficients are prone to error when voltage measurements don’t possess sufficient change leading to ill-conditioned estimation jacobian. In this work, a new algorithm is proposed with a) an adaptive window selection procedure, b) ZIP estimation using hybrid distance (HD-ZIP), Random forest and KNN-Regression, and c) considering data anomalies. The window used for ZIP-estimation adapts according to variation in data measured in terms of variation in identified features and voltage sensitivity. Two proposed methodologies for ZIP parameter estimation are a) a hybrid distance-based ZIP parameter estimation (HD-ZIP) and b) Random forest and KNN-Regression. HD-ZIP is a hybrid of unsupervised and supervised learning and performs intelligent least-squares regression by finding windows having similar numerical features. Random forest and KNN-Regression are used to estimate the load model parameters using a supervised approach to resolve issues with least-squares based parameter estimation. Two sets of scenarios are considered for parameter estimation. The first scenario is positive sensitivity, where power and voltage measurements have a positive correlation. The second scenario is negative sensitivity, where load composition changes lead to the negative correlation between power and voltage measurements. Parameter estimation for various load compositions have been validated using IEEE test systems modeled in the PSS/E and real time digital simulator (RTDS) as well as with the comparative analysis.