System identification techniques and modeling for nonintrusive load diagnostics

Abstract

This thesis addresses the requirements of a system that can detect on/off transients and identify physical parameters of loads connected to a power distribution network. The thesis emphasizes three areas; a transient classifier that recognizes load transients using a pattern matching scheme, parameter estimation techniques suited for use with this classifier, and case studies of modeling and identification motivated by diagnostics and performance monitoring. Together, these areas support applications that can extract detailed load information from centralized, easily accessible parts of a distribution network. A new approach and implementation of pattern-based nonintrusive transient classification is presented. The classifier is nonintrusive in the sense that it uses aggregated measurements at a central location and does not require instrumentation of individual loads. The classifier implementation includes a framework that integrates preprocessors for AC and DC environments, programs that present results, and load-specific parameter identification modules that are executed as their associated transients are classified. An obstacle for these parameter identification programs is that a good initial guess is needed for the iterative optimization routines typically used to find parameter estimates. Two approaches are given to overcome this problem for certain systems. The first extends conventional optimization methods to identify model parameters given a poor initial guess. The second approach treats the identification as a modeling problem and suggests ways to construct "inverse" models that map observations to parameter estimates without iteration. The techniques presented in the thesis are demonstrated with simulation data and in real world scenarios including a dormitory, an automobile, and an experimental building.