System design and dynamic signature identification for intelligent energy management in residential buildings.

Increasing energy demand from residential buildings and evolving utility pricing policy to regulate energy use during peak times require a new paradigm for energy management in residential buildings. As a prototype for intelligent energy management systems of resi- dential buildings, DREAM (Demand Responsive Electrical Appliance Manager), based on a wireless sensor network, was developed. This autonomous system consisting of wireless sensors and actuators, a graphical user interface, and a main control reduces peak electrical demand and ultimately optimizes energy management by identifying house dynamic signa- ture as well as occupant thermal preference and patterns. In summer 2007, functionality and overall performance were evaluated with two field tests and showed promise for the DREAM system.Due to significance of the house dynamic signature learning in an intelligent energy management system, three approaches were studied. Despite the simplicity of the model and success in identifying thermal characteristics of a house, the 1st order differential equation method, which considered thermal influences of five heat sources, showed limitations in representing actual temperature behavior delicately. The tabular method was suggested to capture house nonlinear behavior by learning temperature change rate with respect to different events and periods. The prediction using the tabular method followed the actual measured temperature within a tolerable error range, except for a relatively long heater- on event. The last method, the ARX model fitting method, provided the best prediction result, but the performance was considerably influenced by the choice of sample data for parameter learning.The multiple-model switching algorithm was proposed to minimize performance incon- sistency in the ARX model fitting method. Instead of sticking to one model, it allows several candidates whose parameters are calculated from seven consecutive days, and se- lects one (multiple-model hard switching [MMHS]) or fuses all (multiple-model soft switch- ing [MMSS]). Depending on the criterion to select or weight a candidate, the algorithm is divided into proximity-based model switching and applicability-based model switching. Overall, the MMSS showed better performance than the MMHS and, most of all, the applicability-based MMSS algorithm dramatically improved the prediction quality when anomalies in data were properly filtered.All algorithms in this study were evaluated with the real data that were collected from more than 20 occupied houses in Northern California, Minnesota, and South Australia