Abstract

A literature survey is provided to summarize the existing approaches to building energy prediction. The survey examines both the theory behind each prediction model and practical issues such as data pre- and post-processing. It also points out the pros and cons of each prediction method. Artificial Neural Network (ANN) is identified in the survey as the most popular and effective way to predict building energy demand. The ANN theory is thoroughly reviewed in this thesis. In particular, the ANN prediction model is presented as a generalized nonlinear least squares method. In addition to discussing the architecture and training methods employed by an ANN, we also examine implementation issues such as how to select the input to an ANN through day-typing and how to remove data redundancy and reduce the dimension of the input vector space via principal component analysis (PCA). While most of the existing ANN models for building energy prediction are static in nature, this thesis focuses on developing dynamic ANN models that can evolve over time. The dynamic ANN models developed in this thesis are capable of adapting themselves to unexpected changes in the incoming data. When the dynamic model is combined with an automated data acquisition system, it can be used to provide real-time online building energy prediction. A number of experiments have been performed to test the effectiveness of the ANN models developed in this thesis. (Abstract shortened by UMI.)