This dissertation develops a "neighborhood" based neural network model utilizing wavelet analysis and Self-organizing Map (SOM) to predict building baseline energy use. Wavelet analysis was used for feature extraction of the daily weather profiles. The resulting few significant wavelet coefficients represent not only average but also variation of the weather components. A SOM is used for clustering and projecting high-dimensional data into usually a one or two dimensional map to reveal the data structure which is not clear by visual inspection. In this study, neighborhoods that contain days with similar meteorological conditions are classified by a SOM using significant wavelet coefficients; a baseline model is then developed for each neighborhood. In each neighborhood, modeling is more robust without unnecessary compromises that occur in global predictor regression models. This method was applied to the Energy Predictor Shootout II dataset and compared with the winning entries for hourly energy use predictions. A comparison between the "neighborhood" based linear regression model and the change-point model for daily energy use prediction was also performed. We also studied the application of the non-parametric nearest neighborhood points approach in determining the uncertainty of energy use prediction. The uncertainty from "local" system behavior rather than from global statistical indices such as root mean square error and other measures is shown to be more realistic and credible than the statistical approaches currently used. In general, a baseline model developed by local system behavior is more reliable than a global baseline model. The "neighborhood" based neural network model was found to predict building baseline energy use more accurately and achieve more reliable estimation of energy savings as well as the associated uncertainties in energy savings from building retrofits.