High frequency measurements of water demand at service connections are becoming more common as utilities install smart meter technology. The full range of use for these observations by water suppliers is only beginning to be realized. Potential applications include leak detection, improved demand forecasting, variable water pricing, and improved network operations. Here we develop an approach for the classification of demand patterns and apply this approach to a set of demands collected from smart meters within a single District Metered Area (DMA) of a municipal network. The goal of this work is to develop a robust procedure for classification of demands derived from smart metering and test this procedure on observational data. A fundamental aspect of many feature classification tools is representation of what are often complex and noisy data in a low dimensional feature space that captures the important attributes of the signal. In this work, we employ Gaussian Mixture Models (GMM's) as the basis set for representing demand patterns. GMM's provide a flexible approach to representing the temporal demand patterns with a relatively small number of parameters. The values of these parameters then serve as the feature set for multivariate classification. A data set of hourly demand readings spanning a six-month study period serve as the test case for analysis here. The smart meters record demands to both residential and commercial consumers. Results show that the GMM approach captures variations in the demand patterns between locations. To the first order, the identified patterns appear to be explained by the differences between residential and commercial consumers. The resulting groupings are compared to classifications made using total demand as the sole feature. The stability of the patterns over time is tested by independently clustering each month of data. © 2013 The Authors. Published by Elsevier Ltd.