Application of U-Net for domestic water end-use categorization

Abstract

Policymakers have been keen on understanding domestic water-use patterns in recent years. This knowledge helps to formulate water conservation measures and evaluate the efficacy of such measures. Revealing water-use patterns of a household requires a system to categorize household-level water consumption data, collected by smart meters, into specific end-uses categories, such as showering, basin use, kitchen use and laundry use. Prior studies have suggested identifying individual water-use instances and then applying machine learning techniques to categorize each of these instances (event-based categorization). However, these methods have not been battle-proven to work in all scenarios due to the limited variety of the test datasets. Our data to be categorized are collected locally in Hong Kong, with a relatively low data resolution when compared to data used by previous studies. It is observed that existing methods encountered accuracy issues in our scenario. We, therefore, propose a method that outperforms existing methods without requiring costly high-resolution data collection. Our method applies machine learning techniques, specifically a U-Net model, to categorize the data. Instead of categorizing water-use instances, it takes a household's whole-day water-use profile as input and performs the categorization as a whole (accumulated-use categorization). This method provides more time-related information for the machine learning model to learn from, counteracting the drawbacks of low-resolution data. When the data logging interval is 5 min, our method achieved 84 % for the daily estimated accuracy metric. It outperforms existing methods that achieved 56.7 % and 83 % accuracy in their worst- and best-case scenarios respectively. This thesis has shown that water end-use categorization may not need to be performed per water-use instance. It could be beneficial to take a bigger picture as input, especially when the data resolution is low.