Micro-component disaggregation of domestic water demands using machine learning

Growing population and rapid urbanisation have increased the pressure on water supply in many countries and regions, leading to severe water stress. The escalating impacts of water pollution and climate change are exacerbating the issue of water scarcity. Increasing water supply is generally associated with the development of water infrastructure and investment in alternative water sources. In contrast, addressing water conservation through effective household water management has become a critical issue to tackle water shortage crises. A range of tools for disaggregating water consumption into end-use levels have been developed in the recent years. Many of these studies involve the intrusive approaches, for example, installing sensors at individual water appliances in the household, to collect accurate water end-use data. However, due to cost and privacy concerns, non-intrusive methods are preferred in many cases. Developing non-intrusive method requires high-quality water end-use information, which enables the accurate separation between water end-use types. This study aims to develop and evaluate an effective machine learning-based method for disaggregating household water end-uses. A variety of machine learning algorithms and input features for disaggregation processes have been investigated, evaluated, and discussed in this research. The water end-use disaggregation requires high-resolution water consumption data (typically at intervals of less than a minute) to enable the extraction of detailed water consumption patterns. Devon, a coastal county in the United Kingdom, is the study area for this research. The water utility in the study area offers access to diverse water consumption datasets, collected via pulse interval loggers and automated meter reading (AMR) meters. The resolution of the water consumption data available in the area ranges from 30-minute intervals to recordings of each litre used. To support the disaggregation process, the research also presents the labelling process of each type of water end-use, along with the deployment of online household water consumption survey. The water end-use disaggregation can be broadly divided into two main components, segmenting water consumption data into end-use events and classifying these events with appropriate labels. The first part of the research 4 involves the process of data collection, cleaning and water end-use segmentation. The second part presents the development and evaluation of machine learning methods, including k-nearest neighbours, Random Forest and Artificial Neural Networks, using both physical features and time series patterns of water end-use events. The results from evaluating various machine learning models and input features combinations indicate that Random Forest is a promising method for further research, especially when combing physical features with time series patterns. The final part integrates the models developed for detecting each type of water end-use, as introduced in separate chapters, and applies the integrated model to a wider range of households within the study area. Based on the results from both detection of separate water end-uses and the application of the integrated model, several key findings from this research are summarised as follows: machine learning algorithms may perform differently depending on the specific type of water end-use, indicating that the selection of suitable algorithms is important for improving water end-use classification accuracy; the similarity of water end-use time series patterns proves to be an effective input feature, especially when labelled water end-use data is scarce. A small number of typical water end-use patterns can serve as reference samples for measuring similarity with unlabelled water end-uses. These reference samples function as cluster centroids, with similarity measured as the distance between each end-use and these central patterns. This approach can be applied when multiple patterns exist for a single water end-use type, allowing similarity to be used as an input feature for the training of water end-use disaggregation model alongside physical features, which may show limitations in capturing the water usage variability.<p></p>

Author dc:creator

• Xiaojie Zhou (21039959)

• water end-use
• machine learning
• household water consumption

Statement dc:rights

• All rights reserved
• Open Access after 2027-07-05

Identifier

10779/exe.30998605.v1

OAI identifier oai:identifier

oai:figshare.com:article/30998605