A machine learning framework for predicting downstream water end-use events with upstream sensors. Issue 7 (7th June 2022)

• Record Type: Journal Article
• Authors: Kropp, Ian; Pouyan Nejadhashemi, A.; Julien, Ryan; Mitchell, Jade; Whelton, Andrew J.
• Is Part Of: Water Supply. Volume 22:Issue 7(2022)
• Page Start: 6427
• Page End: 6442
• Publication Date: 2022-06-07
• Subjects: bootstrap aggregated decision tree -- end-use categorization -- feature selection -- innovization -- premise plumbing -- water use
• DOI: 10.2166/ws.2022.226
• Languages: English
• ISSNs: 1606-9749

Abstract

Abstract: Understanding the end-use of water is essential to a plethora of critical research in premise plumbing. However, direct access to end-use data through physical sensors is prohibitively expensive for most researchers, building owners, operators, and practitioners. Therefore, machine learning models can alleviate these costs by predicting downstream end-use events (e.g., sink, shower, dishwasher, and washing machine) via an affordable subset of upstream sensors. Choosing which upstream sensors, as well as data preprocessing methods, are best for machine learning has historically been a manual process. This paper proposes a novel approach to systematically configure the machine platform automatically. The optima were determined through a Pareto analysis of the exhaustive combinations of upstream predictors and preprocessing methods. The model was trained and validated with real-world data obtained from a house that has been extensively monitored for over a year. Results from the analysis suggested that downstream events can be effectively predicted with minimum overfitting error for most categories, using as few as two to four upstream sensors. This study automatically implemented highly accurate machine learning models to predict downstream features within premise plumbing systems, significantly lowering the costs of researching residential plumbing best practices such as water conservation. HIGHLIGHTS: Physical end-use monitoring is prohibitively expensive for most practitioners. Machine learning is a viable method of lowering the cost of end-use categorization. Here we proposed a novel approach to systematically configure the machine platform automatically. Effective end-use prediction is possible with a small subset of upstream features. Graphical Abstract