Author(s): Danillo Lange; Marc Ribalta; Lluis Echeverria; Joshua Pocock

Linked Author(s):

Keywords: Water consumption profiling; Autoencoders; Time-series clustering; K-means; Dynamic time warping

Abstract: Nowadays, water utilities face the rising challenge of ensuring water availability amidst a rapidly growing society and a shifting climate. Our research aims to develop a household clustering solution based on water consumption behaviour in Southwest England, to enable utilities to identify different profiles and enhance customized control of household consumption, resulting in improved resource management. The solution is composed of three modules. The first one is based on a K-Means clustering model, designed to group domestic water use behaviours. This module uses the Dynamic Time Wrapping algorithm as a similarity mechanism to process the high-resolution water meter data. In parallel, the second module processes the market segmentation data through an Autoencoder, a specific Neural Network architecture used to reduce the high dimensionality of such data to a low dimension dataset by extracting its latent encoded space. Finally, to assemble the final household water use profiles, a blending K-Means algorithm is used to merge previous modules outputs, based on the Euclidean distance. The solution provides insightful information to water companies to better understand consumer behaviour, habits, and routines.

DOI: https://doi.org/10.1088/1755-1315/1136/1/012005

Year: 2022