A cyber-physical testing facility for edge-based automatic classification of residential water end uses

Water demand-side management strategies are increasingly recognized as key for urban water conservation, complementing supply-side operations. Recent studies demonstrate that consumption-based feedback can effectively encourage water conservation behaviors. Smart meters, supported by digital innovations like the Internet of Things (IoT) and advanced data analytics, have become central to enabling personalized feedback and reinforcing behavioral changes. These advancements highlight the need for Non-Intrusive Water Monitoring (NIWM) algorithms capable of estimating individual water end uses from aggregate household consumption recorded by single-point smart meters. Existing research offers heuristic and machine learning algorithms to address two primary tasks in NIWM: disaggregation of concurrent end uses and automatic classification of the resulting water end-use data. While many algorithms have been designed, calibrated, and validated using high-resolution temporal data—often synthetically generated or inaccessible due to closed datasets—reproducibility in a realistic environment remains a challenge. Furthermore, most algorithms are tested in virtual settings, overlooking real-world concerns related to data transmission, end users’ privacy, and the intrusiveness of centralized analyses by the water utility or a third party.

In this study, we present a novel cyber-physical testing facility for edge-based, real-time classification of residential water end uses. This facility replicates typical residential water use scenarios and employs machine learning algorithms for on-site, edge computing. Its physical components are modular and include a water tank, two circulation pumps, and piping and valves to simulate flow rate trajectories of various end-use categories. Water consumption is measured using a digital flow meter, with data processed by PyNIWM, an open-source Python framework for NIWM, operating in near real-time on a local computer. By integrating physics-based simulations of water use with edge computing, our test stand supports (i) benchmarking and reproducibility of NIWM algorithms in realistic conditions, (ii) privacy-compliant end-use classification and analysis, (iii) near real-time reporting of NIWM outcomes to users, and (iv) modularity to test various soft- and hardware setups. This approach bridges the gap between virtual testing and practical implementation, addressing key challenges in modern water management while advancing privacy-conscious, user-oriented solutions for smart water metering.