Leveraging Spatiotemporal Redundancy for Sensor Data Imputation in Water Distribution Networks

Abstract

The rapid digital transformation of Water Distribution Networks (WDNs) has led to the collection of multi-sensor time series with high temporal and spatial resolution. However, missing data poses a significant challenge, undermining the usability and effectiveness of data-driven applications. Performing missing data imputation is essential to enhance data quality and support intelligent management. This study first reveals that WDN sensor data in tensor form inherently exhibit spatiotemporal redundancy across three dimensions: inter-sensor similarity, intra-day regularity, and daily recurrence. The redundancy can be algebraically characterized by the low-rank structure of WDN tensor data, providing a robust foundation for imputation. Based on these findings, a novel Low-rank Autoregressive Tensor Completion (LATC) approach is proposed to efficiently impute spatiotemporal WDN data. The LATC combines autoregressive regularization with standard low-rank tensor completion, effectively capturing both global redundancy and local correlation of multi-sensor WDN data. Finally, the LATC is validated on four real-world and simulated WDN data sets under eight different missing scenarios. Extensive experiments show that the LATC significantly outperforms state-of-the-art baseline methods, achieving accurate imputation even under severe corruption and complex missing patterns.