Feasibility study of real-time virtual sensing for water quality parameters in river systems using synthetic data and deep learning models

Abstract

With water quality management crucial for environmental sustainability, multiple techniques for real-time monitoring and estimation of water quality parameters have been developed. However, certain data types, such as airborne images, are only accessible from major river systems and are often unavailable for small river systems. This study assessed the feasibility of virtual sensing combined with deep learning (DL) for the real-time measurement of key water quality parameters-total organic carbon (TOC), total nitrogen (TN), and total phosphorus (TP)-in a small river system. By utilizing data from nine sensor measurement indicators, multiple linear regression models were formulated for the parameters to generate synthetic TOC/TN/TP data. Subsequently, two DL models were implemented using actual and synthetic data, with their results compared. The results revealed that the best DL models achieved an error of less than 0.4 mg/L for all water quality parameters (0.3975 mg/L for TOC, 0.2285 mg/L for TN, and 0.0055 mg/L for TP) compared to the actual data. The study also revealed that actual data are advantageous for short-term accurate estimations, while synthetic data are more suitable for long-term overall predictions. The computation time for estimating the parameters was less than a minute, significantly shorter than on-site measurement. Overall, this study demonstrates that DL-based virtual sensing can be effectively utilized in data-limited river systems.