Combing High-Frequency Surrogate Measurements and Data-Driven Models to Early Warn Water Quality Anomalies

Abstract

It is critical for surface water management systems to provide early warnings of abrupt, large variations in water quality. Anomalous water quality levels may be due to a variety of factors, such as natural accidents, uncertain point sources, and the intentional injection of contaminants. Although water quality monitoring systems generate data continuously using automatic sensors, the full benefits of online monitoring data cannot be obtained without real-time analysis. Smart anomaly detection systems (ADSs) based on real-time monitoring data are in high demand.

Many key water quality indexes cannot be monitored at a high temporal resolution, and some may not be detected by a sensor in real time. Surrogate relationships can be used to estimate water quality concentrations at a much higher temporal resolution. Data-driven models can be used to describe these variations at a high resolution. In this study, a combined approach integrating wavelet-ANN model and high-frequency surrogate measurements is proposed as a method of water quality anomaly detection and warning provision. High-frequency time series of major water quality indexes (TN, TP, COD, etc.) were produced via a regression-based surrogate model. After wavelet decomposition and denoising, a low-frequency signal was imported into a back-propagation neural network for one-step prediction to identify the major features of water quality variations. The precisely trained site-specific wavelet-ANN outputs the time series of residual errors. A warning is triggered when the actual residual error exceeds a given threshold, i.e., baseline pattern, estimated based on long-term water quality variations.

A case study based on the monitoring program was conducted. The integrated approach successfully identified two anomaly events of TP variations at a 15-minute scale from high-frequency online sensors. A storm event and point source inputs likely accounted for these events. Compared to the ANN prediction method, the wavelet-ANN method was more sensitive to sudden water quality anomaly events and avoided the effects of false positive events in many cases. ROC tests based on two hypothetic scenarios yielded a detection accuracy of up to 0.98. Analyses of the performance at different stations and over different periods illustrated the stability of the proposed method. By combining monitoring instruments and surrogate measures, the presented approach can support timely anomaly identification and be applied to urban aquatic environments for watershed management.

We live in the age of big data, and intelligent and precise environmental management is essential. The data-driven method proposed can serve as a state-of-the-art alternative for safeguarding water quality, improving urban and aquatic environmental management, controlling non-point and point source pollution, and enhancing watershed management.

Keywords: surrogate measurements; data-driven models; water quality; anomalies