Data-Driven Monitoring Strategies for Groundwater Quality Protection Through Time Series Clustering of Groundwater Pollution Data.

Over the past few decades, groundwater quality monitoring networks have grown, and considerable time series of quality data are currently available. Hence, a research effort is required to investigate novel methods for exploiting groundwater quality time series. In this work, time series of groundwater pollution data are exploited with the aim of developing data-driven monitoring strategies, and early warning systems to support groundwater resource management for drinking water purposes. The research area is Brescia municipality (N Italy): an urban area with a complex hydrogeological context and numerous superimposing historical pollution sources. The dataset, provided by A2A Ciclo Idrico Spa, consisted in concentrations of Tetrachloroethylene (PCE), Trichloroethylene (TCE) and Cr(VI) on raw water before any potabilization treatment, from 2009 to 2020 for 68 wells and 16 springs. Time series exploratory analysis and time series cluster analysis were performed on groundwater pollution data, to identify groups of wells with homogeneous responses to anthropogenic inputs and highlight critical situations. The exploratory analysis was performed through Mann-Kendall test and Sen's slope estimator for the trend identification and quantification. A multivariate time-series clustering was performed on PCE and TCE, with the Dynamic Time Warping method. The clustering revealed 3 clusters linked to diffuse background contamination and 7 clusters linked to discrete hotspots, distinguished by distinct time profiles. Similarly, Cr(VI) data were subjected to a univariate time series cluster analysis, which revealed 3 background clusters and 7 hotspots, including 4 singletons. Based on the clustering outputs, data-driven monitoring strategies and early warning systems were designed for each group of wells. For the clusters associated with diffuse background contaminations and those with constant trends, the 95th percentile was used to calculate trigger levels, which represent future threshold values for early warnings. Specific monitoring procedures were suggested for the clusters with pluriannual patterns that were either oscillatory or monotonous based on the direction of the trends. The outcomes of the work demonstrate how a detailed time series analysis can support the implementation of data-driven monitoring strategies and early warnings for more efficient, site-specific monitoring networks, able to avoid redundant analyses while focusing on relevant trends.