Saltwater intrusion early warning in Pearl river Delta based on the temporal clustering method

Forecasting of saltwater intrusion in estuaries is challenging due to its nonlinear and nonstationary features. The efficiency of most existing regressive prediction models decreases exponentially with increasing lead steps. To meet the requirements of accuracy and long lead time, a real-time saltwater intrusion early warning framework based on timeseries clustering was proposed in this study. In the example analysis of the Pearl River Delta, four risk levels of saltwater intrusion were defined based on the salinity duration of a string of salinity gauges along the Modaomen channel. A comprehensive 24-h-forehead prognostic value of saltwater intrusion risk was obtained by clustering previous 48-h observations. Results indicated that the latest supervised clustering model, Cluster bAsed iMportancE Learning fOr Time-series (CAMELOT), achieved better predictive performance than traditional unsupervised clustering models and machine learning classifiers. Then, the temporal diversities of various environmental components, including antecedent chlorinity, upstream discharge, tidal level and wind vector, were investigated in each identified cluster. Notably, the variation of saltwater intrusion length was strongly associated with tidal cycles and upstream discharge. The maximum length of saltwater intrusion mainly occurred during the transition between neap and spring (spring and neap) tides, while the minimum length of saltwater intrusion occurred during spring tides. An average river discharge of 2100 m3/s at Makou station and 600 m3/s at Sanshui station contributed to the cluster with the highest propensity of the longest saltwater intrusion above Nanzhen station.

Automated summary

This study proposes a real-time saltwater intrusion early warning framework based on timeseries clustering to improve the accuracy and lead time of predictions. By clustering previous observations, a comprehensive 24-h forecast of saltwater intrusion risk is obtained. The latest supervised clustering model, CAMELOT, outperforms traditional unsupervised clustering models and machine learning classifiers. The analysis reveals that the variation in saltwater intrusion length is strongly associated with tidal cycles and upstream discharge.