A Novel Machine Learning Based Cluster-Then-Forecast Framework for Sensor Placement Optimization and Real-Time Water Level Prediction in Low Impact Development (LID) Stormwater System

Emerging real-time monitoring technologies have been implemented in stormwater system management in recent decades. However, barriers such as limited monitoring resources and transient reaction time for decision making still exist in practical applications. Nonstationary rainfall and urbanization can relocate the flooding hotspots. To bridge these gaps, a novel machine learning based cluster-then-forecast framework is proposed for sensor placement optimization and real-time water level prediction at both monitored and unmonitored nodes in stormwater systems with Low Impact Development. The developed customized clustering algorithm enables the pinpointing of representative centroid nodes for optimized sensor placement. Past 25 min water level data from monitored nodes is used by LSTM models to predict system-wide water levels with a 5 min lead in the case study, in which centroid node information significantly improves the forecast accuracy for all nodes in a spatially holistic manner. To our knowledge, this study is the first work to predict hydraulic information at unmonitored nodes using only monitored data in real time, addressing the challenge of limited sensor coverage in newly emerging flood hotspots.