When Physics Meets Machine Learning for Nowcasting of Hydrological Sensors Fault Detection

Urban flood monitoring requires timely and dependable decisions in fast-evolving, partially observed settings. Sensor-network faults can degrade awareness and delay response, with substantial human and economic consequences. We introduce a Nowcasting Physics-Informed (NPI) framework for detecting faults in streamflow sensors using a 100-min sliding window sampled every 2 min. The approach combines measured sensor signals with outputs from the Storm Water Management Model (SWMM), forms a fused feature set, and feeds it to a stacked long short-term memory (LSTM) model to estimate the probability of a fault at the end of each window. We assess the benefit of coupling physical-model information with data-driven learning by comparing non-physics baselines. Over five cross-validation folds, the physics-informed fusion improves F1 by 3.7 to 12.5 percentage points, raising performance from 0.75 for a data-only LSTM to 0.88 for the complete NPI model. The pipeline is causal, yields auditable predictions via explicit physical features, and generates binary alerts that operators can use directly. Overall, the method offers a practical blueprint for robust warning systems that maintain performance under unseen conditions.