Threshold update for reservoir safety status classification and early warning based on multi-source data fusion of rainwater engineering monitoring

Abstract

To address the issue of false negatives or false positives in static thresholds under complex rainfall conditions, this study proposes a method for updating early warning thresholds for reservoir safety status classification based on the fusion of multi-source data from rainfall monitoring. Based on a four-dimensional indicator system comprising seepage, deformation, hydrometeorological data, and stress, a comprehensive feature sequence is constructed using signal-to-noise ratio (SNR) weighting and sliding window smoothing. The safety status is classified into four levels—normal, watch, alert, and alarm—based on quantiles and rates of change. Bayesian online change-point detection is used to identify statistical change points for delineating stable phases. Within each phase, upper and lower threshold curves are fitted using extreme value quantile regression, enabling adaptive adjustment of thresholds in response to rainfall conditions. Experimental results show that the method achieves an average F1-score of 0.94 and exhibits zero missed and false alarms during typical flood events, validating its effectiveness.