Reliable anomaly early warning for hydropower station sump wells remains challenging due to the strong nonlinearity of water level dynamics and the limited adaptability of conventional fixed-threshold alarms. Here, we present a hybrid deep learning framework—termed CNN–Attention–LSTM–BO—that fuses multi-scale local feature extraction, adaptive temporal weighting, and sequential dependency modeling within a unified architecture, with all critical hyperparameters tuned via Bayesian optimization. A four-dimensional input representation is first constructed from the raw water level signal and its first- and second-order differences together with the drainage pump operating state, capturing both trend and transient information. One-dimensional convolutions at multiple kernel scales encode short-range fluctuation patterns, a Bahdanau-style temporal attention layer selectively amplifies informative time steps, and a stacked LSTM propagates long-horizon risk dependencies. At the decision stage, a dual dynamic thresholding scheme couples an improved 3σ criterion with kernel density estimation (KDE) to partition the smoothed risk score into three graded alert levels (normal/warning/critical), replacing the binary alarm paradigm. Experiments on the SWaT benchmark yield an average area under the ROC curve (AUC) of 0.9246, an average Accuracy of 0.8812, and a best single-well false alarm rate (FAR) of 3.21% (Well-4), with an average FAR of 8.97% across three wells, outperforming both traditional limit-value alarms and ablated variants lacking CNN or attention modules.
Lin et al. (Thu,) studied this question.