A deep learning approach driven by raw monitoring data for earth/rockfill dam seepage prediction and safety assessment

Based on the measured seepage data of earth/rockfill dams, it is of great significance for dam safety assessment to establish an accurate seepage prediction model through reasonable and effective mathematical methods. This paper presents a deep learning framework for dam seepage prediction and safety assessment, including ensemble empirical modal decomposition (EEMD) decomposition of seepage monitoring data, isolated forest and local outlier factors (iForestLOF)-based outlier detection and rejection, bidirectional long and short-term memory (BiLSTM) neural network prediction model training, sparrow search algorithm (SSA)-based prediction model hyperparameter optimization, seepage prediction for earth/rockfill dams. Using seepage monitoring data from a gravel-shelled clay core dam in Wenzhou City, Zhejiang Province, China, for modeling and validation, the case study results show that the EEMD-SSA-BiLSTM seepage prediction model achieves a significant enhancement in accuracy, and especially exhibits higher efficiency in revealing the implicit information within the nonlinear data. The EEMD-iForestLOF outlier detection method is able to combine the local outlier degree with the global data anomaly information when determining the outlier points, which is very effective in improving the accuracy of the detection. The monitoring indicators formulated based on the proposed prediction model provide the safety boundary values that can be achieved when the dam is subjected to loads or effects, and the dam is in good operating condition and there is no potential risk of seepage failure for the time being.