Abstract: Most of the previous studies on the combined prediction of dam seepage pressure are based on a single pressure measurement point for modeling, ignoring the spatiotemporal correlation of multiple measurement points and using a linear combination strategy which suffers problems such as difficulty in capturing nonlinear features between sub-models. This paper constructs a combinatorial deep learning prediction model for dam seepage pressure, considering spatiotemporal correlation. First, the K-nearest neighbor (KNN) is used to optimize the local density function of the density peaks clustering (DPC) algorithm, so as to extract spatiotemporal correlation features from a seepage pressure time series and to achieve adaptive clustering. Then, for the time series, on the basis of its multi-scale refinement by wavelet decomposition (WD), the wavelet neural network (WNN) is used to capture its high-frequency details and construct a highly nonlinear mapping model based on the bidirectional long short-term memory (BiLSTM) for its low-frequency trend characteristics, spatiotemporal characteristics, and external environmental impact factors. Finally, the prediction results of high- and low-frequency feature sequences are combined nonlinearly based on the long short-term memory network (LSTM) to capture the nonlinear characteristics between sub-models. An engineering case analysis shows that our new model raises the prediction accuracy by 75.7% and 41.4%, respectively, compared with the single point prediction model without considering spatiotemporal correlation and the spatiotemporal prediction model using linear combination strategy. This validates its applicability and efficacy as a new approach for dam seepage safety monitoring.

Key words: dam seepage pressure prediction, spatiotemporal correlation, nonlinear combinatorial model, deep learning, density peak clustering