Conventional machine learning models often fail the task of identifying long-term dependencies and local importance when dealing with time-series data. To address this problem, we proposed a coupled model for seepage prediction of earth and rock dams based on bidirectional temporal convolutional network (BiTCN), attention mechanism (Attention), and least squares support vector machine (LSSVM). The model uses BiTCN to capture the long-term dependencies in the time-series data from both forward and backward directions, introduces the attention mechanism to help the model focus on the key local features related to the prediction, and inputs the deeply processed BiTCN-Attention features into the LSSVM model for better prediction results. To analyze the prediction effect, the model was used to process two different datasets. The results of case analysis show that compared with LSSVM, CNN-LSSVM and TCN-LSSVM, the evaluation indexes of BiTCN-Attention-LSSVM model are the best, and the proposed model shows higher accuracy and stability in the prediction of piezometer water level in the earth-rock dam. The combination of BiTCN and Attention can better extract the interdependence relationship in time-series data. Inputting the features extracted by BiTCN-Attention into LSSVM for prediction can achieve better prediction performance. After the data set is expanded, the learning ability of the model is effectively improved.