Abstract: To achieve a rapid and accurate simulation of water level variations in the Dongting Lake region, two deep learning models, i.e., the TRANS-LSTM and TRANS-LSTM-SE models which incorporate the Transformer architecture and the Squeeze-and-Excitation (SE) module, were established based on the Long Short-Term Memory (LSTM) neural network. The proposed models were trained and compared using 6h precipitation, water level, and flow discharge datasets from 301 rainfall stations, 6 water level stations, and 8 hydrological stations from 2010 to 2023, The results indicate that: ① The training and simulation times of the deep learning models are significantly less than those of hydrodynamic models when apply in the same basin; ② The TRANS-LSTM-SE model markedly outperforms both the TRANS-LSTM and the LSTM model, and the Nash-Sutcliffe Efficiency coefficients exceeds 0.996 during both the training and validation periods; ③ The multi-head attention mechanism of the Transformer architecture combined with the SE channel attention mechanism, could synergistically optimize the LSTM model and effectively describe the nonlinear spatiotemporal dependencies of water levels in the complex river-lake system, which provides a new approach for rapid water level simulation and forecasting in the Dongting Lake region.