To address the challenge of achieving high accuracy in tidal level forecasting along the tidal reach of the Yongjiang River Estuary, this study introduces a hybrid model (VMD-LSTM hybrid model) that integrates classical harmonic analysis (T_TIDE), variational mode decomposition (VMD), and long short-term memory network (LSTM). The VMD-LSTM hybrid model utilizes the T_TIDE package to obtain the hourly water level data of the estuary. The tide-simulated water levels are calculated, and the corresponding residual water levels are derived by subtracting tidal levels from the measured data at each station. The VMD model is employed to decompose the residual water levels into 13 Intrinsic Mode Functions (IMFs), specifically IMF₀ through IMF₁₂, which correspond to the D0 to D12 tidal species in sequence. LSTM-based regression is performed on each IMF component and tide level of the residual water levels for step-by-step prediction over a forecast horizon ranging from 12 to 48 hours. The sum of the predicted values of each IMF component and tide level is the predicted value of the estuarine water level. The results showed that: the VMD model can completely separate the D0 to D12 tidal constituent fluctuations in the residual water level for the Yongjiang River Estuary; the root mean square error (RMSE) of the VMD-LSTM hybrid model for short-term water level forecasting at 12 hours, 24 hours, 36 hours, and 48 hours was reduced by at most 0.15, 0.13, 0.16, and 0.16 m, respectively, compared to the LSTM model. In addition, the hybrid model of VMD-LSTM demonstrates the most significant capability for error correction in the D0 and D2 tidal bands. Compared to the LSTM model, this approach can reduce the spectral peak prediction errors of these tidal bands by up to 0.05, 0.04 m·d0.5, respectively.