Aiming to improve the accuracy of regional groundwater depth prediction, a CEEMDAN-hybrid algorithm-LSTM prediction model was proposed. Based on the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method, the groundwater depth data from 15 farms under the jurisdiction of Jiansanjiang Branch office were decomposed into five modal components, effectively reducing the complexity of the input data. Meanwhile, a hybrid optimization algorithm combining the red fox optimization (RFO) algorithm and the whale optimization algorithm (WOA) was employed to optimize key parameters of the long short-term memory (LSTM) neural network model, including time step, number of hidden units, batch size, and learning rate, thereby further enhancing the model's prediction accuracy. Monthly precipitation and paddy field irrigation well volume were used as input factors for the LSTM model to separately predict the five modal components, and the final groundwater depth prediction was obtained by summing the predicted values of each component. The results showed that compared with the back propagation (BP) neural network model and the recurrent neural network (RNN) model, the CEEMDAN-hybrid algorithm-LSTM model reduced the root mean square error (RMSE) by more than 43%, and increased the coefficient of determination R2 and Nash-Sutcliffe efficiency coefficient (NSE) by more than 18%. Prediction results indicated that from 2023 to 2027, the overall groundwater depth in the Jiansanjiang Branch office area would vary by up to 6.22%, with southern farms generally having greater groundwater depths than northern farms.