The decrease in precipitation and excessive exploitation of groundwater resources have led to a significant decline in groundwater levels in many regions worldwide. One of the aquifers experiencing groundwater level depletion is the Nahavand aquifer, located in Hamadan Province, western Iran. To model the groundwater level, the Long Short-Term Memory (LSTM) model and the Group Method of Data Handling (GMDH) were employed. In order to enhance model accuracy, the Complete Ensemble Empirical Mode Decomposition (CEEMD) preprocessing technique was applied, resulting in the development of four models: LSTM, GMDH, CEEMD-LSTM, and CEEMD-GMDH. The results indicated that the GMDH model performed better than the LSTM model. Moreover, integrating these models with the CEEMD preprocessing technique improved their performance. Specifically, the coefficient of determination (R²) for the LSTM model increased from 0.867 to 0.950 in the CEEMD-LSTM model. Similarly, the R² value for the GMDH model improved from 0.885 to 0.945 in the CEEMD-GMDH model. Additionally, the analysis of Root Mean Square Error (RMSE) and Akaike Information Criterion (AIC) also demonstrated that the use of CEEMD preprocessing reduced model errors. Based on these findings, the CEEMD-LSTM hybrid model was identified as the most accurate and was therefore used to forecast groundwater levels for the next six months (the first half of the 2024–2025 water year). Overall, the CEEMD-LSTM hybrid model showed excellent performance in modeling groundwater levels in the Nahavand aquifer and has the potential to be applied to other aquifers as well.