Groundwater level prediction using deep learning-based recurrent neural network and numerical modeling: a comparative study

The use of Artificial Neural Networks (ANNs) for predicting groundwater (GW) levels has shown promising results in previous studies. However, earlier research utilized diverse inputs such as air temperature, pumping rates, precipitation, population data, and other basin-related information as model inputs. This study proposes a Long Short-Term Memory Neural Network (LSTM-NN) model to predict GW levels using only historical GW level data as input, without relying on additional data basin-specific information.The proposed approach applies the LSTM-NN to predict GW levels or in the Qazvin aquifer in Iran. Historical monthly GW levels from six observation wells over a 30-year period were used, with 75% of the data allocated for training and 25% for testing and validation. Additionally, the numerical MODFLOW model was calibrated and validated for a five-year period under steady-state and transient conditions. The results demonstrate that the LSTM model can predict GW levels in the six observation wells with high accuracy up to 12 months ahead. The statistical coefficient of determination (R2) for the wells was recorded as 0.98143, 0.99094, 0.99263, 0.98492, 0.98698, and 0.70637, respectively. The root mean square error (RMSE) values for these wells were calculated as 0.48, 0.71, 0.32, 0.39, 1.24, and 0.55, respectively. In comparison, the RMSE for the transient state of the MODFLOW model was 1.24. The findings highlight the superior performance of the LSTM algorithm over the MODFLOW numerical model. These results emphasize the potential of machine learning (ML) algorithms in groundwater prediction and underline the critical importance of collecting high-quality, long-term GW level data to enhance the sustainable management of groundwater resources.