• Journal of Korea Water Resources Association
• /
• v.57 no.10
• /
• pp.673-686
• /
• 2024

Groundwater is an important water resource used for various purposes along with surface water. Jeju Island relies on groundwater for most of its water use, so predicting and managing groundwater volume is very important for sustainable use of groundwater. In this study, precipitation data from the Baekrokdam Climate Change Observatory was additionally used to accurately predict groundwater levels. We compared and analyzed the improvement in monthly groundwater level prediction performance of the ANN and LSTM models for two observation wells located in the mid-mountainous area of the Pyoseon watershed in Jeju Island. As a result, when Baekrokdam precipitation data was not used, the NSE values of the two artificial intelligence models were over 0.871, showing very high groundwater level prediction performance. The LSTM model showed relatively higher prediction performance at high and low groundwater levels than the ANN model. We found that the prediction performance decreases as the variation characteristics of the groundwater level become more complex. When Baekrokdam precipitation data was additionally used, the NSE values of the two artificial intelligence models were above 0.907, indicating improved prediction performance, and the NSE value was improved by up to 0.036. This means that when additional rainfall in the upstream area is used, the artificial intelligence model can more appropriately interpret the fluctuating characteristics of the groundwater level. In addition, the additional use of Baekrokdam precipitation data further helped improve groundwater level prediction for observation well, where groundwater level prediction is relatively difficult, and artificial intelligence models, which have relatively low groundwater level prediction performance. In particular, when Baekrokdam precipitation data was additionally used for a specific observation well, the groundwater level prediction performance of the ANN model was improved to a level comparable to that of the LSTM model. The methods and results of this study can be useful in future research using artificial intelligence models.