Applying Transfer Learning to Improve the Performance of Deep Learning-based Groundwater Level Prediction Model with Insufficient Training Data

초록

Predicting groundwater levels with data-driven models like artificial neural networks typically requires a substantial amount of data. However, when groundwater monitoring wells are newly developed or when a significant portion of the data is invalid (for example, due to missing values or outliers), acquiring an adequate dataset for training prediction models becomes challenging, leading to diminished prediction accuracy. This study proposes a method based on transfer learning to address the issue of insufficient training data. The Gated Recurrent Unit (GRU) was used as the primary data-driven model for predictions. A GRU-based pretrained network for the transfer learning process was developed using groundwater level and corresponding rainfall data collected from 89 monitoring stations nationwide. Subsequently, this pretrained network was fine-tuned using a small amount of training data obtained from the target monitoring well to develop the final prediction model. To verify the effectiveness of the transfer learning algorithm, two different groundwater level prediction models were evaluated: 1) a GRU-based model trained with insufficient data from the target well, and 2) a GRU-based model utilizing the transfer learning algorithm. Comparative verification was conducted with groundwater level data obtained from wells at two different locations, where the model using the transfer learning algorithm demonstrated superior performance compared to the other. This study confirms that the transfer learning algorithm can significantly enhance the performance of groundwater level prediction models, irrespective of the amount of available training data.

키워드