Groundwater level prediction using modified recurrent neural network combined with meta-heuristic optimization algorithm

Abstract

Groundwater is an important resource for water supply; it fluctuates depending on various factors, and the prediction of groundwater level is very important for water resources. Among various models for predicting groundwater levels, deep learning models have been applied to various water resources fields. Recurrent neural network (RNN) is a deep learning model for sequential data, and optimizers of RNN are important operators for calculating weights. However, existing optimizers of RNN have disadvantages such as convergence of local optimum and absence of weights storage. To improve RNN, new optimizers that combine existing optimizers with a meta-heuristic optimization algorithm were applied to a modified recurrent neural network (MRNN). To verify the accuracy of the MRNN, the groundwater level in Icheon was predicted and compared with the prediction results of RNN. The average temperature, daily precipitation, relative humidity, duration of sunshine, ground temperature, water level of nearby stream, and soil wetness were used as input data for the groundwater level prediction. Correlation analysis and normalization were applied as data preprocessing methods. The accuracy of each model was compared according to the value of mean square error (MSE). Prediction accuracy of MRNN was improved by an average of 43.35 % compared to RNN.