Evaluating deep learning models for estimating groundwater level perturbation sources in areas with non-natural groundwater level regimes

Abstract

The study uses deep learning models such as artificial neural networks (ANN) to tackle applications within the scope of hydrogeology. Time series modelling is one of many such applications used in analysing historical groundwater heads over an area of interest. A perturbed source, i.e. the Haga service tunnel of the West Link commuter rail project is the case study used to examine the model. Reduced groundwater heads are observed from monitoring wells around the perturbed source when plotted against time around the service tunnel. This reveals a certain impact on the groundwater caused by underground constructions. In this thesis, a branch under ANN called RNN (Recurrent Neural Networks) using LSTM (Long- Short Term Memory) network is investigated for the case study and the model is completely data-driven and statistically evaluated. The standard evaluation error metrics used are Mean squared error, Mean absolute error, Coefficient of determination and Pearson correlation coefficient. The predictions made on the observed GWL are checked with the error metrics and the results evaluated the impacts caused by the perturbed source for different cases. The groundwater heads are forecasted for different time periods based on best case identified from sequence to sequence (seq2seq) predictions. From the forecast, the groundwater heads are evaluated with unseen data using the Pearson correlation coefficient. However, shorter forecasting periods showed promising accuracy compared to longer forecasting accuracy indicating the groundwater movements are complex in real world influenced by natural parameters such as precipitation rates, relative humidity, evapo-transpiration.