Groundwater Level Forecasting with Artificial Neural Networks: A Comparison of LSTM, CNN and NARX

Abstract:

Recently, Deep Learning (DL) networks show great potential to deliver high-quality forecasts in subsurface-water applications. However, also shallow neural networks have proven their value in groundwater level forecasting during the last years. In this study we perform a comprehensive comparison of two DL networks, long short-term memory (LSTM) and convolutional neural network (CNN), with a shallow nonlinear autoregressive neural network with exogenous inputs (NARX). We base the analysis on a highly transferable forecasting approach, using weekly groundwater level data as well as only few meteorological input parameters, such as precipitation, temperature and relative humidity. The hyperparameters for all models were optimized with Bayesian optimization. The dataset we use comprises 17 groundwater hydrographs from the Upper Rhine Graben aquifer in central Europe, with lengths between 22 and 49 years of weekly data. Our results show that for sequence-to-value forecasting (one week ahead) all model types can achieve similar results, LSTMs and NARX show higher performance on a case-by-case basis, with LSTMs being more robust to network initialization effects. CNNs, however, show by far the fastest calculation speed, which might make them preferable for certain applications. We also illustrate that low forecasting performance for some time-series is probably due to insufficient input data, not necessarily because of an inadequate modelling approach. Furthermore, we study the influence of training data length on model performance and observe a noteworthy threshold of training data needed at about 10 years in the case of LSTMs and CNNs, below which the forecasting performance becomes significantly worse. We also explore sequence-to-sequence forecasting (3 months ahead) and find that all three models are well suited and are therefore of great potential for short- and mid-term forecasting applications, where no future input data is available for the forecasting period. NARX outperform both LSTMs and CNNs in terms of accuracy and can almost keep up with CNNs in terms of calculation speed, but also show higher initialization dependency, which, however, can be easily compensated by implementing a small forecasting ensemble.