Groundwater level forecasting with artificial neural networks: a comparison of long short-term memory (LSTM), convolutional neural networks (CNNs), and non-linear autoregressive networks with exogenous input (NARX)

Abstract. It is now well established to use shallow artificial neural networks (ANNs) obtain accurate and reliable groundwater level forecasts, which are an important tool for sustainable management. However, we observe increasing shift from conventional ANNs state-of-the-art deep-learning (DL) techniques, but a direct comparison of the performance often lacking. Although they have already clearly proven their suitability, recurrent frequently seem be excluded study design due euphoria about new DL techniques its successes in various disciplines. Therefore, aim provide overview on predictive ability terms levels ANNs, namely non-linear autoregressive with exogenous input (NARX) popular such as long short-term memory (LSTM) convolutional (CNNs). We compare both sequence-to-value (seq2val) sequence-to-sequence (seq2seq) forecasting 4-year period while using only few, widely available easy measure meteorological parameters, makes our approach applicable. Further, also investigate data dependency time series length different ANN architectures. For seq2val NARX models average perform best; however, CNNs much faster slightly worse accuracy. seq2seq mostly outperform even almost reach speed CNNs. least robust against initialization effects, nevertheless can handled easily ensemble forecasting. showed that networks, NARX, should not neglected especially when small amounts training available, where LSTMs CNNs; might substantially better larger dataset, really demonstrate strengths, rarely domain though.