Groundwater level forecasting using empirical mode decomposition and wavelet-based long short-term memory (LSTM) neural networks
Amirhossein Nazari, Moein Jamshidi, Abbas Roozbahani, Behzad Golparvar
Abstract
Groundwater is a vital resource for multiple sectors, but over-extraction has led to significant declines in groundwater levels across many regions. Accurately forecasting groundwater levels is essential for effective planning and management. However, the presence of non-stationarity in groundwater time series, such as trends and fluctuations, can result in poor prediction performance. This study proposes a novel hybrid approach combining Long Short-Term Memory (LSTM) models with Empirical Mode Decomposition (EMD) and Wavelet Transform (WT) to address these challenges. Non-stationary data from three wells in San Bernardino County, California, collected over a five-year period (2017–2022), were used for training and testing the models. The time-series data were preprocessed using EMD and WT to break down complex patterns into simpler components, which were then fed into LSTM models to improve forecasting accuracy. Our results show that the EMD-LSTM model significantly outperforms both the Wavelet-LSTM and traditional Single LSTM models when the error is rooted in a trend factor. According to the Root Mean Squared Error (RMSE) index, The EMD-LSTM reduced forecasting errors by up to 19% and 78% for wells W0804 and W0904, respectively. In contrast, for the well 4905, WT and EMD were not able to increase LSTM accuracy when fluctuations happened randomly. These findings demonstrate that the EMD-LSTM model is a powerful tool for forecasting groundwater levels, especially in cases where non-stationarity is prevalent. This approach can be applied to enhance groundwater management strategies, helping decision-makers ensure sustainable water resource planning, particularly in regions facing unsustainable groundwater withdrawals. • Groundwater level data from multiple wells underwent ADF testing before forecasting. • EMD and Wavelet methods decomposed non-stationary GWL time-series. • EMD-LSTM outperformed WL-LSTM and Single-LSTM when trends caused non-stationarity. • EMD-LSTM reduced forecasting errors by 19% and 78% for W0804 and W0904.