Groundwater level forecasting in response to climate change scenarios in southwestern Saskatchewan using wavelet decomposition and artificial neural networks

Global warming has intensified extreme climate events, including prolonged droughts, altering precipitation patterns and threatening groundwater resources. Additional stresses from economic development, population growth, and land use changes exacerbate groundwater depletion, often exceeding recharge rates. This study develops a hybrid artificial neural networks (ANN) and wavelet decomposition (WA) model to forecast long-term groundwater levels (GWLs) until 2100 under future climate scenarios in southwestern Saskatchewan to inform sustainable groundwater management strategies in a data-scarce region with a complex disconnected aquifer system. Monthly gridded precipitation and temperature data were combined with monthly GWLs from three wells in two aquifers. Three machine learning ANN models were applied and evaluated to forecast GWLs: i) nonlinear autoregressive network with exogenous input (NARX), ii) nonlinear autoregressive network (NAR), and iii) nonlinear input-output network (NIO). WA was integrated with NIO and NAR for signal denoising. Moreover, two base models were applied to each well: i) linear regression (LR), and ii) autoregressive integrated moving average (ARIMA) to quantify WA-ANN added value. Three learning algorithms, Levenberg-Marquardt (LM), Bayesian Regularization (BR), and scaled conjugate gradient (SCG), trained the models with varying neurons and delay times. Results show that NARX trained with BR produce the most accurate predictions for all wells. The applicability of WA-NIO and WA-NAR trained with LM in data-sparse settings remains largely exploratory with potential for improvement. GWLs are least impacted under SSP1-2.6, moderately affected under SSP2-4.5, and severely impacted under SSP5-8.5. These findings support decision-making through informing aquifer sustainability management plans under changing climate conditions.