A hybrid DWTNN–MARS framework for monthly lake water level forecasting in the lake Volta basin under data-scarce conditions

Abstract

Accurate forecasting of Lake Water Level (LWL) fluctuations is essential for sustainable reservoir management, particularly in data-scarce tropical environments. This study develops and evaluates a hybrid Discrete Wavelet Transform Neural Network–Multivariate Adaptive Regression Splines (DWTNN–MARS) framework for monthly LWL prediction in the Lake Volta Basin, Ghana, using 28 years (1992–2020) of satellite altimetry data. The proposed approach integrates multi-resolution wavelet decomposition, nonlinear neural network pre-processing, and adaptive spline regression with generalized cross-validation-based pruning to effectively capture complex, non-stationary hydrological dynamics while controlling model complexity. A time-series-based validation strategy was adopted to prevent data leakage and ensure realistic predictive assessment. Model performance was evaluated using Prediction Correction Index (PCI), Arithmetic Mean Absolute Error (AMAE), Arithmetic Mean Square Error (AMSE), Arithmetic Mean Absolute Percentage Error (AMAPE), and Arithmetic Standard Deviation (ASD), and benchmarked against a DWTNN–ARIMA hybrid model. Results show that the DWTNN–MARS model achieved a PCI of 0.0215 m and ASD of 0.0003 m, indicating strong agreement between observed and predicted values, while the DWTNN–ARIMA model exhibited higher prediction bias (PCI = 0.2152 m) and greater residual dispersion (ASD = 0.0420 m). These findings demonstrate that the structured decomposition-regression architecture enhances predictive accuracy and improves the representation of seasonal variability and long-term storage dynamics. The study highlights the robustness and applicability of the proposed framework for large tropical reservoirs under limited hydro-meteorological data conditions and contributes to advancing data-driven hydrological forecasting methodologies.