Deep learning models for multi-step prediction of water levels incorporating meteorological variables and historical data

Abstract

Precise multi-step water level predictions are crucial for managing water resources and mitigating the effects of extreme weather. This study introduces a novel approach by integrating Variational Mode Decomposition (VMD), Whale Optimization Algorithm (WOA), and Long Short-Term Memory (LSTM) to forecast variations in water levels, employing both endogenous and exogenous environmental variables. Furthermore, this research proposes two additional fusion algorithms, each possessing unique potential for enhancement: Multivariate Long Short-Term Memory (MLSTM) and an advancement in the Residual Sequence (RESID). The predictive accuracy of these diverse algorithms is assessed using data from the water levels in Jinan Baotu Spring, China. The findings indicate that the VMD-WOA-LSTM model presents the most robust results for both long-term and short-term predictions. For multi-step, ultra-short-term forecasts, VMD-WOA-MLSTM proves to be a pragmatic algorithm. However, the refined algorithm that incorporates RESID does not significantly improve and, indeed, may diminish prediction accuracy. Conclusively, the VMD-WOA-LSTM, exemplifying a data-driven predictive algorithm, boasts high accuracy and demonstrates versatility in water level forecasting across various scenarios.