Water temperature plays a pivotal role in shaping riverine ecosystems, exerting significant influence on a range of water quality parameters. However, accurately forecasting multi-temporal daily data remains challenging due to the non-stationary and nonlinear characteristics of hydrological time series. To address these challenges, the study proposes a novel deep learning framework that integrates Recursive Feature Elimination (RFE) with Multivariate Variational Mode Decomposition (MVMD), a multi-channel decomposition scheme that extracts meaningful sub-signals across multiple correlated variables. The decomposed features are processed using an Elman neural network integrated with a Bidirectional Gated Recurrent Unit (BIGRU) to capture both bidirectional and feedback-driven temporal dynamics. The model is applied to Fanno Creek and the McKenzie River in the western United States, demonstrating superior predictive performance under dynamically evolving hydrological conditions. RFE selected key input variables (discharge, pH, specific conductance, dissolved oxygen) from five years of data (2017-2021). The MVMD multi-channel scheme decomposes input lags into sub-sequences for each forecast horizon. The primary model (MVMD-ELMAN-BIGRU) was validated using elastic net (ELNET) regression and a Convolutional neural network coupled with BIGRU (CNN-BIGRU) as comparative machine learning (ML) models. Evaluation facilities include statistical indices, vulnerability assessments, and diagnostic visualizations. Additionally, for a reasonable evaluation of the models, a novel Multi-criteria decision-making (MCDM) method, namely the Multi-Objective Optimisation method based on Ratio Analysis (MOORA), was adopted to consolidate the metric performance across scenarios. The results indicated that MVMD-ELMAN-BIGRU, regarding the least value of MOORA (T+1:0.1096; T+3: 0.0096; T+7: 0.0478) for the Fanno Creek Rivers (T+1:0.00; T+3: 0.0296; T+7: 0.0932) for the McKenzie River, was superior to the MVMD-CNN-BIGRU and MVMD-ELNET models, respectively. This approach presents a promising solution for multi-temporal water temperature forecasting, which is crucial for effectively managing river ecosystems and water resources.
Jamei et al. (Tue,) studied this question.