A deep complementary learning framework for surface water temperature forecasting.

Area of Science:

• Environmental Science
• Hydrology
• Data Science

Background:

• River water temperature is a critical factor influencing aquatic ecosystems and water quality.
• Forecasting daily river water temperature data is challenging due to the non-stationary and nonlinear nature of hydrological time series.

Purpose of the Study:

• To develop and validate a novel deep learning framework for accurate multi-temporal daily river water temperature forecasting.
• To address the complexities of hydrological time series prediction using advanced decomposition and recurrent neural network techniques.

Main Methods:

• A deep learning framework integrating Recursive Feature Elimination (RFE) for feature selection and Multivariate Variational Mode Decomposition (MVMD) for signal decomposition.
• Utilized an Elman neural network with a Bidirectional Gated Recurrent Unit (BIGRU) to capture temporal dynamics.
• Applied the model to Fanno Creek and McKenzie River data, comparing it with MVMD-ELNET and MVMD-CNN-BIGRU using statistical indices and the Multi-Objective Optimisation method based on Ratio Analysis (MOORA).

Main Results:

• Recursive Feature Elimination identified key variables: discharge, pH, specific conductance, and dissolved oxygen.
• The proposed MVMD-ELMAN-BIGRU model demonstrated superior predictive performance compared to benchmark models across different forecast horizons (T+1, T+3, T+7).
• The MOORA method confirmed the superiority of MVMD-ELMAN-BIGRU, showing the least complex value for both Fanno Creek and McKenzie River datasets.

Conclusions:

• The novel MVMD-ELMAN-BIGRU framework offers a robust and accurate solution for multi-temporal water temperature forecasting in riverine environments.
• This approach enhances the ability to manage river ecosystems and water resources effectively by providing reliable temperature predictions.
• The study highlights the potential of integrating advanced signal decomposition and deep learning for complex hydrological time series analysis.