A unified deep learning framework for water quality prediction based on time-frequency feature extraction and data...

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Deep learning models are effective for predicting water quality due to their ability to map nonlinear relationships.
• Variations in model selection and construction lead to differing prediction accuracy and performance.
• A unified framework is needed to standardize and improve deep learning-based water quality prediction.

Purpose of the Study:

• To establish a unified deep learning framework for accurate water quality prediction.
• To integrate data processing, feature enhancement, and prediction modules for comprehensive analysis.
• To evaluate the framework's performance using real-world water quality data.

Main Methods:

• Data processing using wavelet transform to decompose complex hydrological and meteorological data into frequency domain signals.
• Feature enhancement via Informer Encoder for improved encoding of time series data and discovery of global time-dependent features.
• Water quality prediction using a stacked bidirectional long and short term memory network (SBiLSTM) to capture local correlations in feature sequences.

Main Results:

• The framework was applied to the Lijiang River, achieving maximum relative errors of 12.4% for dissolved oxygen (DO) and 20.7% for chemical oxygen demand (COD_Mn).
• The model demonstrated superior prediction accuracy compared to other methods, with low RMSE, MAE, and SMAPE values for DO and COD_Mn.
• Ablation tests validated the effectiveness and necessity of each module within the framework.

Conclusions:

• The developed unified deep learning framework offers a robust and accurate approach for water quality prediction.
• The integration of wavelet transform, Informer Encoder, and SBiLSTM significantly enhances prediction performance.
• This framework provides a valuable tool for environmental monitoring and water resource management.