Application of the Two-Layer Regularized Gated Recurrent Unit (TLR-GRU) Model Enhanced by Sliding Window Features in...

Area of Science:

• Environmental Science
• Data Science
• Water Resource Management

Background:

• Traditional water quality monitoring faces limitations in real-time assessment due to temporal-spatial coverage and cost.
• Deep learning models struggle with the inherent complexity and noise in raw hydrological time series.

Purpose of the Study:

• To develop a robust framework for high-precision, real-time prediction of key water quality parameters.
• To address challenges posed by complexity and noise in hydrological time series data for improved water quality assessment.

Main Methods:

• Integration of sliding window feature enhancement, Principal Component Analysis (PCA) for dimensionality reduction, and a Two-Layer Regularized Gated Recurrent Unit (TLR-GRU) model.
• Utilized Sample Entropy (SampEn) to quantify time series regularity and assess the impact of noise reduction techniques.
• Collected and analyzed 4970 water quality records from a typical aquaculture-irrigated water body between 2020-2023.

Main Results:

• The proposed TLR-GRU framework significantly improved prediction accuracy for dissolved oxygen (DO) and total phosphorus (TP).
• Sliding window enhancement reduced time series complexity (SampEn) and noise, leading to better model performance.
• The enhanced dataset resulted in improved R2 values (e.g., DO from 0.82 to 0.93) and reduced RMSE (e.g., TP by 55.6%) compared to baseline models.

Conclusions:

• The developed framework offers a high-precision solution for real-time water quality prediction, outperforming existing state-of-the-art deep learning models.
• The approach is effective in managing complex and noisy hydrological data, supporting sustainable water resource management in aquaculture and beyond.
• Future research will focus on model optimization and integration of multi-source data for enhanced water quality monitoring.