A multi-step water quality prediction model based on the Savitzky-Golay filter and Transformer optimized network

Area of Science:

• Environmental Science
• Data Science
• Hydrology

Background:

• Water quality prediction is crucial for sustainable water resource management.
• Complex water environments and noisy data hinder accurate long-term trend analysis and multi-step prediction.
• Existing loss functions can cause prediction lag, impacting forecast accuracy.

Purpose of the Study:

• To develop an effective multi-step water quality prediction model for watersheds.
• To address challenges posed by noisy data and prediction lag in water quality forecasting.
• To improve the accuracy and generalization ability of water quality prediction models.

Main Methods:

• Utilized the Savitzky-Golay (SG) filter to smooth noise and enhance data trend analysis.
• Employed a Transformer network with a sequence-to-sequence framework, position encoding, and self-attention for multi-step prediction.
• Introduced the DIstortion Loss including shApe and TimE (DILATE) loss function to mitigate prediction lag.

Main Results:

• The proposed model demonstrated superior accuracy in multi-step water quality prediction across four monitoring stations.
• Predictions exhibited correct shape and temporal positioning, outperforming a benchmark model.
• The model effectively captured serial correlations in water quality data.

Conclusions:

• The combined SG filter and Transformer model offers a robust solution for complex water quality prediction tasks.
• The DILATE loss function significantly improves prediction accuracy by addressing shape and time errors.
• This approach provides a valuable decision-making tool for watershed water quality and pollutant control.