Making waves: Enhancing stormwater quality models through memory bias dynamics.

Area of Science:

• Environmental science
• Hydrology
• Machine learning

Background:

• Process-based build-up/wash-off models are crucial for urban stormwater management.
• Model simplifications and environmental changes introduce biases, reducing decision reliability.

Purpose of the Study:

• To explore reframing model biases as information assets.
• To propose a model-agnostic machine learning error correction strategy to enhance predictive credibility.

Main Methods:

• Recasting model-observation errors as structured signals with "memory bias".
• Implementing a machine learning correction layer to learn dynamic error patterns.
• Applying the framework to a study catchment and testing across two catchments.

Main Results:

• The error correction strategy significantly improved pollutant Nash-Sutcliffe Efficiency (NSE) by ~20% with hydrologic inputs and ~66% with lagged errors.
• The proposed framework demonstrated superior accuracy and stability compared to traditional and pure machine-learning baselines.
• The approach enables dynamic output correction, moving beyond static model recalibration.

Conclusions:

• Systematic model biases can be transformed into learnable, usable information assets.
• This method enhances the long-term validity and reliability of stormwater models in dynamic environments.
• The approach creates responsive models capable of real-time adjustments to environmental fluctuations.