Calibration of stormwater quality regression models: a random process?

Area of Science:

• Environmental science
• Hydrology
• Water quality modeling

Background:

• Regression models are crucial for estimating pollutant event mean concentrations (EMC) in urban wet weather discharges.
• Understanding the calibration sensitivity of these models is vital for accurate water quality assessments.

Purpose of the Study:

• To investigate the sensitivity of EMC regression models to the size and content of calibration datasets.
• To analyze how model results change when re-calibrated with evolving and expanding datasets over time.

Main Methods:

• Developed and analyzed four Total Suspended Solids (TSS) EMC regression models (two log-linear, two linear) using data from 64 rain events.
• Employed iterative re-weighted least squares for calibration, comparing it with ordinary least squares.
• Investigated three calibration strategies: chronological order and random sampling.

Main Results:

• Iterative re-weighted least squares yielded more robust and less sensitive model calibration compared to ordinary least squares.
• The best-performing nonlinear model demonstrated high sensitivity to both the size and composition of the calibration dataset.
• Model predictions varied significantly when re-calibrated with updated or expanded datasets.

Conclusions:

• The accuracy and reliability of EMC regression models are significantly influenced by the characteristics of the calibration data.
• Careful consideration of dataset size and content is essential for robust wet weather pollution modeling in urban catchments.
• Dynamic recalibration strategies are necessary to maintain model accuracy as new data becomes available.