Reliable and Adaptive Probabilistic Forecasting for Event-Driven Water-Quality Time Series Using a Gated...

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Reliable water quality (WQ) forecasting is vital for environmental management.
• Modeling time-varying uncertainty (heteroscedasticity) is challenging, especially during disruptive events like storms.
• Standard probabilistic models often fail to accurately represent uncertainty during volatile periods.

Purpose of the Study:

• Introduce the Gated Hybrid-Mixture Density Network (GH-MDN) for adaptive uncertainty estimation in WQ forecasting.
• Develop a model that can modulate prediction interval width in response to event-precursor signals.
• Improve the trustworthiness and reliability of environmental forecasting systems.

Main Methods:

• Developed a novel Gated Hybrid-Mixture Density Network (GH-MDN) architecture.
• Incorporated a gating network to adaptively adjust prediction intervals.
• Evaluated the GH-MDN on synthetic and real-world WQ datasets using cross-validation.

Main Results:

• The GH-MDN demonstrated robust calibration and trustworthy adaptive coverage.
• The model successfully widened prediction intervals to capture extreme WQ events where benchmarks failed.
• Analysis revealed that aggregate metrics like CRPS can obscure over-confident behavior during rare events, highlighting the need for calibration-focused evaluation.

Conclusions:

• The GH-MDN offers a science-informed approach to modeling heteroscedasticity in WQ forecasting.
• Prioritizing reliable risk coverage over aggregate error minimization is crucial for trustworthy environmental systems.
• This work represents a significant advancement in developing more dependable environmental forecasting tools.