3 min read
HypoxAI

HypoxAI forecasts hypoxia (dangerously low dissolved oxygen) in Narragansett Bay using machine learning. The best model predicts dissolved oxygen two weeks out with 5.02% error. It won Best in Fair at the Rhode Island Science and Engineering Fair.

Why

When the bay is overly enriched with nutrients, algae bloom and deplete its dissolved oxygen, a condition known as hypoxia. In August 2003, a low-oxygen event killed over one million fish in Greenwich Bay.

The state monitors the bay with buoy sensors that take readings every 15 minutes: dissolved oxygen, temperature, salinity and more.

What I did

I fit two kinds of models to that sensor record. One kind predicts dissolved oxygen from the other indicators. The other forecasts dissolved oxygen 14 days out.

I compared tree-based models, a neural network and time-series models by their mean absolute percentage error (MAPE) on test data. Tree-based models were chosen because they are resilient to missing data; the neural network needed imputation.

Results

Forecasting dissolved oxygen 14 days out at the GSO Dock station (lower is better):

Model Error (MAPE)
Prophet 5.02%
XGBoost 7.52%
Exponential smoothing 7.86%
AutoARIMA 8.12%

Predicting dissolved oxygen from the other indicators:

Model Surface Bottom
Neural network (MLP) 3.49% 4.67%
HistGradientBoosting 4.76% 9.61%
ExtraTrees 4.85% 10.03%
XGBoost 5.00% 9.89%
Random forest 5.04% 10.03%

The real data is owned by RI DEM, the Narragansett Bay Commission and URI GSO, and isn’t mine to redistribute, so the public notebooks run on synthetic data. The full write-up is in the report (PDF).