A comparison of nonlinear regression and neural network models for ground-level ozone forecasting

Area of Science:

• Environmental science
• Atmospheric chemistry
• Computational modeling

Background:

• Ground-level ozone (O3) is a significant air pollutant with adverse health effects.
• Accurate forecasting of O3 concentrations is crucial for public health advisories and environmental management.
• Predictive modeling plays a key role in understanding and mitigating O3 pollution.

Purpose of the Study:

• To compare the performance of a nonlinear regression (NLR) model and a neural network (NN) model for O3 forecasting.
• To evaluate model accuracy in both forecast (using predicted weather) and hindcast (using observed weather) modes.
• To assess the models' ability to predict high O3 concentration events.

Main Methods:

• Developed and applied a hybrid nonlinear regression (NLR) model and a neural network (NN) model.
• Utilized meteorological data for two O3 seasons (1998 and 1999) in Louisville, KY.
• Compared model predictions in forecast mode (using forecasted meteorological data) and hindcast mode (using observed meteorological data).

Main Results:

• In forecast mode, NLR and NN models exhibited similar performance, with mean absolute errors of 12.5 ppb and 12.3 ppb, respectively.
• Both models achieved a 42% detection rate for O3 threshold exceedances (120 ppb) in forecast mode.
• In hindcast mode, the NLR model outperformed the NN model, showing a lower mean absolute error (11.1 ppb vs. 12.9 ppb) and a higher detection rate (92% vs. 75%).

Conclusions:

• Both NLR and NN models are viable tools for forecasting ground-level O3 concentrations.
• The NLR model demonstrated superior performance in hindcasting O3 levels and detecting high-concentration events.
• Model selection may depend on whether the application prioritizes forecasting or hindcasting accuracy for O3 prediction.