Deep Learning‐Based Ensemble Forecasts and Predictability Assessments for Surface Ozone Pollution

The impacts of weather forecast uncertainties have not been quantified in current air quality forecasting systems. To address this, we developed an efficient 2‐D convolutional neural network‐surface ozone ensemble forecast (2DCNN‐SOEF) system using 2‐D convolutional neural network and weather ensemble forecasts, and we applied the system to 216‐hr ozone forecasts in Shenzhen, China. The 2DCNN‐SOEF demonstrated comparable performance to current operating forecast systems and met the air quality level forecast accuracies required by the Chinese authorities up to 144‐hr lead time. Uncertainties in weather forecasts contributed 38%–54% of the ozone forecast errors at 24‐hr lead time and beyond. The 2DCNN‐SOEF enabled an “ozone exceedance probability” metric, which better represented the risks of air pollution given the range of possible weather outcomes. Our ensemble forecast framework can be extended to operationally forecast other meteorology‐dependent environmental risks globally, making it a valuable tool for environmental management. Plain Language Summary Weather forecasts are intrinsically uncertain, but the impacts of that uncertainty on air quality forecasts are not explicitly quantified in current air quality forecast systems. We proposed here a surface ozone ensemble forecast system, analogous to modern weather ensemble forecast systems, to represent the probability distribution of forecasted surface ozone concentrations given 30–50 possible future weather outcomes. The computation costs of this surface ozone ensemble forecast system were greatly reduced using deep learning techniques that emphasized the spatial patterns of weather. We showed that the surface ozone ensemble forecast system's accuracy met the Chinese operational requirements. However, half of the ozone forecast error was due to weather forecast uncertainties, which cannot be completely eliminated even with perfect pollutant emission estimates and chemistry models. This weather‐induced innate uncertainty in air quality forecasts should be considered for effective air quality management. Key Points We built a deep‐learning surface ozone ensemble forecast system to quantify pollution risks given the range of possible weather outcomes Deep‐learning models accentuating the spatial patterns of weather effectively represented the ozone‐meteorology relationship Weather forecast uncertainties contributed 38%–54% of the ozone forecast errors at 24‐hr lead time in Shenzhen