Applying Machine Learning Technique for Winter Ozone Forecasting in the Uintah Basin

Abstract:

We apply the recurrent neural network (RNN) with long short-term memory (LSTM) to predict ozone concentration out to 48 hours in advance during winter season in the Uintah Basin, Utah. Ozone concentration in the Basin frequently exceeded the national ambient air quality standard during past winter seasons and is mainly caused by emission from oil and gas productions and persistent atmospheric stable condition. The ability to predict ozone exceedances has crucial meaning in controlling ozone pollution. Ozone alert program for the Uintah Basin has been developed and operated by Bingham Research Center (BRC) which perform its educated guess of future ozone concentration basing on current condition of ozone and weather and on weather forecasts. The oil and gas operators have been relying on this ozone alert program to mitigate emissions of ozone precursors when high ozone is forecasted. BRC also has developed quantitative ozone forecast model basing on multi-variable regression of ozone and its diving factors (e.g., oil and gas production, meteorology) and uses numerical weather forecasts as inputs to predict discrete value of ozone. The model, however, has not been successful in forecasting ozone due to uncertainties in weather forecasts and the strong variation of ozone. The RNN model, after being properly trained, forecasts ozone based on current observations of ozone and weather conditions. Principle component analysis and decision tree techniques are performed to screen for most important input parameters of the RNN. Performance of the RNN model will be evaluated against the random forest regression technique and observed ozone in winter 2019-2020.