Effectiveness and Limitations of Machine Learning for Predicting Lower Atmosphere Temperature Profiles

Abstract Temperature profiles in the lower atmosphere rapidly evolve during the morning transition, and predicting their details over complex urban environments is particularly important for pollutant dispersal. Temperature profiles can be predicted by operational weather models but are sensitive to model parameterizations. Machine learning models can be used, but require high-resolution profiles that are often unavailable. To overcome this, pre-dawn temperature and wind speed observations from aircraft over Dallas-Fort Worth, TX are used to predict the morning transition. Initially, only temperature is used as an input to assess several types and configurations of machine learning models. Each model’s prediction is evaluated based on various methods because standard metrics, such as mean square error (MSE), may be minimized but miss important features, such as strong temperature gradients. A long short-term memory network with an AdamW optimizer does not result in the lowest MSE but does better capturing the temperature structure. The model tends toward the annual mean by underestimating the onset of mixing and near-surface temperature in summer and overestimating these in the winter. Including wind speed does not have a statistically significant improvement in MSE but performs better sequentially and with the inversion at the top of the boundary layer. Further refinements are likely to improve rapidly updated short-term forecasts, which may provide an additional forecast tool, but large gradients, extreme means, and external forcing remain an inherent challenge. These results highlight the importance of holistic evaluation of machine learning techniques to ensure adequate representation of predicted temperature profiles.