Temperature forecasts for the continental United States: a deep learning approach using multidimensional features

Accurate weather forecasts are critical for saving lives, emergency services, and future developments. Climate models such as numerical weather prediction models have made significant advancements in weather forecasts, but these models are computationally expensive and can be subject to inaccurate representations of complex natural interconnections. Alternatively, data-driven machine learning methods have provided new dimensions in assisting weather forecasts. In this study, we used convolutional neural networks (CNN) to assess how geopotential height at different levels of the troposphere may affect the predictability of extreme surface temperature (t2m) via two cases. Specifically, we analyzed temperature forecasts over the continental United States at lead times from 1 day to 30 days by incorporating z100, z200, z500, z700, and z925 hPa levels as inputs to the CNN. In the first case, we applied the framework to predict summer temperatures of 2012, which contributed to one of the extreme heatwave events in the U.S. history. The results show that z500 leads to t2m forecasts with relatively less root mean squared errors (RMSE) than other geopotential heights at most of the lead time under consideration, while the inclusion of more atmospheric pressure levels improves t2m forecasts to a limited extent. At the same lead time, we also predicted the z500 patterns with different levels of geopotential height and temperature as the inputs. We found that the combination of z500, t2m, and t850 (temperature at 850 hPa) is associated with less RMSE for the z500 forecasts compared to other inputs. In contrast to the 2012 summer, our second case examined the wintertime temperature of 2014 when the upper Midwest and Great Lakes regions experienced the coldest winter on record. We found that z200 contributes to better t2m predictions for up to 7-days lead times whereas z925 gives better results for z500 forecasts during this cold event. Collectively, the results suggest that for long-range temperature forecasts based on the CNN, including various levels of geopotential heights could be beneficial.