Predicting Surface Temperatures Using Compound Geopotential Heights in a Deep Learning Framework

Predicting surface temperatures accurately and with longer lead times is critically essential for saving lives, emergency services, and future developments. Climate models such as numerical weather prediction models have made significant advancements in weather and climate forecasts, but the computational costs of running these models are high and they can be subject to inaccurate representations of the complex natural interconnections. Alternatively, data-driven machine learning methods have provided new dimensions in improving weather forecasts. In this study, we used Convolutional Neural Networks (CNN) to assess the role of geopotential height at different levels in enhancing the predictability of surface temperature (t2m). Geopotential height, which approximates the actual height of a pressure surface above mean sea level is one of the most important factors for explaining a long-term forecast but often goes overlooked. Herein we used z100, z200, z500, z700, and z925 hPa levels as our inputs to the CNN and analyzed the temperature forecasts over the continental United States at lead time from 1 day to 30 days. We applied the framework to predict the summer temperature of 2012, which contributed to one of the extreme heat wave events in U.S. history. The results show that Z500 leads to t2m forecasts having relatively fewer 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. Overall, the RMSE for the t2m forecast for the 2012 summer varies from 1.4 to 2.9 °F from 1-day to 30-day lead times. For the same time duration, we also predicted the z500 patterns using different levels of geopotential height and temperatures as inputs to the CNN. We found that t2m and T850 (temperature at 850 hpa) have less RMSE for z500 forecasts compared to other geopotential heights. The RMSE for z500 using all input variables ranges from 226 to 855 m²s^-2 from 1-day to 30-day lead times. For the 2014 winter analysis, we found that z200 contributes to better t2m predictions for up to 10-day lead times whereas z925 gives better results for z500 forecasts. Overall, RMSE for t2m in winter varies from 2.2 to 6.2 °F and z500 ranges from 258 to 1460 m²s^-2.