Predicting Heat Extremes over the Contiguous United States using Deep Learning

Predicting extreme temperatures is becoming crucial with changing climate. The accuracy of predictions is paramount in saving lives, emergency preparedness and for businesses and environment. In this study we have used deep learning to predict the summer hot conditions over the contiguous United States. We have defined hot conditions using Standardized Temperature Index (STI). STI is used to characterize the severity of temperature extremes. We predicted the STI using Deep Convolutional Neural Network (CNN). CNN is trained to predict the future STI values at a lead time from 1-day to up to 30-days. We used ERA-5 hourly data of temperature and geopotential heights to train, validate and test the CNN. We used multiple levels of geopotential heights such as z200, z500, z700 to capture the influence of different atmospheric levels on prediction performance. We evaluated the accuracy of CNN using Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Anomaly Correlation Coefficient (ACC). Our results show CNN can predict the STI thresholds with reasonable accuracy at short lead times while a combination of geopotential heights impact the forecasts at longer lead times. This new deep learning approach can be a valuable tool for heat events prediction.