Early Predictions of Extreme Heat Events in the Eastern United States Using Machine Learning

Early predictions of extreme heat events can alleviate their negative impacts on public health, agriculture and regional economies; however, the current seasonal and sub-seasonal forecasting models exhibit modest skills in predictions of heat waves. In this study, various machine learning techniques are used for long-lead forecast of heat events in the eastern United States using the Pacific Ocean sea surface temperatures anomalies. Using historical measurements for the eastern United States and sea surface temperature from reanalysis, the machine learning models were trained to predict heat events in the eastern United States 20, 30, 40, and 50 days in advance. Results of cross-validation show that the machine learning models have significant statistical skill in forecasting extreme heat waves for binary forecasts. Relative Operating Characteristic (ROC) score was used for accessing the models’ skills due to its applicability to both binary and unique events. Furthermore, the machine learning models were evaluated using hot days in 2016, 2017 and 2018.