Turning up the Heat: Assessing 2‐m Temperature Forecast Errors in Artificial Intelligence Weather Prediction Models During Heat Waves

Abstract

Extreme heat is the deadliest weather‐related hazard in the United States. Furthermore, heat extremes are increasing in intensity, frequency, and duration, making skillful forecasts vital to protecting life and property. Traditional numerical weather prediction (NWP) models struggle with extreme heat forecasts on longer (i.e., medium‐range‐ and subseasonal) timescales. Meanwhile, artificial intelligence‐based weather prediction (AIWP) models are progressing rapidly. However, it is largely unknown how well AIWP models predict extreme heat, especially for longer‐range forecasts. This study investigates 2‐m temperature forecasts for 60 heat waves across four contiguous United States regions (Northwest, Midwest, Northeast, Southeast) and the four boreal seasons at medium‐range to subseasonal (i.e., 10–20‐day) lead times, using two deterministic AIWP models (GraphCast and Pangu‐Weather) and one probabilistic AIWP model (GenCast). All three models exhibit persistent mean cold biases throughout the forecast evolution, especially near the start of the heat wave. However, the GenCast ensemble mean consistently exhibits smaller biases and errors than the two deterministic models. In all three AIWP models, the smallest regional average errors during the heat wave period are in the Northwest, with the largest in the Midwest. In all boreal seasons, mean cold biases occur before the heat waves in the three AIWP models except Pangu‐Weather in winter, which exhibits a mean warm bias before heat wave onset. A brief comparison of 2‐m temperature forecast errors is then illustrated among the AIWP models and a traditional NWP model (NOAA United Forecast System Global Ensemble Forecast System (UFS GEFS)), for which GenCast particularly performs strongly. Overall, results show that the probabilistic AIWP model GenCast offers the most promising accuracy toward medium‐range and subseasonal forecasts of extreme heat, albeit with a consistent slight mean cold bias.