Leveraging machine learning and climate phase analysis to enhance seasonal forecasts of cold extremes
Abstract
Extreme cold weather events can have devastating impacts on energy production, and human livelihood. In 2021, winter storm Uri brought record cold to Texas, ultimately leading to a widespread failure of the power system. Prediction of extreme cold events at longer leads is important for anticipating energy demand and preparing resources in the event of extreme cold. At seasonal timescales, machine learning based forecasts of surface temperature have shown promise over traditional dynamical forecasts. Large-scale climate phenomena are known to modulate temperature in Texas and may help capture asymmetric relationships between global climate variables and extreme cold events. Here, a climate phase analysis approach is used to construct season-ahead prediction models for the number of extreme cold days (<−6.67 °C) from December—February for the different phases of three major climate phenomena. This approach improves hindcast skill relative to a single-phase model and selects distinct, phase-dependent regions of predictor variables.