Sub-Seasonal Forecasting From Necessity to Improvement: Identifying Climate Shifts, Evaluating Predictability, and Advancing with Neural Networks

Abstract

As climate change accelerates, extreme weather events are occurring with increased frequency and severity, highlighting an urgent need for improved sub-seasonal forecasting. This dissertation investigates sub-seasonal forecasting as an adaptation tool by examining the Time of Emergence (TOE) when climate change signals exceed natural variability. Specifically, the study finds that peak summer temperatures in South Korea and water storage in the western United States are nearing climate thresholds sooner than anticipated, marking the emergence of a new normal in these regions. To address current forecasting challenges, this work introduces a novel metric that combines mean state and seasonal cycles, providing an approach to evaluating forecast performance. This metric reveals the relationship between the performance in simulating the seasonal cycle and the sub-seasonal predictability of Sub-seasonal to Seasonal (S2S) models, suggesting a new direction for improving these models. Furthermore, this research applies neural network-based post-processing techniques—specifically, U-Net architectures—to refine forecasts generated by the European Centre for Medium-Range Weather Forecasts (ECMWF) model. These advanced deep-learning techniques substantially increase forecast accuracy and enable downscaling to finer spatial resolutions. However, challenges remain in reliably predicting extreme events, which require further model refinement. Despite these constraints, enhancing sub-seasonal forecasting is crucial in responding to the upcoming new normal, and further development and effort are needed.