Bridging ML and Subseasonal-to-Seasonal Forecasting: Exploring Sources of Predictability of Hydroclimatic Extremes Using Explainable Neural Networks

Hydroclimatic extremes such as droughts and heatwaves can be supported by precipitation and temperature forecasts (White et al. 2022). These predictions can aid the decision-making process across a wide range of industries at the subseasonal-to-seasonal (S2S) timescale (White et al. 2017), which includes lead times spanning two weeks up to a few months. In particular, they are highly sought by sectors such as public health (Disera et

al. 2020), disaster mitigation (De Perez et al. 2016), water resource management (Baker et al. 2019), energy and utilities (Soret et al. 2019), emergency management and response (Lala et al. 2022), and agriculture (Flohr et al. 2017). Despite this need, there is still a well-known gap in the predictive skill of forecasts in the S2S range, commonly referred to as the S2S prediction gap (Mariotti et al. 2018). The challenge with generating reliable forecasts

at this timescale arises because the lead times begin after atmospheric memory starts to fade, yet are too short for the longer-term oceanic memory to dominate (Vitart & Robertson, 2018).

A promising way to enhance the predictive skill of hydroclimatic variables and extreme events is through hybrid modelling. This is a type of forecasting approach that uses data-driven methods such as machine learning to combine a wide variety of predictors from physics-based models (Slater et al. 2023). Hybrid models leverage the strength of both methodologies by incorporating empirical knowledge about past events and an understanding of the physical processes involved.

A lot of fundamental questions on how to best train these models to extract the maximum information available from the predictors have not yet been answered. Such questions have been described in more detail in the "Intended Outcome" section below, and we aim to address them in this project. By doing so, we hope to find ways to increase the predictive skill of the forecasts at this timescale.