Hybrid forecasting: using statistics and machine learning to integrate predictions from dynamical models

Abstract

Abstract. Hybrid hydroclimatic forecasting systems employ data-driven (statistical or machine learning) methods to harness and integrate a broad variety of predictions from dynamical, physics-based models – such as numerical weather prediction, climate, land, hydrology and Earth system models – into a final prediction product. They are recognised as a promising way of enhancing prediction skill of meteorological and hydroclimatic variables and events, including rainfall, temperature, streamflow, floods, droughts, tropical cyclones, or atmospheric rivers. Hybrid forecasting methods are now receiving growing attention due to advances in weather and climate prediction systems at sub-seasonal to decadal scales, a better appreciation of the strengths of machine learning, plus expanding access to computational resources and methods. Such systems are attractive because they may avoid the need to run a computationally-expensive offline land model, can minimize the effect of biases that exist within dynamical outputs without explicit bias correction and downscaling, benefit from the strengths of machine learning models, and can learn from large datasets, while combining different sources of predictability with varying time-horizons. Here we review recent developments in hybrid hydroclimatic forecasting and outline key challenges and opportunities. These include obtaining physically-explainable results, assimilating human influences from novel data sources, integrating new ensemble techniques to improve predictive skill, creating seamless prediction schemes that merge short to long lead times, incorporating modelled initial land surface and ocean/ice conditions, acknowledging spatial variability in landscape and atmospheric forcing, and increasing the operational uptake of hybrid prediction schemes.

Skillful Decadal Flood Prediction

“…We used a statistical-dynamical framework (Slater et al, 2022) to predict multiyear average winter (DJFM) high flows in the UK using precipitation and temperature. Decadal climate predictions are currently only publicly available as monthly aggregates.…”

“…Hybrid methods have several advantages compared to dynamical methods. They are computationally efficient, and can integrate a wide variety of nonstationary predictor variables, including climate forecasts, large-scale climate indices, and teleconnections (Slater et al, 2022).…”

“…Most operational flood forecasts are produced by forcing a conceptual or physics-based hydrological model with dynamical daily or sub-daily climate inputs (Emerton et al, 2016). Statistical-dynamical, or hybrid, methods present an attractive alternative approach to produce streamflow forecasts, as they avoid the need to run an offline land model and benefit from advances in statistical modeling (AghaKouchak et al, 2022;Slater et al, 2022). Streamflow quantiles are predicted using statistical or machine-learning models driven by dynamical weather or climate predictions.…”

Resilience of UK crop yields to compound climate change

“…Third, to assess future changes in crop yields, we drive the same multiple regression model with the bias-corrected projections of the same variables, computed from the 12 mem-bers of the UKCP Local simulations (i.e. we employ a hybrid approach; see Slater et al, 2022). This approach allows us to fuse together the data-driven regression model with the meteorological simulations for higher greenhouse gas emissions.…”

A comparison of seasonal rainfall forecasts over Central America using dynamic and hybrid approaches from Copernicus Climate Change Service seasonal forecasting system and the North American Multimodel Ensemble

“…Some studies for instance use Canonical Correlation Analysis (CCA) to relate observed SST in the Atlantic and Pacific oceans to regional rainfall, demonstrating the important role SST fields play in rainfall over Central America (Giannini et al ., 2000; Alfaro, 2007; Maldonado et al ., 2013; 2016; 2017; Alfaro et al ., 2016a), including for predictions of low‐ and high‐rainfall events (e.g., rainfall above the 80th and below the 10th percentiles of the monthly climatology; Maldonado et al ., 2013). Hybrid methods then combine dynamic and statistical methods (Slater et al ., 2022). Some hybrid methods generate indirect rainfall forecasts by extracting related variables (e.g., SST) from the GCMs and then statistically translate those values to target variables (e.g., Alfaro et al ., 2018; Strazzo et al ., 2019; Colman et al ., 2020).…”