[Paper Review] Extreme Precipitation Seasonal Forecast Using a Transformer Neural Network

Daniel Civitarese, Daniela Szwarcman+2 more|arXiv (Cornell University)|Jul 14, 2021

Meteorological Phenomena and Simulations25 citations

TL;DR

The paper uses a Temporal Fusion Transformer (TFT) to forecast weekly maximum daily precipitation quantiles up to six months ahead, outperforming the ECMWF SEAS5 ensemble and matching climatology in some cases.

ABSTRACT

An impact of climate change is the increase in frequency and intensity of extreme precipitation events. However, confidently predicting the likelihood of extreme precipitation at seasonal scales remains an outstanding challenge. Here, we present an approach to forecasting the quantiles of the maximum daily precipitation in each week up to six months ahead using the temporal fusion transformer (TFT) model. Through experiments in two regions, we compare TFT predictions with those of two baselines: climatology and a calibrated ECMWF SEAS5 ensemble forecast (S5). Our results show that, in terms of quantile risk at six month lead time, the TFT predictions significantly outperform those from S5 and show an overall small improvement compared to climatology. The TFT also responds positively to departures from normal that climatology cannot.

Motivation & Objective

• Motivate improved sub-seasonal to seasonal prediction of extreme precipitation under climate change.
• Develop a quantile-based forecast framework for weekly maximum daily rainfall across multiple locations.
• Leverage the Temporal Fusion Transformer to handle static, historical, and future inputs for multi-horizon forecasting.
• Provide interpretable outputs (multiple quantiles) to aid risk assessment and decision making.

Proposed method

• Forecast maximum daily precipitation quantiles (0.1, 0.5, 0.9) for weekly steps up to 26 weeks ahead using TFT.
• Incorporate static inputs (location, altitude), known future inputs (month, ENSO-related predictors), and historical inputs (past precipitation, temperature, soil moisture, geopotential).
• Train TFT with joint quantile loss across all quantiles and horizons (quantile loss QL, q-risk normalization).
• Compare TFT against climatology (1981-2010) and a calibrated ECMWF SEAS5 ensemble (50 members) using region-specific sub-regions and grid-point splits.
• Evaluate using q-risk differences at quantile 0.9 and analyze spatially across Rio de Janeiro and Florida.

Experimental results

Research questions

• RQ1 Can TFT-based quantile forecasting improve extreme precipitation predictions at lead times up to six months compared to state-of-the-art dynamical ensembles?
• RQ2 How does TFT performance vary across regions (Rio de Janeiro and Florida) and against climatology?
• RQ3 What is the value of multi-quantile TFT outputs for risk assessment in sub-seasonal to seasonal forecasts?

Key findings

• TFT significantly outperforms S5 in terms of q-risk for quantile 0.9 across all locations.
• TFT shows overall small improvements over climatology for some targets, and clearly improves upon S5, especially for the 0.9 quantile.
• For Rio, TFT reduces q-risk relative to climatology for all quantiles; relative to S5, TFT improves by up to 29.54% at 0.9 quantile.
• For Florida, TFT improves over S5 across quantiles, with notable gains at 0.9 quantile (41.87% in S5−TFT).
• TFT captures departures from normal better than climatology, by adapting predictions through input variables and time dynamics.