An interpretable machine learning model for seasonal precipitation forecasting

Abstract Seasonal climate forecasting is important for societal welfare, as it supports decision-makers in taking proactive steps to mitigate risks from adverse climate conditions or to take advantage of favorable ones. Here, we introduce TelNet, a sequence-to-sequence machine learning model for sho...

Full description

Saved in:
Bibliographic Details
Main Authors: Enzo Pinheiro, Taha B. M. J. Ouarda
Format: Article
Language:English
Published: Nature Portfolio 2025-03-01
Series:Communications Earth & Environment
Online Access:https://doi.org/10.1038/s43247-025-02207-2
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Abstract Seasonal climate forecasting is important for societal welfare, as it supports decision-makers in taking proactive steps to mitigate risks from adverse climate conditions or to take advantage of favorable ones. Here, we introduce TelNet, a sequence-to-sequence machine learning model for short-to-medium lead seasonal precipitation forecasting. The model takes past seasonal precipitation values and climate indices to predict an empirical precipitation distribution for every grid point of the target region for the next six overlapping seasons. TelNet has a simple encoder-decoder-head architecture, allowing the model to be trained with a limited amount of data, as is often the case in climate forecasting. Its deterministic and probabilistic performance is thoroughly evaluated and compared with state-of-the-art dynamical and deep learning models in a prominent region for seasonal forecasting studies due to its high climate predictability. The training, validation, and test sets are resampled multiple times to estimate the uncertainty associated with a small dataset. The results show that TelNet ranks among the most accurate and calibrated models across multiple initialization months and lead times, especially during the rainy season when the predictable signal is strongest. Moreover, the model allows instance- and lead-wise forecast interpretation through its variable selection weights.
ISSN:2662-4435