An interpretable machine learning model for seasonal precipitation forecasting.

Area of Science:

• Climate Science
• Machine Learning
• Meteorology

Background:

• Seasonal climate forecasting is crucial for societal welfare and risk management.
• Accurate precipitation forecasts enable proactive mitigation of adverse climate conditions or exploitation of favorable ones.
• Existing models face challenges with limited data common in climate science.

Purpose of the Study:

• Introduce TelNet, a novel sequence-to-sequence machine learning model for seasonal precipitation forecasting.
• Evaluate TelNet's deterministic and probabilistic performance against state-of-the-art models.
• Assess TelNet's interpretability for instance- and lead-wise forecast analysis.

Main Methods:

• Developed TelNet, a simple encoder-decoder-head architecture for sequence-to-sequence learning.
• Utilized past seasonal precipitation and climate indices as model inputs.
• Employed resampling techniques to estimate uncertainty with limited climate data.
• Compared TelNet with dynamical and deep learning models in a high-predictability region.

Main Results:

• TelNet demonstrates high accuracy and calibration across various initialization months and lead times.
• The model performs exceptionally well during the rainy season, where predictable signals are strongest.
• TelNet ranks among the top-performing models, outperforming some state-of-the-art methods.
• Instance- and lead-wise forecast interpretation is enabled through variable selection weights.

Conclusions:

• TelNet offers a robust and accurate solution for short-to-medium lead seasonal precipitation forecasting.
• The model's architecture is suitable for training with limited climate data.
• TelNet provides valuable insights into forecast drivers, enhancing decision-making capabilities.