Enhancing AI and Dynamical Subseasonal Forecasts with Probabilistic Bias Correction

2020-01-03 · Source: stat.ML updates on arXiv.org · Field: Science & Research — Environmental Science & Earth Systems, Mathematics & Computational Sciences, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

A new machine learning framework, Probabilistic Bias Correction (PBC), significantly enhances subseasonal weather forecasts (2-6 weeks ahead) by correcting historical probabilistic predictions. Applied to leading models from the European Centre for Medium-Range Weather Forecasts (ECMWF), PBC doubles the subseasonal skill of the AI Forecasting System (AIFS-SUBS) and improves the operationally-debiased dynamical model's skill for 91% of pressure, 92% of temperature, and 98% of precipitation targets. PBC also consistently outperforms existing operational debiasing protocols and other AI-based subseasonal models like FuXi-S2S. In ECMWF's 2025 AI Weather Quest competition, an ensemble using PBC (MicroDuet) placed first globally for all weather variables and lead times, outperforming 34 other teams and multiple operational forecasting centers. This framework also substantially boosts the early detection of extreme weather events, improving flood forecasting skill.

Key takeaway

For AI Scientists and Machine Learning Engineers working on weather prediction, PBC offers a robust post-processing solution to overcome the "predictability desert" at subseasonal timescales. You should consider implementing PBC to enhance the skill of your existing dynamical or AI forecasting systems, particularly for improving the accuracy of extreme weather event predictions and overall subseasonal forecast quality. This framework provides a low-overhead strategy for significant skill gains.

Key insights

Probabilistic Bias Correction (PBC) significantly improves subseasonal weather forecast accuracy and extreme event prediction.

Principles

• Systematic forecast errors are predictable and correctable.
• Probabilistic bias correction can enhance both dynamical and AI models.

Method

PBC converts input ensembles into initial probabilistic forecasts, then applies Debias++ and Persistence++ machine learning models in parallel to generate complementary corrections, which are then averaged.

In practice

• Apply PBC to existing dynamical or AI weather models.
• Use PBC for improved extreme weather event detection.
• Integrate PBC into operational forecasting systems.

Topics

• Probabilistic Bias Correction
• Subseasonal Weather Forecasting
• ECMWF AI Forecasting System
• AI Weather Quest Competition
• Extreme Weather Prediction

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Rigorous uncertainty quantification of probabilistic AI weather forecasts with conformal prediction — Online conformal prediction rigorously calibrates AI weather forecasts, guaranteeing statistical coverage without performance loss.
• Hybrid NARX-LLM for Greenland Iceberg Discharge: Prompt-Driven Residual Correction — A hybrid NARX-LLM framework uses physics-informed prompts to correct iceberg discharge predictions, enhancing accuracy for complex climate dynamics.
• Ensemble High-Resolution Weather Forecasting on AMD Instinct GPU Accelerators — Diffusion models enable robust ensemble weather forecasting by generating diverse atmospheric states to quantify uncertainty.
• A Free Baseline That Beats Amazon’s Neural Net — Probabilistic forecasting should always validate complex models against naive baselines to ensure true value.
• Hybrid Probabilistic Forecasting of Under-Five Malaria Admissions in Ghana: A Gaussian Process Regression with Holt-Winters Smoothing — Combining GPR with Holt-Winters provides robust probabilistic malaria forecasting, capturing non-linearity and seasonality.
• ForecastBench-Sim: A Simulated-World Forecasting Benchmark — ForecastBench-Sim offers a controlled, simulated environment for evaluating probabilistic reasoning in dynamic systems, overcoming real-world forecasting limitations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.