Beyond the Predictability Desert: A Data-Driven Approach to 5-Week Temperature Forecasting

Introduction

Accurate subseasonal forecasting of weather conditions is essential for anticipating extreme events like heatwaves, yet it remains a major gap for traditional dynamical models. We propose a data-driven neural approach for predicting temperature anomalies at 4- and 5-week lead times, designed to rival the European Centre for Medium-Range Weather Forecasts (ECMWF)'s operational model, the Integrated Forecasting System (IFS).

Methods

The forecasting task is framed as a sequence-to-sequence problem, mapping a 12-week historical lookback of ten environmental variables to the subsequent five weeks of temperature anomalies. To handle the highly heterogeneous nature of the input features, we designed a multi-encoder residual U-Net. Rather than early concatenation, ten autonomous temporal encoders independently compress the trajectory of each variable before spatial mixing. To respect the global topology, all convolutions utilize custom longitudinal periodic padding, eliminating artificial meridian boundaries. Training follows a two-stage strategy: pre-training on a low-resolution 2000-year paleo-climate simulation, followed by sliding-window fine-tuning on the high-resolution ERA5 reanalysis. The objective incorporates Huber loss paired with spatial gradient loss; forcing the model to forecast weeks 1–3 alongside the target weeks implicitly regulates long-term stability.

Results

Developed for ECMWF's "AI Weather Quest", initial evaluations across 20 years (2006–2025) demonstrate a lower Root Mean Square Error (RMSE) than IFS when validated against ERA5, starting at week 4. At week 5 over continental areas, our model achieved an RMSE of 2.695 K versus 2.814 K for IFS, outperforming it across 64.5% of the landmass.

Conclusions

Our findings indicate that our model successfully outperforms IFS for the evaluated period, while reducing computational requirements to a workstation scale. While IFS performance degrades sharply over time, our model maintains greater stability at longer lead times. Analyzing spatial error maps reveals our network leverages different physical relationships than the numerical model, highlighting potentially overlooked mechanisms in the physical domain.