TSSM triples weather forecast accuracy with 80% missing data resilience
New AI model beats baselines by 61% on extreme weather metrics...
A team of researchers from multiple institutions (including Songru Yang and Lei Bai) has introduced TSSM, a novel Triaxial State Space Model designed specifically for Global Station Weather Forecasting (GSWF). Traditional methods struggle with chaotic weather dynamics because they rely too heavily on short-term look-back windows. TSSM addresses this by incorporating period-aligned historical weather data through a Temporal-Variable-Historical paradigm. It stacks historical samples into batches aligned by seasonal cycles, allowing the model to capture long-term, large-scale periodic patterns beyond the immediate past.
TSSM achieves state-of-the-art performance on Weather-5K, the largest station weather dataset to date, with 10% improvement in overall accuracy and a striking 61% gain in extreme event metrics. It also obtains 95% best or second-best results on human-involved datasets. The model excels in long-horizon forecasting—37.5% better at 240 hours—and iterative settings show up to 103.5% improvement under a 48-hour by 5-step rollout. Crucially, TSSM maintains >90% performance even when 80% of observations are missing, far surpassing baselines (which drop below 43%). This robustness makes it highly practical for global in-situ observation networks.
- TSSM uses period-aligned historical data to overcome short-term pattern reliance, improving extreme weather prediction by 61%.
- On Weather-5K, it achieves 10% accuracy gains and retains >90% performance under 80% missing observations vs. <43% for baselines.
- Iterative forecasting shows up to 103.5% improvement (48h×5), demonstrating strong long-horizon capability.
Why It Matters
Enables reliable extreme weather warnings from sparse sensor networks, potentially saving lives and infrastructure costs globally.