Abstract
Nuclear radiation, the energy released during atomic decay, poses persistent risks to public health and the environment, heightened by concerns since the Fukushima accident and the recent commencement of treated-water discharge. Modern monitoring networks now record radiation levels and accompanying weather conditions at thousands of stations, paving the way for nationwide forecasting to inform emergency response, agricultural advisories, and routine public-safety decisions. However, transforming this abundance of monitoring data into reliable forecasts is challenging for three reasons:
- Non-stationary Time Series: The time series at each station are highly non-stationary, influenced by radioactive decay, weather variability, and irregular human interventions.
- Uneven Spatial Distribution: Monitoring stations are severely unevenly distributed, with roughly 78% of Japan's stations clustered within less than 6% of the country, violating the assumptions of standard graph-based models.
- Co-evolution with Heterogeneous Context: Radiation co-evolves with heterogeneous context such as wind, temperature, and humidity through atmospheric transport processes, which purely data-driven models struggle to capture from observations alone.
In this study, we introduce NRFormer+, a spatio-temporal Transformer for nationwide nuclear radiation forecasting. NRFormer+ couples non-stationary temporal attention and density-adaptive spatial attention with a new atmospheric diffusion module that estimates how meteorology drives radiation dispersion and injects this physical signal into the network as an architectural prior. NRFormer+ achieves state-of-the-art accuracy across all 13 baselines, reducing sudden-change MAE by up to 19.1% over the strongest baseline at comparable inference latency. Our code and datasets are publicly available at GitHub.
Blogger's Review: The introduction of NRFormer+ effectively addresses the complexities in nuclear radiation forecasting, particularly by integrating meteorological data with deep learning architecture, enhancing model accuracy and practicality. This innovative approach provides robust support for emergency response and public safety decisions, making it worthy of broader application across various fields.