NRFormer: Nationwide Nuclear Radiation Forecasting with Spatio-Temporal Transformer

TL;DR

NRFormer addresses nationwide nuclear radiation forecasting under non-stationary and spatially imbalanced conditions by introducing a spatio-temporal Transformer with three key components: non-stationary temporal attention, imbalance-aware spatial attention, and radiation propagation prompting. The model fuses per-station instance-wise normalization, macro- and proximity-based spatial modeling, and context-aware prompts from location and meteorology to guide predictions for the next $K$ time steps. On two real-world Japan datasets (Japan-4H and Japan-1D), NRFormer outperforms 11 baselines across MAE, RMSE, and MAPE, and exhibits robustness in deployment with 1–24 day forecasts. The work delivers a scalable, real-time forecasting system for emergency planning and public safety, alongside publicly releasing datasets to spur further research in nationwide nuclear radiation forecasting.

Abstract

Nuclear radiation, which refers to the energy emitted from atomic nuclei during decay, poses significant risks to human health and environmental safety. Recently, advancements in monitoring technology have facilitated the effective recording of nuclear radiation levels and related factors, such as weather conditions. The abundance of monitoring data enables the development of accurate and reliable nuclear radiation forecasting models, which play a crucial role in informing decision-making for individuals and governments. However, this task is challenging due to the imbalanced distribution of monitoring stations over a wide spatial range and the non-stationary radiation variation patterns. In this study, we introduce NRFormer, a novel framework tailored for the nationwide prediction of nuclear radiation variations. By integrating a non-stationary temporal attention module, an imbalance-aware spatial attention module, and a radiation propagation prompting module, NRFormer collectively captures complex spatio-temporal dynamics of nuclear radiation. Extensive experiments on two real-world datasets demonstrate the superiority of our proposed framework against 11 baselines.

Figures (11)

• Figure 1: Collaborative nuclear radiation forecasting for national-wide stations in Japan. (a) Spatial distribution of monitoring stations in Japan, high-risk areas (i.e., with more nuclear power plants) have more stations. (b) Time-varying radiation levels at different monitoring stations.
• Figure 2: Empirical study of the imbalanced spatial distribution problem. (1) Full station distribution of our constructed Japan-1D dataset. (2) Evenly sampled station distribution (from Japan-1D dataset). (3) The forecasting performance of GWN on Japan datasets, where GWN-w/o SPM is the model variant that removes the spatial propagation module (SPM).
• Figure 3: Distributions of the radiation and meteorological datasets: (a) spatial distribution of radiation monitoring stations; (b) temporal distribution of daily average radiation in Japan; (c) spatial distribution of meteorological stations; (d) temporal distribution of the daily average value of each meteorological station across Japan.
• Figure 4: Architecture of deployed forecasting system.
• Figure 5: Comparison of radiation patterns in adjacent prefectures, which demonstrates that radiation patterns in neighboring areas can be either very similar or quite dissimilar. (a) and (c) shows the geographical distribution of radiation stations. (b) and (d) present the variation of average radiation levels in corresponding prefectures.

Theorems & Definitions (2)

• Definition 1
• Definition 2