A novel transformer-based method for predicting air absorbed dose rates in nuclear radiation environmental monitoring.

Many studies have used various methods to estimate future nuclear radiation levels to control radiation contamination, provide early warnings, and protect public health and the environment. However, due to the high uncertainty and complexity of nuclear radiation data, existing prediction methods face the challenges of low prediction accuracy and short warning time. Therefore, accurate prediction of nuclear radiation levels is essential to safeguard human health and safety. This study proposes a novel Mixformer model to predict future hourly nuclear radiation data. The seasonality and trend of nuclear radiation data are extracted by data decomposition. To address the slow speed problem common in traditional methods for long-time series prediction tasks, Mixformer simplifies the decoder with convolutional layers to speed up the convergence of the model. The experiments consider the air-absorbed dose rate of nuclear radiation data, spectral data, six climatic conditions, and two other conditions. We use MSE and MAE metrics to verify the effectiveness of Mixformer prediction. The results show that the Mixformer proposed in this paper has better prediction performance compared to the currently popular models. Therefore, the proposed model is a feasible method for industrial nuclear radiation data processing and prediction.