Dose equivalent rate forecasting: A comparison of time series methods and machine learning approaches

Accurate prediction of Dose Equivalent Rate (DER) is essential for radiation monitoring and risk assessment. This study explores multiple modeling approaches, ranging from traditional statistical techniques to advanced deep learning methods, to forecast DER in San Antonio, Texas. We evaluate five models: the Persistence Model, Lasso Regression, K-Nearest Neighbors, Evolutionary Neural Architecture Search, and Recurrent Neural Networks, using DER data from January to December 2019 for training and early 2020 for testing. Performance is assessed based on correlation coefficient (r) and mean squared error across different DER levels. This study highlights the challenge posed by the stochastic nature of DER variations, emphasizing the necessity of longer datasets to enhance predictive reliability. These findings contribute to the development of more robust radiation forecasting models, improving decision-making in radiation protection and environmental safety. Results indicate that Recurrent Neural Networks achieve the best balance between predictive accuracy and error minimization, effectively capturing temporal dependencies in DER fluctuations.