Study on field strength prediction using different models on time series from urban continuous RF-EMF monitoring

Abstract

State-of-the-art electromagnetic field (EMF) monitoring networks, such as the latest one – the Serbian EMF RATEL system, are able to provide continuous and daily monitoring of radio-frequency (RF) EMF levels, which is especially important for urban areas where people may spend many hours and, consequently, experience increased sensitivity to RF-EMF exposure. By generating considerable time series sets of RF-EMF data, these monitoring networks are initiating a new research topic – near-future RF-EMF prediction, which is valuable for a number of public health activities, from supplementing EMF monitoring in high-risk areas, to proactively reducing exposure times, and towards advancing pre-testing of EMF compliance. This paper investigates the impact of different models on the prediction of field strength in urban environments, where Seasonal Auto-Regressive Integrated Moving Average (SARIMA), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Extreme Learning Machine (ELM), Partial Least Squares Regression (PLS) and Transformer models are considered. The prediction performance of each model is analyzed on a case study of EMF-sensitive areas in the Serbian city of Novi Sad, i.e., two kindergartens and an elementary school; however, the established framework has strong potential for generalization to other urban environments. Based on two-year long monitoring data sets, a comprehensive comparison of the six models on prediction accuracy, performance degradation rate, extreme value prediction accuracy, and training time is made, showing that the PLS model outperforms other models in predicting EMF exposure. This preliminary study may be a valuable reference for large-scale deployment in real-time monitoring systems for public health protection and may trigger additional research on this ultimate EMF topic.