Predicting Belgrade Air Pollution Peaks using a Simple Neural Network: A Low-Cost Early Warning System

This study presents the development of a Long Short-Term Memory (LSTM) neural network model designed to predict daily PM2.5 concentrations in Belgrade. The model successfully captures the temporal patterns and seasonal variations of air pollution, closely aligning with observed data from local monitoring stations. It demonstrates strong predictive capabilities for both regular air quality trends and critical peak pollution events, which are vital for protecting public health. The model’s performance surpasses traditional statistical forecasting methods by effectively learning complex temporal dependencies inherent in urban air pollution data. Despite some minor temporal offsets in peak detection, the LSTM model exhibits robustness and balanced accuracy across different pollution levels. These results highlight the potential of deep learning-based approaches as effective early warning tools, supporting better environmental management and urban planning to mitigate the adverse effects of air pollution.