genetic algorithm-optimized deep recurrent networks for pm2.5 forecasting: a case study on urban air pollution in tehran

Air pollution, especially fine particulate matter (pm2.5), remains a major public health and environmental concern in megacities like tehran. this study proposes a hybrid deep learning framework that integrates recurrent neural networks (rnn, lstm, gru) with genetic algorithm-based hyperparameter optimization to forecast pm2.5 concentrations. to address missing values and inconsistent timestamps in the raw monitoring data, we reconstructed a complete and regular hourly-resolution time series by generating all 24 hourly records for each station and applying forward- and backward-filling imputation. additionally, we introduced station-wise embeddings to capture spatial variability across monitoring sites. the dataset was divided into 80% training and 20% testing, with 20% of the training data further reserved for validation. the proposed method significantly outperformed baseline models, achieving a mean absolute error (mae) of 0.0045, root mean square error (rmse) of 0.0102, and r² score of 0.9966. these findings emphasize the effectiveness of evolutionary hyperparameter tuning and spatio-temporal modeling for air quality prediction systems.