Artificial neural network modeling for air pollution prediction

Accurate prediction of air pollutant concentrations remains a critical challenge for environmental monitoring and public health, demanding robust and adaptive artificial intelligence approaches. This study investigates the effectiveness of various types of artificial neural networks (ANNs), including Long Short-Term Memory networks (LSTM) and networks based on the Levenberg-Marquardt algorithm (LM) and its variant with Bayesian regularization (LMBR), in predicting air pollution under different data conditions. Since LSTM networks are based on first derivative loss function algorithms and the LM algorithm is usually superior to this type of algorithm, a comparison of these networks was conducted. This is further supported by the limited coverage of this topic in the existing literature. ANNs were tested on two different datasets: the Air Quality dataset, where the target variable was the concentration of benzene (C6H6) and the Beijing PM2.5 dataset, where the target was the concentration of PM2.5 particles. The performance metrics of the ANNs were the root mean square error (RMSE), the mean absolute error (MAE), and the mean absolute percentage error (MAPE). It is shown that, in the case of the Air Quality dataset, the values of these parameters RMSE = 0.11 μg/m3, MAE = 0.09 μg/m3, MAPE = 1% for LSTM networks and RMSE = 0.14 μg/m3, MAE = 0.092 μg/m3, MAPE = 1% for LMBR networks, were competitive. For LM networks, these values were significantly higher: RMSE = 0.57 μg/m3, MAE = 0.2 μg/m3, MAPE = 2%. Contrastingly, in the case of the Beijing database, the values of all parameters were drastically higher: RMSE = {45.74 μg/m3, 64.94 μg/m3, 65.68 μg/m3}, MAE = {30.32 μg/m3, 42.83 μg/m3, 54.5 μg/m3} and MAPE = {52%, 72%, 74.25%} for LSTM, LMBR, and LM networks, respectively. In this case, the benzene concentration values exhibited a strictly linear correlation with the input variables. For the Beijing dataset, the relationships between PM2.5 concentration and predictor values were non-monotonic, leading to a drastic drop in the performance of all networks. Findings reveal that, in this case, LSTM networks were more robust compared to LMBR and LM networks, as the values of their RMSE, MAE, and MAPE parameters were significantly lower. Furthermore, it is shown that the input variable selection technique (IVS) can be used to detect seasonal trends in the input data.