Hybrid multilayer perceptron models optimized by evolutionary algorithms for urban air quality forecasting: a case study of Shiraz, Iran.

Accurate air quality predictions are vital for environmental management and public health, especially in urban areas facing rising pollution levels. This study investigates the potential of hybrid machine learning models based on Multilayer Perceptron neural networks, optimized using three nature-inspired metaheuristic algorithms: Optics-Inspired Optimization, Biogeography-Based Optimization, and the Multiverse Optimizer. The aim is to enhance the prediction of key air quality indicators, carbon monoxide, ozone, nitrogen dioxide, sulfur dioxide, and the Air Quality Index, using long-term air quality measurements collected from Shiraz Province, Iran, from April 2019 to November 2024. These data were used to train and evaluate the models. Each metaheuristic was applied to optimize the Multilayer Perceptron training across varying population sizes, and the performance of the resulting models was evaluated using the root mean squared error and the coefficient of determination on both the training and test datasets. Among the tested models, the hybrid combining the Multiverse Optimizer with the Multilayer Perceptron achieved the best generalization performance, with a test root mean squared error of 6.71 and a test coefficient of determination of 0.9903. The Biogeography-Based Optimization with Multilayer Perceptron model also yielded accurate results, confirming the strength of nature-inspired optimization in neural network training. This study provides a novel comparative framework for air quality forecasting by integrating underexplored metaheuristic algorithms with neural models. The results suggest strong potential for application in smart environmental monitoring systems. Future research could explore hybrid deep learning methods and real-time predictive models using broader environmental datasets.