Bio-Inspired Optimization-Driven Air Quality Prediction Using Machine Learning

Abstract

Precise air quality forecasting is vital to environmental management and health safety of people but traditional machine learning solutions tend to experience reduced performance as a result of large-dimensional input features and inefficient hyper-parameter settings. To overcome these shortcomings, this paper suggests a new bioinspired optimization-based machine learning framework where both feature selection and hyper-parameter optimization are done through a single co-optimization. The model presents a multiphase hybrid search methodology combining Whale Optimization Algorithm, which is a global exploration method, with the use of the Gay Wolf Optimizer, which is an adaptive convergence method, and the Particle Swarm Optimization, which is a fine-grained exploitation method, to facilitate efficient search of complex solution spaces. In order to develop Air Quality Index predictions models, real-world air quality data in the form of multivariate pollutant concentrations and meteorological variables were used. The optimized learning parameters were found to be very effective in the errors of forecasting and the complexity of the model. The experimental findings show reductions of errors by over 35 percent compared to traditional machine learning methods and a steady improvement in performance as compared to single-optimizer methods. The XG-Boost optimization model performed best on the predictive performance with an R 2 of 0.951 using a smaller magnitude of features (up to 50% reduction in features) which shows the efficiency of the proposed dimensionality-aware optimization framework. The results suggest that bio-inspired co-optimization could be used to provide reliable, scalable, and computationally efficient air quality prediction systems to be applied in real-time environmental monitoring and smart cities.