A hybrid deep learning model–based LSTM and modified genetic algorithm for air quality applications

Over time, computing power and storage resource advancements have enabled the widespread accumulation and utilization of data across various domains. In the field of air quality, analyzing data and developing air quality models have become pivotal in safeguarding public health. Despite significant progress in modeling, the critical need for accurate pollutant predictions persists. In addressing this challenge, deep learning models have garnered substantial attention in research due to their outstanding performance across diverse applications. However, the optimization of hyperparameters and features remains a challenging task. This study seeks to leverage historical data to construct the long short-term memory–based model for forecasting multistep PM₁₀. To refine its architecture, a modified genetic algorithm is employed for automatic design. Furthermore, we explore principal component analysis and exhaustive feature selection to identify the optimal feature set. This paper introduces a novel hybrid deep learning model named EFS-GA-LSTM, tailored for multistep hourly PM₁₀ forecasting. To assess its performance, we compare it with other hyperparameter optimization algorithms, including particle swarm optimization, variable neighborhood search, and Bayesian optimization with Gaussian process. The input dataset comprises hourly PM₁₀ concentrations, meteorological variables, and time variables. The results reveal that for 3-h-ahead forecasting tasks, the EFS-GA-LSTM network demonstrates improvements in root mean square error, mean absolute percentage error, correlation coefficient, and coefficient of determination.