A comparative analysis of hyperparameter optimization using LSTM-based deep learning models for urban air quality predictions

Air pollution poses significant threats to human health and the environment, necessitating accurate prediction models for effective management and mitigation strategies. This study presents a comprehensive analysis of hyperparameter optimization techniques for Long Short-Term Memory (LSTM) based deep learning models in urban air quality forecasting. We focus on predicting concentrations of four key pollutants: carbon monoxide (CO), nitrogen oxides (NO_X), nitrogen dioxide (NO₂), and particulate matter (PM₁₀). The study employs and compares three prominent hyperparameter optimization methods: Random Search, Bayesian Optimization, and Hyperband. Using air quality data from Sakarya, Turkey, collected between January 2020 and September 2022, we first addressed missing data through comparative analysis of mean imputation and k-Nearest Neighbors (kNN) imputation methods. Our results demonstrate that kNN imputation generally outperforms mean imputation, except for NO_X predictions. The hyperparameter-optimized LSTM models consistently outperformed baseline models across all pollutants. Notably, the Hyperband Search algorithm excelled in NO_X prediction, while Bayesian Optimization showed superior performance for other pollutants. Our analysis also revealed temporal trends in pollutant concentrations during the COVID-19 pandemic, including significant reductions in PM₁₀ and CO levels. This study contributes to AI-driven environmental monitoring by comparing hyperparameter optimization techniques in urban air quality modeling. The improved prediction accuracy offered by our optimized models has significant implications for public health protection, environmental policymaking, and smart city initiatives. Our findings underscore the importance of tailored optimization approaches for different pollutants and highlight the potential of advanced machine learning techniques in addressing environmental challenges.