AiM: urban air quality forecasting with grid-embedded recurrent MLP model

Urban air pollution poses a considerable risk to both public health and environmental sustainability, highlighting the importance of precise, low-latency forecasting systems that can be integrated into real-world infrastructures. This research presents the Urban Air Quality Forecasting with Grid-Embedded Recurrent MLP Model (AiM), a hybrid model that merges a recurrent Multi-Layer Perceptron (R-MLP) with a Grid-Embedded framework that takes spatial factors into account, aiming to improve air quality predictions across space and time. It employs a grid-based partitioning strategy for urban monitoring areas, allowing it to effectively capture localized patterns in pollutant dispersion, while the R-MLP aspect addresses the complex temporal dependencies found in multi-pollutant and meteorological time series data. A customized feature engineering pipeline is designed to incorporate pollutant interactions, meteorological variability, and grid adjacency relationships, facilitating the robust learning of cross-regional correlations. Experiments conducted on multi-season, multi-station datasets reveal that AiM achieves superior forecasting accuracy compared to conventional LSTM, GRU, and CNN-RNN hybrids, reducing RMSE by as much as 12.4% and inference latency by 35% on edge devices. Furthermore, the architecture demonstrates high scalability, accommodating dynamic grid reconfiguration and integration with low-power IoT nodes, making it well-suited for real-time deployment in smart city air quality management systems.