Graph-enhanced Multi-scale Mamba Network with Whale Optimization for Air Quality Forecasting

Abstract

Accurate forecasting of fine particulate matter (PM2.5) concentration is crucial for public health and environmental management. Existing deep learning methods often struggle to capture complex spatio-temporal dependencies, especially non-Euclidean spatial relationships and multi-scale temporal dynamics. To address these challenges, this paper proposes a Graph-enhanced Multi-scale Mamba Network with Whale Optimization Algorithm (WOA-GMamba). The model first constructs a dynamic graph by fusing a static prior graph (based on geographical distance) and a dynamic posterior graph (based on time-series correlations), then applies sparsification and graph convolutional networks to aggregate spatial features. For temporal modeling, a multi-scale Mamba module is designed to capture short-term fluctuations, periodic changes, and long-term trends simultaneously. Finally, the Whale Optimization Algorithm is employed to globally optimize key structural and training hyperparameters. Experiments on 36 air quality monitoring stations in Beijing demonstrate that WOA-GMamba achieves superior performance in both single-step and multi-step forecasting tasks. Compared to WOA-STMamba (a previous Mamba-based model), WOA-GMamba reduces RMSE by 2.03% and MAE by 2.56%, while ablation studies confirm the effectiveness of each component.