Multi-spatial Multi-temporal Air Quality Forecasting with Integrated Monitoring and Reanalysis Data

TL;DR

This paper introduces M2G2, a dual multi-scale framework for air quality forecasting that jointly models spatial interactions at station and city scales (via MS-GCN) and temporal dynamics across multiple time scales (via MT-GRU). By enabling bidirectional fusion between spatial scales and parallel, dynamically weighted temporal updates, M2G2 achieves superior accuracy for PM2.5, PM10, NO2, and O3 across 3-, 24-, 48-, and 72-hour horizons. Extensive experiments on 41 northern China cities with 152 stations, using meteorological data and KNN-imputed measurements, demonstrate notable improvements over nine baselines and validate the contributions of both MS-GCN and MT-GRU through ablations. The results highlight the importance of multi-spatial and multi-temporal information fusion for robust, long-horizon air quality forecasting with broader applicability to multiple pollutants. The approach offers practical value for urban air quality management by providing more reliable forecasts across diverse pollutants and timeframes.

Abstract

Accurate air quality forecasting is crucial for public health, environmental monitoring and protection, and urban planning. However, existing methods fail to effectively utilize multi-scale information, both spatially and temporally. Spatially, there is a lack of integration between individual monitoring stations and city-wide scales. Temporally, the periodic nature of air quality variations is often overlooked or inadequately considered. To address these limitations, we present a novel Multi-spatial Multi-temporal air quality forecasting method based on Graph Convolutional Networks and Gated Recurrent Units (M2G2), bridging the gap in air quality forecasting across spatial and temporal scales. The proposed framework consists of two modules: Multi-scale Spatial GCN (MS-GCN) for spatial information fusion and Multi-scale Temporal GRU(MT-GRU) for temporal information integration. In the spatial dimension, the MS-GCN module employs a bidirectional learnable structure and a residual structure, enabling comprehensive information exchange between individual monitoring stations and the city-scale graph. Regarding the temporal dimension, the MT-GRU module adaptively combines information from different temporal scales through parallel hidden states. Leveraging meteorological indicators and four air quality indicators, we present comprehensive comparative analyses and ablation experiments, showcasing the higher accuracy of M2G2 in comparison to nine currently available advanced approaches across all aspects. The improvements of M2G2 over the second-best method on RMSE of the 24h/48h/72h are as follows: PM2.5: (7.72%, 6.67%, 10.45%); PM10: (6.43%, 5.68%, 7.73%); NO2: (5.07%, 7.76%, 16.60%); O3: (6.46%, 6.86%, 9.79%). Furthermore, we demonstrate the effectiveness of each module of M2G2 by ablation study.

Figures (11)

• Figure 1: An overview of M2G2. The orange module: M2G2 follows a sequence-to-sequence structure, using historical data to predict future air quality. The inputs are air quality and meteorological indicators to predict the air quality at the next moment. M2G2 consists of Multi-scale Spatial GCN (MS-GCN) and Multi-scale Temporal GRU (MT-GRU). In this framework, $t$ means the current time. The blue module: MS-GCN consists of two scales, station and city, and each of the two scales conducts spatial feature extraction by GCN, which results in cross-scale feature interaction. The output of the MS-GCN will be passed into the MT-GRU. The green module: we have improved the update mechanism of the hidden state in GRU by slicing the hidden state and updating it at different intervals. As shown in the figure the hidden state $\bm{H}^t$ is cut into 3 parts: $\bm{H}_1^t, \bm{H}_2^t, \bm{H}_3^t$. The solid line represents the update, and the dashed line represents the current iteration step to keep the original value. In addition, $\bm{H}_1^t, \bm{H}_2^t \ {\rm and} \ \bm{H}_3^t$ each have a learnable dynamic weight, which corresponds to different temporal scales with different significance for the current prediction.
• Figure 2: The core idea of M2G2: At the spatial scale, we employ the bidirectional fusion module to learn feature information that mutually enhances the city-scale and station-scale, using an end-to-end approach. After spatial feature extraction, the relevant information is fed into the temporal dimension module. In this module, components of different scales are utilized to extract time-series features with distinct periodicities. Finally, these features are aggregated using dynamically learnable weights and produce the final predictions.
• Figure 3: The map of the distribution of air quality monitoring stations
• Figure 4: The prediction for different hour prediction horizons. Subfigure (a) represents the results of predicting pollutant concentrations for the next 3 hours, while (b) to (d) correspond to the next 24 hours, 48 hours, and 72 hours, respectively.
• Figure 5: The fine-grained comparison for 3 hour prediction horizon. The black line represents the ground truth, while the gray lines indicate all techniques save the current one. The remaining colored line reflects the method that corresponds to the current subplot.