AirFlow is proposed, a pollutant-aware dual-stream framework that operates on station multivariate observations without additional graph propagation or predefined signal decomposition, achieving high forecasting accuracy with low computational overhead.
Abstract
Accurate air quality forecasting is essential for public health and urban environmental management, but remains challenging because pollutant channels differ in periodicity and distribution drift, while their concentration trajectories contain both multi-scale dependencies and rapid changes. Recent methods have improved spatial dependency learning and meteorological covariate modeling. However, pollutant channels are still passed through the same normalization rule and temporal backbone, using a shared latent representation for channel-specific distributions and changes at different rates. To address this limitation, we propose AirFlow, a pollutant-aware dual-stream framework that operates on station multivariate observations without additional graph propagation or predefined signal decomposition. Specifically, AirFlow designs two novel blocks: (1) a statistic-guided normalization routing mechanism that selects a normalization path for each pollutant according to its 24-hour autocorrelation and distribution drift; and (2) a hierarchical dual-stream state model that combines multi-scale state space propagation with learnable response coefficients, where gated bidirectional cross-attention exchanges information and adaptively fuses the resulting representations. Experiments on real-world data from multiple cities show that AirFlow achieves the best performance in 34 of 36 metrics comparisons, with reductions of up to 11.11% root mean square error over the state-of-the-art baseline. AirFlow also requires only 0.0483M parameters and 0.0215G FLOPs, achieving high forecasting accuracy with low computational overhead.
A Temporal-Aware Multi-Task Learning (TMTL-AQI) framework to assess urban air quality via structured data that outperforms single-task and baseline multi-task models with an accuracy of 0.7605 and an F1-score of 0.7422.
Iman Youssif Ibrahim, D. M. Ahmed· Dasinya Journal for Engineer...· 0 citations
Disentangling the relative roles of emissions and atmospheric processes in controlling air-pollutant concentrations remains a central challenge in air-quality management. Although machine-learning (ML) models can accurately predict pollutant concentrations, the temporal evolution of the importance of individual drivers is often difficult to interpret. Here, we introduce a framework for reconstructing time-resolved feature importance (FI) in ML air-quality models. Hourly NO2 and PM2.5 concentrations were simulated across clusters of observations, defined along concentration trajectories in a state–space spanned by source intensity and meteorological variables. Within each cluster, predictor importance is quantified and mapped back onto the corresponding time points, yielding continuous FI time series for all predictors. The framework is demonstrated using observations from the nationwide air-quality network in Israel, together with traffic-related source indicators and meteorological parameters. The dominant drivers differ markedly between the two pollutants: NO2 variability is primarily associated with local emissions, mechanical transport, and turbulent mixing, whereas PM2.5 variability reflects predictors that are related to nucleation, coagulation, hygroscopic growth, long-range transport, and chemical transformation. The feature importance exhibits pronounced seasonal, regional, and diurnal variability, including modulation around traffic rush hours. These results demonstrate the value of time-resolved interpretability for diagnosing drivers of air-pollutant variability and improving the representation of processes in statistical air-quality models.
Yuval, Yoav Levi, P. Khain et al.· Atmosphere· 0 citations
Fine particulate matter (PM2.5) concentration is a critical indicator of air quality and is closely related to human health and ecological environments. Accurate multi-site PM2.5 forecasting still faces considerable challenges: PM2.5 exhibits non-stationary distribution drift, monitoring sites are unevenly distributed, and temporal evolution is tightly coupled with spatial interactions. To tackle these problems, this paper proposes an Adaptive Spatiotemporal Graph Transformer (AST-GT) framework for multi-site PM2.5 multi-step forecasting. The framework incorporates five core components: adaptive non-stationary normalization to alleviate time-varying distribution drift and feature scale mismatch; Transformer-based temporal representation learning to capture long-range temporal dependencies; multi-source context encoding to fuse geographic location, meteorological conditions, and co-pollutant information; graph-based dual-scale spatial attention and Temporal-Spatial Cooperative Attention mechanism to jointly model spatial correlations and spatiotemporal interactions, particularly under uneven site distribution. Extensive experiments are conducted on multi-site datasets from Beijing and India to validate the performance of the proposed AST-GT. The results demonstrate the model's effectiveness, accuracy, and robustness across different datasets, highlighting its generalization capability under irregular spatial distributions, and offering a reliable solution for multi-site multi-step air quality prediction and refined environmental management.
Yi-Di Shi, Jun Yang, Dunwang Qin· Journal of Environmental Man...· 0 citations
ABSTRACT Atmospheric pollution remains a major concern for public health and city sustainability, especially in densely populated metropolitan regions with complex environmental dynamics. Traditional prediction techniques often rely on static feature selection methods that disregard the spatial and temporal variations embedded in pollutant behaviour. In this work, we introduce FAST (Feature-Adaptive Selection and Transformation), a novel approach for dynamic feature selection for spatiotemporal air quality forecasting. FAST applies attention-based methods to dynamically update relevant features based on spatial and meteorological conditions at any given time while improving interpretability and predictive performance. We evaluate FAST over three of the largest Indian metropolitan areas, Delhi, Mumbai, and Bengaluru, based on real-world air quality measurements from 90 + monitoring stations. The system is benchmarked against conventional filter, wrapper, and embedded feature selection techniques and demonstrates consistent improvements in forecasting accuracy, with up to 18% lower Mean Absolute Error (MAE) and up to 12% improvement in R2 values. FAST demonstrates consistent performance across the evaluated 6-hour and 48-hour forecasting horizons while also providing interpretable insights into seasonal and spatial patterns of pollution. By dynamically adapting to the evolving nature of urban air quality, FAST provides an approach for data-driven air quality forecasting in urban environments. GRAPHICAL ABSTRACTAn infographic summarizing a dynamic feature adaptive learning framework for spatiotemporal urban air pollution forecasting in India.The figure shows an infographic describing a dynamic feature adaptive representation learning framework for spatiotemporal air pollution forecasting. At the top, the title reads Dynamic Feature Adaptive Representation Learning for Spatiotemporal Air Pollution Forecasting. On the left, a rounded box labeled Background lists bullet points about urban air pollution as a public health concern, limitations of traditional static feature models, the design of a new attention based deep learning framework for dynamic feature selection in space and time, and recalibration of input relevance based on context. Below this, two small outline maps depict India and a highlighted region, and a larger map titled Air Pollution Monitoring Stations shows dense station locations across the highlighted area. In the center, an illustration shows an urban neighborhood with buildings, factories emitting smoke, trees, scattered trash, and people walking, representing polluted city conditions. Beneath this illustration, a block labeled FAST Framework connects four colored rectangles: one for pollutant, geospatial, and meteorological data; one for FAST feature engineering; one for dynamic feature importance estimation; and one for a deep forecasting block. On the right, a rounded box labeled Key Findings contains bullet points about achieving up to 18 percent lower mean absolute error and 12 percent higher coefficient of determination compared with traditional techniques, reporting lowest total loss values of 0.13 for Delhi, 0.17 for Mumbai, and 0.19 for Bengaluru, generalizing across diverse urban environments, and reducing forecasting error while improving interpretability. Farther right, a section labeled Forecasting Results contains several small line graphs with curves over time and several three dimensional surface plots that resemble low hills and peaks, representing spatial and temporal pollution forecasts; axes and numeric scales are too small to read. At the bottom, a full width text bar labeled Conclusion states that the FAST framework improves forecasting accuracy and interpretability, outperforms traditional methods across major Indian cities and forecasting horizons, and offers scalable support for data driven environmental policy making.
Air pollution poses a serious hazard to public health in Cities of India; Among all sources, those resulting from Energy Consumption dominate the deaths caused by air pollution. Economic growth has intensified air pollution and climate change issues at the same time. Existing prediction methods focus on isolated single-site time-series forecasting, ignoring spatial dependencies and cross-city pollution transport, limiting regional management effectiveness. Based on the hourly air quality data of key pollutants in seven cities across India during 2015-2020, this paper first explores their concentrations and correlations; Then it builds a Deep Learning forecasting system consisting of Long Short-Term Memory (LSTM), a convolutional neural network, and long short-term memory networks (CNN-LSTM) and a transformer to investigate its AQI-prediction performance. LSTM can capture long-term temporal dependencies; CNN-LSTM combines spatial-temporal information model integration; Transformers explore long-range relationships through attention mechanisms. Based on the experimental results of this paper, it can be concluded that CNN-LSTM is more effective than other systems. It also has some advantages over single-LSTM and the transformer.
Short-term prediction of hazardous gas concentrations is crucial for industrial air monitoring, but conventional approaches often fail to capture abrupt local fluctuations and nonlinear temporal dependencies, limiting prediction accuracy. To address these limitations, this study develops a multi-task residual Transformer-based framework for short-term concentration forecasting. First, historical high-frequency H2S measurements are processed using a sliding-window approach to form input sequences for the model. Next, a shared Transformer encoder extracts temporal features, while task-specific branches perform residual concentration prediction and concentration-based emission-state classification. Within this multi-task framework, an adaptive weighting mechanism emphasizes high-variation samples during training to improve sensitivity to rapid concentration changes. Experiments conducted on data from the South Coast Air Quality Management District demonstrate that, averaged over three random seeds, the model achieves an MAE of 0.133±0.001, an RMSE of 0.237±0.000, and an R2 of 0.810±0.001 for one-observation-step forecasting. These results show that the proposed framework effectively captures abrupt rises and peak concentrations, providing a reliable tool for industrial emission monitoring and early warning applications.
Ning Jin, Zhiying Wang, Ruohan Ma· Mathematics· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
