A multi-task stations cooperative air quality prediction system for sustainable development

The study addresses the pressing need for reliable air quality prediction (AQP) to support sustainable development and public health. Traditional physically based models (e.g., WRF-CMAQ) often struggle due to uncertainty in pollutant formation mechanisms, difficulty in integrating complex meteorological features, and inadequate cooperation among spatially distributed monitoring stations. Statistical and machine learning approaches can improve accuracy but still face challenges in extracting meteorological features across regions and leveraging inter-station information. The authors propose an integrated system, MTSC-AQP, to improve AQI forecasting by combining multi-source meteorological and pollutant data, learning temporal-spatial dependencies with LSTM, and enabling cooperation among stations via a gravity-model-based mechanism.

The literature on AQP broadly splits into: (1) monitoring instrument measurement inference methods (physically based), and (2) parametric or non-parametric statistical methods, including modern machine learning and deep learning. Physically based approaches (e.g., WRF-CMAQ) infer air quality via chemical and physical formation mechanisms and meteorology. While foundational, they suffer from fuzzy/complex secondary pollutant chemistry (e.g., O3 formation), model-process complexity, and can yield low predictive accuracy or inconsistencies between predicted and observed concentrations. Subsequent efforts improved representation of processes (e.g., gas/aerosol chemistry, deposition), yet complexity remains a barrier to accurate forecasts. Statistical/ML methods, including linear models, tree-based ensembles, CNN/RNN/LSTM-based models, spatial-temporal ensembles, and hybrid deep architectures, often outperform physically based models in accuracy by learning from data, but they can struggle to extract meteorological features across areas and typically underutilize inter-station cooperation. The review motivates a hybrid data-driven approach that fuses meteorology with pollutant histories and explicitly models station cooperation.

The MTSC-AQP methodology comprises three tasks supported by multi-source data (six pollutants: CO, SO2, NO2, O3, PM10, PM2.5; meteorology: temperature, pressure, wind speed, humidity, wind direction). Data are hourly, cleaned for outliers (forward fill for invalid values such as negatives/zeros/nulls/‘−’ and concentrations >2600 µg/m³), and normalized via Min-Max scaling to [0,1].

Task 1: Air quality analysis

• IAQI calculation: For each pollutant P with concentration Cp, compute IAQI using linear interpolation between bounding breakpoints (BP_L, BP_H) and their IAQI values (IAQI_L, IAQI_H). O3 concentration uses the maximum 8-hour moving average; other pollutants use 24-hour averages. AQI is the max over pollutant IAQIs. AQI levels (I–VI) map to ranges [0–50], [51–100], [101–150], [151–200], [201–300], [301+]. Primary pollutant is defined as the pollutant with the highest IAQI when AQI>50.
• Correlation analysis: Pearson correlation is used to assess relationships between IAQIs and meteorological variables, yielding a correlation matrix to understand meteorological drivers of pollutant variability and AQI.

Task 2: WRF-LSTM module for initial pollutant prediction

• WRF supplies mesoscale meteorological forecasts over broad spatial scales; these forecasts are aligned hourly with measured pollutants to form fused time sequences F_mix.
• An LSTM network ingests sequences of fused features to extract temporal patterns and predict initial pollutant concentrations. The network uses standard LSTM gates (input, forget, output) and dense layers to output multi-pollutant predictions aligned to the original shape.
• Loss function: Mean Absolute Error (MAE) over time steps and pollutants is minimized using Adam. Training uses batch size 32, dropout 0.5, initial learning rate 1e-4 with early stopping, on a system with Intel i9-13900k, NVIDIA 4080, TensorFlow 2.5, CUDA 11.8, Python 3.8. Approximately 20 minutes for 200 epochs.

Task 3: Multi-station cooperative AQP system

• Inputs: combines initial pollutant predictions from Task 2 with multi-source measured data at multiple stations via an LSTM module similar to Task 2 to refine station-level predictions.
• Gravity model for station cooperation: Models stations as nodes in an undirected weighted graph with edge weights based on Euclidean distances. Final target station predictions aggregate sub-station predictions using a gravity mechanism where influence is proportional to correlation and inversely related to distance. Two formulations are described: (i) A' = Σ (k_i A_i / d_i) and (ii) a regression-style form A' = β + Σ k·d_i·A_i, where k (or k_i) are gravity coefficients and d_i distances. This encodes spatial cooperation among stations.
• Data split for Task 3: Multi-source data from 12 July 2020 to 12 July 2021 are split into 8760 hourly sequences; 6873 for training and the remainder for testing. Seven-day windows (168 sequences) are used: six days as input and the 7th day target (daily prediction at 0:00 sequence), with a 24-sequence interval.

Evaluation

• Metrics: RMSE and MAPE over pollutants and time steps; lower values indicate better performance.
• Baselines: Linear regression (Lasso and OLS), WRF-CMAQ, XGBoost, MLP, LSTM, and a spatial-temporal ensemble (STE). Comparisons are made at stations A, B, and C (mutual influence approximated negligible due to >100 km separation).

• Air quality characterization (Station A, Apr 16, 2019–Jul 13, 2021): 358 days (≈44%) were Excellent (AQI≤50); in Good (335 days) the primary pollutants were mainly NO2 and O3; Moderate and Heavy pollution days were driven entirely by O3 as the primary pollutant; no Serious pollution days observed.
• Correlations: Humidity strongly negatively correlates with SO2 (r≈−0.57); O3 is the only pollutant positively correlated with temperature; wind speed, wind direction, and humidity are generally negatively correlated with all pollutants; AQI correlates strongly with O3 (r≈0.90) and also with PM2.5 (≈0.70) and PM10 (≈0.68), indicating O3 is a dominant driver of AQI at Station A.
• Ablation (all variants share a basic LSTM): • WRF only: MAPE≈0.312, RMSE≈32.781 (poorest performance). • Gravity model only: MAPE≈0.269, RMSE≈27.288. • Basic LSTM only: MAPE≈0.261, RMSE≈27.128. • WRF-LSTM only: MAPE≈0.236, RMSE≈23.841. • WRF-LSTM + Gravity (full MTSC-AQP): best with MAPE≈0.228, RMSE≈22.851, demonstrating clear gains from combining mesoscale-informed LSTM with station cooperation.
• Baselines: Across stations A–C, MTSC-AQP outperformed traditional (WRF-CMAQ, LR) and ML/DL baselines (XGBoost, MLP, standalone LSTM, STE) in RMSE and MAPE (Figure 6), with stations B and C showing similar efficiency to A.
• Case analyses: Over Jun 13–Jul 12, 2021, MTSC-AQP predictions closely matched measured concentrations for all six pollutants, notably improving over the WRF-LSTM initial predictions for low-level SO2 and CO. For PM10 at stations A–C, daily averages from MTSC-AQP nearly fit ground truth, including capturing a sudden increase at Station A on Jul 10 and stable ranges around 20–35 µg/m³ otherwise.

The findings show that integrating mesoscale meteorological forecasts with historical pollutant measurements via LSTM enhances temporal feature learning, while the gravity-model-based cooperation among stations contributes essential spatial information, yielding the best overall accuracy. The strong link between AQI and O3 indicates that ozone dynamics are critical for air quality management in the studied region; thus, O3-focused interventions can meaningfully improve AQI outcomes. Correlation results confirm expected physical relationships: higher humidity and wind facilitate pollutant dispersion; temperature elevates O3 levels. Ablations highlight that the gravity model substantially reduces errors by exploiting inter-station information, and that while WRF informs initial predictions, its marginal benefit in the final stage may diminish when rich measured meteorology is available, suggesting redundancy in mesoscale signals within the refined model. Overall, MTSC-AQP effectively addresses prior challenges in meteorological feature integration and inter-station cooperation, translating to superior predictive performance over diverse baselines.

The paper introduces MTSC-AQP, a three-task system that: (1) computes AQI/IAQIs and analyzes pollutant–meteorology relationships; (2) produces initial pollutant forecasts via a WRF-informed LSTM; and (3) refines predictions using multi-station cooperation with a gravity model. On real-world datasets from multiple Chinese monitoring stations, MTSC-AQP achieves lower MAPE and RMSE than traditional and deep learning baselines, and ablation studies verify the contribution of both the WRF-LSTM initial predictions and the gravity-based station cooperation. The approach provides actionable AQI forecasts and pollutant-specific warnings that can support city management and protect vulnerable populations. Future work will explore optimized fusion of heterogeneous temporal features, assess impacts of external human events on air quality, and further mine periodic patterns to enhance prediction robustness.

• Spatial scope and station selection: Experiments focus on three primary stations and their sub-stations within China, with inter-station interactions approximated negligible for sites separated by >100 km. Generalizability to denser networks or other regions may require further validation.
• Gravity model assumptions: Use of Euclidean distance and linear aggregation (fixed or learned coefficients) may oversimplify complex transport and topographical influences; non-linear or direction-dependent effects are not explicitly modeled.
• Data quality and preprocessing: Equipment failures necessitated forward-filling and outlier handling, which, while practical, can introduce bias in short-term dynamics. Results may be sensitive to imputation strategies.
• Meteorological redundancy: The limited incremental effect of WRF in the final stage suggests potential redundancy with measured data; performance under sparse or missing measured meteorology was not explicitly tested.
• Evaluation reporting: While figures indicate strong performance against baselines, detailed per-pollutant numerical comparisons are limited in the text, and the ablation table formatting suggests potential reporting ambiguities that warrant careful interpretation.