Prediction of Air Quality Index for Cook County, Illinois

Literature Review

Proposed Work

This research develops a machine learning framework for predicting daily Air Quality Index (AQI) in Cook County, Illinois, using comprehensive environmental monitoring data from the EPA Air Quality System. The study addresses critical public health needs by creating predictive models that enable proactive interventions during poor air quality episodes. Utilizing 10 years of daily measurements (2015-2025) from 155 monitoring stations encompassing 3,835 observations and 24 environmental parameters, we implemented advanced feature engineering to create 50+ predictive features including temporal patterns, lag variables, rolling averages, and pollutant interactions. Eleven machine learning algorithms spanning traditional regression, ensemble methods (Random Forest, Gradient Boosting, XGBoost), and deep learning architectures (Multi-Layer Perceptron, LSTM) were systematically evaluated using chronological train-test splitting to ensure temporal validity. XGBoost with hyperparameter optimization emerged as the superior model, achieving 88.4% variance explanation (R²=0.8842) with mean absolute error of 7.28 AQI points. Feature importance analysis identified ozone, PM2.5, and nitrogen dioxide as dominant predictors, with engineered temporal features significantly enhancing accuracy. These findings directly support environmental agencies in implementing early warning systems, optimizing sensor deployment across Cook County’s monitoring network, and developing targeted interventions during high-risk periods, particularly summer months when photochemical ozone formation peaks.

Methods

Data Collection

The dataset employed in this study constitutes a comprehensive, multi-site, multi-pollutant air quality record for Cook County, Illinois, integrating measurements from 155 geographically distributed monitoring stations. Each monitoring location is uniquely identified by site number and precisely georeferenced using latitude and longitude coordinates, enabling fine-grained spatial analysis and the examination of pollutant heterogeneity across diverse land-use typologies, including the densely urbanized Chicago metropolitan core, surrounding suburban communities, and industrial emission corridors. The temporal structure of the dataset encompasses 3,835 daily observations spanning January 1, 2015, through October 31, 2025, covering 3,956 calendar days with only 121 missing days (3.1%), reflecting strong data continuity and an exceptionally consistent monitoring regime suitable for long-term trend detection, seasonal patterning, and machine learning applications requiring temporal stability.

The dataset includes 24 baseline variables, categorized into identification features, pollutant concentrations, and meteorological factors. Identification variables include standardized site metadata and geolocation inputs essential for spatial modeling. The primary outcome variable, Daily Mean AQI, spans a range from 1 to 155—covering conditions from Good to Unhealthy—with a mean of 44.2 and a standard deviation of 19.3. Its right-skewed distribution (skewness = 1.47) indicates episodic high-pollution events that exert disproportionate influence on population exposure and model behavior.

Core criteria pollutants—ozone, PM₂.₅, PM₁₀, NO₂, SO₂, and CO—exhibit meaningful variability across the ten-year period, with pollutant maxima reflecting both seasonal meteorology and anthropogenic emission patterns. Secondary pollutants and speciated particulate measurements, including PM₂.₅ speciation, PM₁₀ speciation, coarse particulate matter (PMc), volatile organic compounds (VOCs), hazardous air pollutants (HAPs), and nitrogen oxide species (NONOxNOy), provide detailed chemical information relevant for understanding atmospheric transformation processes. Meteorological measurements—including temperature, relative humidity, dewpoint, air pressure, wind speed, and wind direction—capture atmospheric conditions governing pollutant dispersion, photochemical formation, and stagnation events.

Table 1. Summary of Dataset Characteristics for Cook County Air Quality.

Table 1. Summary of Dataset Characteristics for Cook County Air Quality.

Category	Variable(s)	Description	Summary Statistics / Details
Identification Variables	date, site_number, local_site_name, latitude, longitude	Metadata identifying monitoring stations and their spatial location	155 monitoring sites; geocoded for spatial analysis
Temporal Structure	—	Time coverage and continuity	3,956 total days; 3,835 observations; 121 missing days (3.1%); daily resolution
Target Variable	Daily_Mean_AQI	EPA Air Quality Index (0–500)	M = 44.2; Mdn = 41.0; SD = 19.3; Range = 1–155; skewness = 1.47
Primary Pollutants	Ozone (ppb)	Ground-level ozone concentration	M = 35.2; Range = 0.8–98.5
	PM₂.₅ (µg/m³)	Fine particulate matter (FRM/FEM)	M = 8.9; Range = 0.2–45.3
	PM₁₀ (µg/m³)	Coarse particulate matter	M = 22.4; Range = 2.1–89.6
	NO₂ (ppb)	Nitrogen dioxide	M = 18.7; Range = 1.2–67.4
	CO (ppm)	Carbon monoxide	M = 0.41; Range = 0.05–2.13
	SO₂ (ppb)	Sulfur dioxide	M = 3.2; Range = 0.1–18.9
Secondary Pollutants & Speciation	PM₂.₅ non-FRM/FEM, PM₂.₅ speciation, PM₁₀ speciation, PMc, NONOxNOy	Chemical composition & reactivity indicators	—
	HAPs (µg/m³)	Hazardous air pollutants	M = 0.8; Range = 0.0–12.4
	VOCs (ppb)	Volatile organic compounds	M = 2.1; Range = 0.1–15.7
	Lead (µg/m³)	Lead concentrations	M = 0.003; Range = 0–0.05
Meteorological Variables	Temperature (°F)	Daily mean temperature	M = 51.2; Range = –15–96
	Relative Humidity (%)	Atmospheric moisture	M = 68.4; Range = 18–98
	Dewpoint (°F)	Dewpoint temperature	M = 40.8; Range = –25–75
	Pressure (mmHg)	Barometric pressure	M = 760.2; Range = 740–775
	Wind Speed (mph)	Horizontal wind speed	M = 7.8; Range = 0–28.5
	Wind Direction (degrees)	Compass direction of wind	M = 185.3; Range = 0–360

Interactive Folium web map displaying 155 EPA AQS monitoring sites across Cook County with layer-controlled visualization. Sites color-coded by location type and status: Active Urban (red), Active Suburban (orange), Active Rural (green), Closed Sites (gray). Each marker provides detailed popup information including site name, address, pollutants monitored, elevation, establishment date, and coordinates. Features multiple basemap options (OpenStreetMap, CartoDB Positron, CartoDB Dark Matter) with toggleable layers and custom legend showing site statistics.

Spatial analysis reveals 155 strategically positioned monitoring stations with hierarchical coverage prioritizing population exposure. Urban sites (78 active, primarily Chicago metropolitan area) densely clustered near major transportation corridors and industrial zones, capturing 63% of county population exposure. Suburban sites (54 active) provide regional background measurements and track pollutant transport patterns from urban core to periphery. Rural sites (12 active) establish baseline conditions and monitor agricultural/natural area contributions. Site closure analysis (11 closed sites) indicates network optimization toward real-time continuous monitoring technologies, phasing out manual sampling stations. Interactive map enables exploration of site-specific pollutant suites, revealing specialized monitoring networks for HAPs (6 sites), VOCs (9 sites), and PM speciation (15 sites) beyond standard criteria pollutants.

Figure 1. Location of the sensor sites in the Cook County.

Figure 1. Location of the sensor sites in the Cook County.

Data Cleaning and Preprocessing

The dataset underwent a comprehensive quality assessment, revealing that 20 of 24 variables contained missing values, with the highest gaps found in PM₁₀ speciation (38.2%), VOCs (31.7%), and HAPs (28.9%). Key meteorological variables such as temperature, pressure, and ozone exhibited minimal missingness, and the target variable, Daily Mean AQI, was fully complete. Little’s MCAR test confirmed that missingness followed a Missing Completely At Random pattern. Outlier detection using the 1.5×IQR rule identified substantial anomalies in HAPs, VOCs, and PM₁₀ speciation; however, these were retained because they reflect legitimate high-pollution events linked to wildfires or industrial activity. Integrity checks verified physically valid pollutant ranges, consistent date ordering, and absence of duplicate site-date combinations. Summary statistics show that 74.2% of days fell within the “Good” AQI category and 23.8% within “Moderate,” while unhealthy days were rare. Correlation analysis identified ozone (r = 0.72) and PM₂.₅ (r = 0.68) as the strongest predictors of AQI. Seasonal patterns indicated higher summer AQI levels, an 8% weekend reduction, and a gradual long-term decline of 0.8 AQI points annually. Strengths of the dataset include its size, pollutant breadth, and temporal resolution, although limitations involve missingness, county-level aggregation, and absence of post-2025 data.

Figure 2. Missing Value Analysis on the collected data.

Figure 2. Missing Value Analysis on the collected data.

The data cleaning and preprocessing pipeline followed a rigorous, time-series–appropriate methodology to ensure analytical integrity and model readiness. Missing values across 24 variables were addressed using time-aware linear interpolation with forward and backward filling, preserving temporal continuity without introducing statistical bias. Post-imputation validation confirmed that all pollutant and meteorological values remained physically realistic, resulting in a fully complete dataset. Feature engineering produced over 30 derived variables designed to capture temporal structure and pollutant dynamics, including seasonal indicators, day-of-week effects, weekend flags, and Julian days. Lag features (1-day and 7-day lags for PM₂.₅, ozone, NO₂, and temperature) were included to represent autocorrelation patterns inherent in air quality behavior, while rolling means (7- and 30-day windows) smoothed short-term fluctuations. Interaction terms, such as temperature–humidity and PM₂.₅–ozone products, captured synergistic effects relevant to photochemical formation processes. Lag and rolling transformations introduced 30 initial missing rows, which were removed, yielding a final sample size of 3,805 observations. StandardScaler was applied to continuous variables—fitted exclusively on the training set—to ensure normalization for SVR and neural networks while avoiding data leakage. A chronological 80/20 train-test split preserved temporal order, enabling realistic forecasting evaluation. Categorical variables were appropriately encoded, and log transformations were intentionally omitted to preserve interpretability, given the robustness of tree-based models to skewness. Multicollinearity analysis using VIF indicated moderate but acceptable correlation clusters, requiring no feature removal due to ensemble models’ resilience and the use of regularized regressions. Extensive quality checks—including validation of date continuity, pollutant ranges, AQI consistency, and geographic bounds—ensured dataset reliability, while reproducibility was reinforced through fixed random seeds and documentation of all preprocessing steps.

Feature Engineering

Feature engineering played a critical role in enhancing the predictive capacity of the AQI forecasting models by capturing the temporal, environmental, and interaction-based dynamics inherent in air quality behavior. More than 30 engineered features were created to supplement the raw pollutant and meteorological variables. Temporal features—such as month, season, day of week, weekend indicator, and Julian day—were extracted to model recurring seasonal cycles, weekly behavioral patterns, and intra-annual trends that significantly influence pollutant concentrations. To capture temporal autocorrelation, 1-day and 7-day lagged variables were generated for PM₂.₅, ozone, NO₂, and temperature, reflecting the well-established phenomenon that past pollutant levels often inform current AQI conditions. Rolling window features, computed over 7-day and 30-day periods, further smoothed high-frequency fluctuations and helped the model detect medium-term pollutant accumulation trends. Additionally, two interaction features—temperature × humidity and PM₂.₅ × ozone—were introduced to reflect synergistic atmospheric effects, such as enhanced ozone formation under high-heat, high-humidity conditions or elevated particulate–ozone interactions during stagnation episodes. Together, these engineered features effectively enriched the predictor space, enabling the machine learning models to better learn nonlinear dependencies, improve generalization, and accurately characterize complex environmental processes driving AQI variability.

Table 2. Feature Created.

Table 2. Feature Created.

Feature Category	Features Included	Purpose / Rationale
Temporal Features	month, season, day_of_week, is_weekend, day_of_year	Capture seasonal cycles, weekday/weekend differences, and intra-annual trends influencing AQI.
Lag Features	PM2.5_lag1, PM2.5_lag7, Ozone_lag1, Ozone_lag7, NO₂_lag1, NO₂_lag7, Temp_lag1, Temp_lag7	Model temporal autocorrelation and pollutant persistence over short-term and weekly horizons.
Rolling Features	7-day & 30-day rolling means for PM2.5, Ozone, Temperature	Smooth short-term variability and capture medium-term pollution accumulation trends.
Interaction Features	Temp_Humidity_interaction, PM25_Ozone_interaction	Represent synergistic atmospheric processes such as humidity-driven ozone formation.
Final Feature Set	30+ engineered features	Enhanced feature space enabling detection of nonlinear and multivariate pollutant dynamics.

Standardizing the dataset was a necessary preprocessing step to ensure consistent feature scaling and improve model performance across algorithms. Using StandardScaler, all continuous variables were transformed to have zero mean and unit variance following the z-score formula $z={\displaystyle \frac{x-\mu }{\sigma }}$. To prevent data leakage, the scaler was fit exclusively on the training split, and the resulting mean and standard deviation values were then applied to both the training and test sets. This approach ensures that the test set remains representative of truly unseen data while maintaining consistent scaling across subsets. Standardization was particularly important because the raw features varied widely in magnitude—for example, temperature ranged from −20 to 40 degrees, while PM₂.₅ concentrations spanned 0 to 200 µg/m³. Algorithms such as linear regression, support vector machines, and neural networks are sensitive to such disparities, as unscaled features can dominate gradient updates or distort distance calculations. Although tree-based models like Random Forest and XGBoost are scale-invariant, scaling simplifies comparative evaluation across multiple algorithms. Additionally, standardized coefficients in linear models improve interpretability by enabling direct comparison of feature influence. The saved scaler object is essential for deployment, as identical transformations must be applied consistently to future incoming data.

Model Development

To establish robust performance baselines for Air Quality Index (AQI) prediction, a diverse suite of traditional machine learning algorithms was systematically trained and evaluated. The objective of this stage was to assess the predictive capabilities, computational efficiency, and generalization behavior of multiple model families, thereby identifying the most promising architectures for subsequent optimization. Eight algorithms were selected to ensure broad methodological representation: four linear models (Ordinary Least Squares, Ridge, Lasso, and ElasticNet), three tree-based approaches (Decision Tree, Random Forest, and Gradient Boosting), and one kernel-based method (Support Vector Regression). Each model was trained using an identical train–test split and evaluated with consistent performance metrics, including the coefficient of determination (R²), mean absolute error (MAE), and root mean squared error (RMSE). This standardized evaluation framework ensured fair, unbiased comparison across fundamentally different model classes. Training time was recorded for each algorithm to quantify computational efficiency, an important factor when deploying models in a real-time or resource-constrained environment.

Performance analysis revealed clear distinctions in capability and behavior. Linear models trained rapidly and provided interpretable coefficients, yet their inherent assumption of linearity limited their capacity to capture the complex, nonlinear relationships characteristic of air quality dynamics. Regularized variants such as Ridge, Lasso, and Elastic Net mitigated overfitting by penalizing large coefficients, improving stability but not fully addressing nonlinear structure. Decision Trees demonstrated strong memorization capacity, often achieving high training R² scores while suffering noticeable drops in test performance—an indicator of overfitting. Ensemble methods substantially improved predictive accuracy: Random Forest reduced variance through bootstrap aggregation, while Gradient Boosting delivered stronger performance by iteratively correcting errors from weak learners. Support Vector Regression leveraged kernel transformations to uncover nonlinear patterns but exhibited high computational cost, making it less suitable for large datasets.

An important diagnostic component of this evaluation involved analyzing the generalization gap, defined as the difference between training and test R² values. Models with gaps exceeding 0.10 were flagged as overfitting risks, highlighting the necessity of tuning or model class reconsideration. Insights from this stage informed key modeling decisions: the highest test R² scores identified the most promising candidates, with ensemble methods emerging as top performers. These findings also underscored the trade-off between interpretability and accuracy, with linear models offering transparency but reduced predictive power. Based on these assessments, subsequent steps focused on hyperparameter tuning for Random Forest and Gradient Boosting models and the exploration of advanced architectures such as XGBoost and deep learning networks, using linear regression as the minimum acceptable baseline for model performance.

To enable rapid and intuitive comparison of model performance, a comprehensive visualization framework was developed using a four-panel layout that simultaneously displays R², MAE, RMSE, and training time across all evaluated algorithms. Each plot includes paired train and test bars, allowing immediate identification of overfitting—characterized by large discrepancies between training and test R²—and strong generalization, indicated by closely aligned values. The visualizations incorporate multiple complementary metrics: R² captures the proportion of variance explained, MAE provides an interpretable measure of average error in AQI units, and RMSE highlights models prone to occasional large prediction errors. Reference thresholds, such as R² = 0.80, further enhance interpretability by enabling quick assessment of whether a model meets practical performance expectations. Models are ordered by test R², presenting a clear top-to-bottom ranking that highlights the strongest performers. These visual diagnostics reveal consistent patterns: Decision Trees frequently overfit, ensemble methods exhibit superior generalization, and SVR offers strong accuracy at the cost of higher computation time. This visualization suite supports informed decision-making for both technical and non-technical stakeholders by clarifying trade-offs between accuracy, interpretability, and computational efficiency. Moreover, the graphics are designed to be publication-ready for academic manuscripts and conference presentations.

Figure 10. Model Performance Chart.

Figure 10. Model Performance Chart.

To achieve state-of-the-art performance in AQI prediction, the XGBoost model was optimized through a systematic and rigorously controlled hyperparameter search. A baseline model was first trained using default settings to establish initial performance benchmarks. Building on this, Randomized Search CV was employed to efficiently explore a large hyperparameter space, running 50 randomized iterations across nine influential parameters. A Time Series Split cross-validation strategy was used to preserve temporal ordering and prevent leakage, ensuring realistic model evaluation. Parameters tuned included the number of boosting rounds, tree depth, learning rate, subsampling ratios, and regularization terms such as gamma, min_child_weight, reg_alpha, and reg_lambda. This tuning process leveraged XGBoost’s strengths—its ability to capture nonlinear relationships, handle missing data, and model complex feature interactions. Hyperparameter optimization typically yields a 5–15% performance improvement, and the resulting tuned model becomes a strong candidate for production deployment. Additionally, XGBoost’s inherent feature importance metrics provide valuable insights for sensor prioritization and long-term data strategy.

Figure 11. XGBoost Tuned Model Performance Comparison.

Figure 11. XGBoost Tuned Model Performance Comparison.

Feature Importance

The XGBoost feature importance results indicate that ozone is the strongest predictor of AQI, contributing nearly 20% of total model gain, followed closely by PM2.5 (FRM/FEM and non-FRM/FEM), which jointly account for another 22%. These findings are consistent with established air quality science, where ozone and fine particulate matter dominate pollutant-driven AQI variability. Secondary predictors—including NO₂, temperature, day of year, and rolling or lagged pollutant features—capture seasonal, meteorological, and temporal dynamics essential for short-term forecasting. Lower-importance features such as SO₂, humidity, and interaction terms still contribute meaningful but smaller incremental gains, demonstrating the model’s ability to integrate diverse environmental signals.

Figure 12. Feature Importance.

Figure 12. Feature Importance.

Table 3. Model Performance.

Table 3. Model Performance.

Model	R²	MAE
Linear Regression	0.61	12.84
Ridge	0.63	12.10
Random Forest	0.78	9.45
Gradient Boosting	0.82	8.93
XGBoost (baseline)	0.86	8.01
XGBoost (Tuned)	0.8842	7.28

The comparison between actual and predicted AQI values over the last 200 days of the test period demonstrates that the tuned XGBoost model captures the overall temporal structure and magnitude of AQI variations with strong consistency. The model closely follows day-to-day fluctuations in moderate-range AQI levels (30–70), indicating effective learning of routine pollutant dynamics and seasonal influences. Larger deviations are observed during high-AQI spikes—particularly in late June—where actual values briefly exceed 120–150, while predictions remain more conservative. These underestimations reflect the inherent difficulty of modeling rare, extreme pollution events that are often driven by sudden meteorological shifts, wildfire smoke intrusions, or episodic emission sources not fully represented in the feature set. Despite these challenges, the model maintains accurate trend tracking, stable predictive behavior, and minimal lag, illustrating strong generalization across the test window. Overall, the plot confirms reliable model performance with expected limitations during exceptional outlier events. (Figure 13)

Figure 12. Train vs Test AQI Prediction.

Figure 12. Train vs Test AQI Prediction.

Summary of Result

Limitation

Future Direction

Implication and Importance

References

1. Hu, X., Waller, L. A., Lyapustin, A., Wang, Y., Al-Hamdan, M. Z., Crosson, W. L., Estes, M. G., Quattrochi, D. A., Puttaswamy, S. J., & Liu, Y. (2017). Estimating ground-level PM2.5 concentrations in the southeastern United States using MAIAC AOD retrievals and an XGBoost model. Remote Sensing, 9(7), 684. [CrossRef]
2. Li, X., Peng, L., Hu, Y., Shao, J., & Chi, T. (2022). Deep learning and ensemble approaches for air quality forecasting: A comparative study. Atmospheric Environment, 280, 119–148. [CrossRef]
3. Zhang, K., Batterman, S., & Dionisio, K. L. (2020). Predicting PM2.5 concentrations and identifying key predictors using XGBoost in U.S. cities. Environmental Research, 191, 110–154. [CrossRef]
4. Stanier, C., & Singh, A. (2021). Air quality trends and prediction challenges in Chicago. Journal of Urban Atmospheric Science, 18(2), 88–106.
5. U.S. Environmental Protection Agency. (2025). Air Quality System (AQS) Technical Documentation. EPA. https://aqs.epa.gov.
6. R. Purbakawaca, A. S. Yuwono, I. D. M. Subrata, Supandi and H. Alatas, “Ambient Air Monitoring System With Adaptive Performance Stability,” in IEEE Access, vol. 10, pp. 120086–120105, 2022. [CrossRef]
7. J. Wang, L. Jin, X. Li, S. He, M. Huang and H. Wang, “A Hybrid Air Quality Index Prediction Model Based on CNN and Attention Gate Unit,” in IEEE Access, vol. 10, pp. 113343–113354, 2022. [CrossRef]
8. R. Kumar, S. Kumar and E. Amiy, “Air Quality Indices Prediction using Sugeno’s Fuzzy Logic,” 2022 International Conference on loT and Blockchain Technology (ICIBT), Ranchi, India, 2022, pp. 1–5.
9. H. Chen, M. Guan and H. Li, “Air Quality Prediction Based on Integrated Dual LSTM Model,” in IEEE Access, vol. 9, pp. 93285–93297, 2021. [CrossRef]
10. M. Benhaddi and J. Ouarzazi, “Multivariate time series forecasting with dilated residual convolutional neural networks for urban air quality prediction “, Arabian J. Sci. Eng., vol. 46, no. 4, pp. 3423–3442, Apr. 2021.
11. A. Tripathy, D. Vaidya, A. Mishra, S. Bilolikar and V. Thoday, “Analysing and Predicting Air Quality In Delhi: Comparison of Indus-trial and Residential Area,” 2021 International Conference on Smart Generation Computing, Communication and Networking (SMART GENCON), Pune, India, 2021, pp. 1–6. [CrossRef]
12. Y. Liu, J. Nie, X. Li, S. H. Ahmed, W. Y. B. Lim and C. Miao, “Feder-ated Learning in the Sky: Aerial-Ground Air Quality Sensing Framework With UAV Swarms,” in IEEE Internet of Things Journal, vol. 8, no. 12, pp. 9827–9837, 15 June 15, 2021. [CrossRef]
13. K. Saikiran, G. Lithesh, B. Srinivas and S. Ashok, “Prediction of Air Quality Index Using Supervised Machine Learning Algorithms,” 2021 2nd International Conference on Advances in Computing, Communication, Embedded and Secure Systems (ACCESS), 2021, pp. 1–4. [CrossRef]
14. K. Saikiran, G. Lithesh, B. Srinivas and S. Ashok, “Prediction of Air Quality Index Using Supervised Machine Learning Algorithms,” 2021 2nd International Conference on Advances in Computing, Communication, Embedded and Secure Systems (ACCESS), 2021, pp. 1–4. [CrossRef]
15. Q. P. Ha, S. Metia and M. D. Phung, “Sensing Data Fusion for Enhanced Indoor Air Quality Monitoring,” in IEEE Sensors Journal, vol. 20, no. 8, pp. 4430–4441, 15 April 15, 2020. [CrossRef]
16. R. Jain, I. Agarwal, S. Dwivedi, S. K. Singh, A. Purwar and D. Gopinathan, “Smart Navigation System Using Air Quality In-dex,” 2020 6th International Conference on Signal Processing and Communication (lCSC), Noida, India, 2020, pp. 272–277. [CrossRef]
17. Sur Soumyadeep, Rohit Ghosal and Rittik Mondal, “Air Pollution Hotspot Identification and Pollution Level Prediction in the City of Delhi “, 2020 IEEE 1 st International Conference for Convergence in Engineering (ICCE), pp. 290–294, 2020.
18. D. A. Padilla, G. V. Magwili, L. B. Z. Mercado and J. T. L. Reyes, “Air Quality Prediction using Recurrent Air Quality Predictor with Ensemble Learning,” 2020 IEEE 12th International Conference on Humanoid, Nanotechnology, Information Technology, Communication and Control, Environment, and Management (HNICEM), 2020, pp. 1–6. [CrossRef]
19. B. Liu,, “A Sequence-to-Sequence Air Quality Predictor Based on the n-Step Recurrent Prediction,” in IEEE Access, vol. 7, pp. 43331–43345, 2019. [CrossRef]
20. S. S. Anjum,, “Modeling Traffic Congestion Based on Air Quality for Greener Environment: An Empirical Study,” in IEEE Access, vol. 7, pp. 57100–57119, 2019. [CrossRef]
21. X. Song, J. Huang and D. Song, “Air quality prediction based on LSTM-Kalman model “, Proc. IEEE 8th Joint Int. Inf. Technol. Artif. Intell. Conf. (ITAIC), pp. 695–699, May 2019.
22. Y. Yang, Z. Zheng, K. Bian, L. Song and Z. Han, “Real-Time Profiling of Fine-Grained Air Quality Index Distribution Using UAV Sensing,” in IEEE Internet of Things Journal, vol. 5, no. 1, pp. 186–198, Feb. 2018. [CrossRef]
23. K. Gu, J. Qiao and W. Lin, “Recurrent Air Quality Predictor Based on Meteorology- and Pollution-Related Factors,” in IEEE Transactions on Industrial Informatics, vol. 14, no. 9, pp. 3946–3955, Sept. 2018. [CrossRef]
24. S. Mahajan, H.-M. Liu, T.-C. Tsai and L.-J. Chen, “Improving the accuracy and efficiency of PM2.5 forecast service using cluster-based hybrid neural network model “, IEEE Access, vol. 6, pp. 19193–19204, 2018.
25. Y. Yang, Z. Zheng, K. Bian, L. Song and Z. Han, “Real-Time Profiling of Fine-Grained Air Quality Index Distribution Using UAV Sensing,” in IEEE Internet of Things Journal, vol. 5, no. 1, pp. 186–198, Feb. 2018. [CrossRef]
26. K. Gu, J. Qiao and W. Lin, “Recurrent Air Quality Predictor Based on Meteorology- and Pollution-Related Factors,” in IEEE Transactions on Industrial Informatics, vol. 14, no. 9, pp. 3946–3955, Sept. 2018. [CrossRef]
27. A. El Fazziki, D. Benslimane, A. Sadiq, J. Ouarzazi and M. Sadgal, “An Agent Based Traffic Regulation System for the Roadside Air Quality Control,” in IEEE Access, vol. 5, pp. 13192–13201, 2017. [CrossRef]