⚡ electrical engineering

Hourly Air Quality Forecasting Using Ensemble Learning: Insights from an Arid Urban Atmosphere

This study develops a machine learning-based hourly air quality forecasting system for Kerman, Iran, utilizing a multi-stage preprocessing pipeline and ensemble tree models to effectively predict six criteria pollutants while identifying key temporal patterns and meteorological drivers in an arid urban environment.

Original authors: Abdullah Kaviani Rad, Mohammad Javad Nematollahi

Published 2026-07-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Abdullah Kaviani Rad, Mohammad Javad Nematollahi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the air around us as a giant, invisible soup. Sometimes, this soup is clear and fresh, but other times, it gets thick with invisible ingredients like car exhaust, dust, and chemicals that can make us sick. Scientists have long tried to predict when this "soup" will get too thick, much like a weather forecaster predicts rain. To do this, they use two main tools: old-school math that tries to follow every single chemical reaction (like tracking every drop of rain in a storm), and a newer, smarter tool called Machine Learning. Think of Machine Learning as a super-smart detective that looks at millions of past clues—like how the wind blew yesterday or how hot it was last Tuesday—to guess what the air will do tomorrow. This is especially tricky in dry, dusty cities where the air behaves differently than in rainy places. If we can predict these dirty air days, we can warn people, especially kids and the elderly, to stay inside and keep their lungs safe.

Now, let's zoom in on a specific detective story happening in Kerman, a city in Iran with a hot, dry climate. A team of researchers decided to build a super-smart forecasting system for six different types of air pollution: Carbon Monoxide (CO), Ozone (O3), Nitrogen Oxides (NO and NO2), Sulfur Dioxide (SO2), and tiny dust particles called PM2.5. They gathered hourly data from 2020 to 2024, but the data was messy, like a puzzle with missing pieces. First, they had to fix the gaps. They tried five different ways to fill in the missing numbers, acting like a detective trying to guess what a missing puzzle piece looked like based on the pieces around it. They found that a method called "Extra Trees" was the best guesser, filling in the blanks with an average accuracy score (R²) of 0.843.

Once the puzzle was complete, they looked at the patterns. They discovered that the air pollution in Kerman has a very strict daily schedule, almost like a clock. Carbon Monoxide (CO) and Nitrogen Oxides (NO, NO2) act like commuters; they spike twice a day, once in the morning rush hour (around 7:00–9:00 AM) and again in the evening (around 8:00–11:00 PM), peaking at about 1.5 ppm for CO. This is because cars are the main culprit. Ozone (O3), however, is a sun-chaser. It doesn't care about traffic; it loves the sun. It builds up slowly in the morning and hits its highest point in the afternoon (around 2:00–4:00 PM), reaching about 38–39 ppb, because sunlight cooks the chemicals together to make it. Meanwhile, the tiny dust particles (PM2.5) act like winter guests, showing up in huge numbers during the cold months, reaching a high of 31.31 µg/m³ in January, likely due to heating and trapped air.

To predict these patterns, the researchers tested seven different "detective" algorithms. They found that the old-school, simple math models (linear models) were often too dumb to catch the complex tricks of the air. Instead, the "ensemble" methods—which are like a team of detectives working together—won every time. For most pollutants, a model called Extra Trees was the champion, correctly predicting CO with a score of 0.713 and PM2.5 with 0.726. Another team, Hist Gradient Boosting, was the best at guessing Sulfur Dioxide (SO2) with a score of 0.669. For Nitrogen Dioxide (NO2), a model named LightGBM took the lead with a score of 0.622.

However, there was one tricky case: Ozone. Even though the models were great at guessing the past, they struggled to predict the future for Ozone, sometimes getting it so wrong that their accuracy score went negative. The researchers suggest this is because Ozone is a chemical magician that depends on a very specific mix of sunlight and heat that the current models didn't fully capture. They also found that the most important clue for predicting almost everything was simply "what happened one hour ago." The air has a memory; if it was dirty an hour ago, it's likely to be dirty now. But for Ozone and dust, the direction the wind was blowing and the amount of sunlight were also critical clues.

In the end, the paper suggests that while we can't perfectly predict every chemical trick the air plays yet, we have built a powerful, cost-effective tool that works well for most pollutants in dry cities. It's a scalable way to keep an eye on the air, helping cities like Kerman give better warnings to protect public health. The researchers admit that to get even better at predicting Ozone, they might need to add more chemical clues to the mix in the future, but for now, this system is a solid step forward for keeping the air clean and the people safe.

Technical Summary: Hourly Air Quality Forecasting Using Ensemble Learning in an Arid Urban Atmosphere

Problem Statement
Air pollution in arid urban environments presents a critical health threat, yet localized, high-resolution forecasting frameworks remain scarce. In Kerman, Iran, a city characterized by a semi-arid to arid climate, rapid urbanization, and industrial growth, air quality frequently exceeds national standards and WHO guidelines. The unique meteorological dynamics of arid regions—specifically infrequent rainfall limiting wet deposition and the amplification of temperature inversions—create complex atmospheric conditions where pollutants persist and accumulate. While traditional chemical transport models offer mechanistic insights, they are often computationally intensive and sensitive to emission inventory uncertainties. Conversely, data-driven machine learning (ML) approaches have shown promise but require validation within specific arid contexts to address the non-linear relationships between meteorological variables, temporal cycles, and pollutant levels. This study addresses the research gap in developing a comprehensive, localized predictive modeling framework for Kerman's air quality.

Methodology
The study developed a machine learning-based predictive system for hourly concentrations of six criteria pollutants: Carbon Monoxide (CO), Ozone (O3), Nitrogen Monoxide (NO), Nitrogen Dioxide (NO2), Sulfur Dioxide (SO2), and Particulate Matter (PM2.5). The research utilized data from June 15, 2020, to March 19, 2024, integrating hourly pollutant concentrations from a city monitoring station with reconstructed hourly meteorological data from the Kerman synoptic station.

The methodology followed a rigorous multi-stage pipeline:

Preprocessing and Imputation: The dataset underwent temporal synchronization, converting Persian dates to Gregorian formats and resampling to a precise hourly interval. To address missing values without introducing statistical bias, five imputation methods (Random Forest, Extra Trees, Bayesian Ridge, K-Nearest Neighbors, and Linear Regression) were benchmarked. The Extra Trees algorithm was selected as the optimal imputer based on the highest average $R^2$ (0.843) and lowest RMSE.
Feature Engineering: The pipeline explicitly encoded temporal context by extracting time-based features (hour, day of week, month, etc.). Crucially, lagged variables (t-1, t-2, t-3, t-24) were generated for both pollutant and meteorological variables to capture temporal autocorrelation and atmospheric inertia. Interaction features were also created.
Modeling and Evaluation: Seven regression models were evaluated: Ridge Regression, Extra Trees, Bayesian Ridge, Hist Gradient Boosting, Decision Tree, LightGBM, and Random Forest. The data was split chronologically into training and hold-out test sets to ensure unbiased generalization assessment. Performance was measured using $R^2$ , RMSE, and MAE.

Key Results

Temporal Patterns: Distinct diurnal and seasonal patterns were identified. CO and NOx exhibited bimodal rush-hour peaks, while O3 showed a photochemically driven afternoon maximum. PM2.5 concentrations peaked in winter (mean: 31.31 µg/m³ in January), attributed to heating emissions and stable atmospheric conditions.
Model Performance: Ensemble tree-based methods consistently outperformed linear models for most pollutants.
- CO, NO, and PM2.5: Extra Trees achieved the highest test $R^2$ values (0.713 for CO, 0.320 for NO, and 0.726 for PM2.5).
- SO2: Hist Gradient Boosting performed best ( $R^2$ = 0.669).
- NO2: LightGBM yielded the best results ( $R^2$ = 0.622).
- O3: All models demonstrated poor generalization on the test set, resulting in negative $R^2$ values. This indicates that despite capturing training patterns, the models failed to predict unseen O3 data, likely due to the complex, non-linear photochemical interactions not fully captured by the available features.
Feature Importance: Lagged pollutant concentrations (specifically at t-1) were the dominant predictors for all pollutants, confirming strong temporal autocorrelation. For O3, photochemical variables (solar radiation and temperature) and their interactions were critical. For PM2.5, wind direction emerged as a secondary driver, suggesting the influence of regional transport and dust events.

Significance and Claims
The authors claim that this study offers a scalable, cost-effective, and replicable framework for operational air quality forecasting in arid cities. By successfully integrating advanced imputation and ensemble learning, the framework provides a foundation for public health advisories and environmental management in Kerman and similar climatic zones.

The paper modestly acknowledges limitations, noting that the poor generalization for O3 highlights the need for incorporating specific photochemical indicators (such as VOC levels or NO2/NO ratios) which were not available in the dataset. Furthermore, the reliance on a single meteorological station limits the representation of spatial heterogeneity. Consequently, the authors do not claim the model is a perfect solution for all pollutants but rather a robust starting point that effectively captures the dynamics of primary pollutants (CO, NOx, PM2.5, SO2) while identifying specific areas for future refinement in photochemical modeling. The study concludes that ensemble methods are superior to linear approaches for capturing the complex, non-linear interactions governing air pollutant dynamics in arid urban environments.

More like this