Forecast-Aware Explainable Hybrid Stacking for Forecasting Across Heterogeneous Indian Cities

Air quality is a matter of life and death in many parts of the world, yet the invisible clouds of pollution that hang over our cities are notoriously difficult to predict. The trouble is that the air does not behave the same way everywhere. A smoggy morning in a dense, industrial capital follows different rules than a hazy afternoon in a coastal port or a humid southern metropolis. Furthermore, the factors that drive pollution—traffic patterns, weather shifts, and the sheer volume of vehicles and factories—are constantly shifting, creating a complex puzzle that changes from hour to hour. For decades, scientists have tried to build computer models to forecast these conditions, hoping to give city planners and health officials a heads-up before dangerous levels of fine particles, known as PM2.5, settle over a population. The challenge has always been balancing accuracy with clarity: a model might predict the future perfectly, but if it is a "black box" that no one understands, it is hard to trust or act upon.

This is the problem tackled by a team of researchers at Pondicherry University in India. They set out to build a new kind of forecasting system that does not just guess the future but explains how it reached that guess. Their work focuses on three very different Indian cities: Delhi, a sprawling northern capital known for heavy winter smog; Mumbai, a bustling western port; and Chennai, a southern coastal city. By studying hourly data from 2021 to 2025, the researchers developed a method they call an "Explainable Hybrid Stacking Framework." In simple terms, they did not rely on a single computer program to do the heavy lifting. Instead, they created a team of different algorithms, each with its own way of looking at the data, and taught a master algorithm how to combine their insights. The result is a system that not only predicts the air quality one hour ahead with high accuracy but also reveals exactly which factors—like the previous hour's pollution levels or the current humidity—are driving the prediction for each specific city.

The researchers began by gathering a massive amount of raw data from official monitoring stations in the three cities. This data included the concentration of fine particles, along with a host of other details like the time of day, the day of the week, wind speed, and humidity. To make sense of this raw stream, they first had to teach the computer to see the patterns hidden within the numbers. They created new variables, or "features," that captured the rhythm of the city. They added features that told the computer what time of year it was, features that tracked how pollution had behaved in the last few hours, and features that calculated the average pollution over the last few days. They even created special combinations that looked at how weather and pollution interacted. This process generated hundreds of potential clues, but many of them were repetitive or redundant.

To find the most useful clues, the team employed a rigorous selection process. Instead of trusting just one method to decide which features mattered, they used four different statistical approaches to rank them. They looked at which features had the strongest statistical link to the pollution, which ones a tree-based computer model found most useful, which ones caused the biggest drop in accuracy if they were removed, and which ones contributed most to the final decision. By combining the results of these four methods, they reached a "consensus." This ensured that the final list of features was not biased by the quirks of a single mathematical technique. For Delhi, they narrowed the list down to 173 key features; for Mumbai, 159; and for Chennai, 178. This step was crucial because it stripped away the noise, leaving only the signals that truly mattered for each specific location.

With the best features selected, the researchers built their prediction engine. They did not use a single model, which might be good at seeing some patterns but miss others. Instead, they trained three different types of machine learning models—Random Forest, Extra Trees, and XGBoost—on the data. Think of these three models as three different experts. One might be excellent at spotting long-term trends, another at catching sudden spikes, and the third at understanding complex interactions between variables. Each expert made their own prediction for the next hour's air quality. However, to prevent the system from using data it had not seen during its training, the researchers used a careful technique where each expert was tested on data it had not seen during its training.

The final step was to bring these experts together. A fourth model, called CatBoost, acted as the team leader or "meta-learner." Its job was to listen to the predictions of the three experts and learn how to combine them into a single, superior forecast. This leader learned that sometimes the first expert was right, sometimes the second, and sometimes the best answer was a blend of all three. This "stacking" approach allowed the system to capture a much wider range of patterns than any single model could on its own. The researchers tested their system on data from late 2024 and 2025 that the models had never seen before, ensuring the results were a true test of predictive power.

The results were striking. The new system significantly outperformed both simple guesses and the individual expert models. For Delhi, the system explained 93.83% of the variation in air quality, with an average error of about 30 micrograms per cubic meter. In Mumbai, it explained 88.79% of the variation with an error of roughly 11 micrograms. The most dramatic improvement was seen in Chennai, a city where air quality is notoriously difficult to forecast. While the best single model could only explain about 58% of the variation, the new combined system explained 76.96%, reducing the error from over 22 micrograms to just 16.7. This proved that by combining different learning strategies, the system could handle the unique and complex dynamics of each city much better than a one-size-fits-all approach.

Perhaps just as important as the accuracy was the transparency of the system. Because the researchers used a method called SHAP (which stands for SHapley Additive exPlanations), they could open the "black box" and see exactly what the model was thinking. For Delhi, the system showed that the total pollution load and the concentration of larger particles were the biggest drivers of the forecast. In Mumbai, the current level of larger particles and what they had been doing an hour ago were the most critical factors. In Chennai, the system relied heavily on the air quality index from the recent past and the total pollution load. This level of detail is vital for decision-makers. It means that if a city official sees a forecast for high pollution, they can look at the explanation and know whether to blame a sudden spike in traffic, a change in the wind, or a buildup of industrial emissions.

The study also highlighted that there is no single "best" model for every city. The algorithm that worked best for Delhi was not the best for Mumbai, and neither was the best for Chennai. This finding reinforces the idea that air pollution is a local phenomenon, deeply tied to the specific geography, climate, and human activity of a place. A model that works in one city cannot simply be copied and pasted to another; it must be adapted. The researchers' framework succeeded because it was flexible enough to learn the unique rules of each city while maintaining a consistent, rigorous method for building the forecast.

While the results are promising, the researchers are careful to note the limits of their work. The system was tested on only three cities over a specific five-year period. While it performed well there, it has not yet been proven to work in every city across India or the world. The data used came from official government monitors, which is reliable, but the system did not incorporate other potential sources of information like satellite images or real-time traffic data. The authors suggest that future work should test the system in different climates and with different types of data to see if it can be truly universal. They also emphasize that while the system can predict pollution levels with high accuracy, it is a tool for forecasting, not for identifying the specific source of a pollution event.

Ultimately, this research offers a new blueprint for how we might manage the air we breathe. It moves beyond the old idea of trying to find a single perfect formula for the world's air. Instead, it embraces the complexity of the problem, using a team of diverse digital experts to navigate the unique challenges of each city. By combining high accuracy with clear explanations, the system provides a tool that is not only powerful but also trustworthy. It shows that with the right approach, we can not only see the future of our air quality but understand the forces that shape it, giving cities the knowledge they need to protect their residents before the smog even arrives.