AI model may warn Hyderabad before air turns foul

Hyderabad: A city resident may soon know about a pollution spike before stepping out of the house. A study published in ‘Scientific Reports’ has developed a data-driven model that can predict urban air pollution levels using five-year air quality and weather data, offering a potential early warning system for cities battling deteriorating air quality.

The study, 'Deciphering Seasonal Dynamics and Time Series Forecasting of Urban Air Quality: A Case Study in Hyderabad, Vijayawada and Visakhapatnam’, was published in May. It was conducted by Shreyas D, Nishith B, N Neelima, TV Smitha and Vivek Venugopal from Amrita Vishwa Vidyapeetham, and Tolga Ozer from Afyon Kocatepe University, Turkiye.

The study developed and compared two machine-learning approaches to forecast the air quality index (AQI). The direct forecasting method used a deep feedforward neural network with residual blocks to predict the final AQI value in one step. The indirect method first predicted individual pollutants, including PM2.5, PM10 and NO2, using models such as Random Forest, and then calculated the AQI from those estimates.

Bollaram data used for Hyderabad

The researchers used five years of continuous data from Jan 1, 2020, to Dec 31, 2024, obtained from Central Pollution Control Board regulatory-grade monitoring stations. The study covered Hyderabad in Telangana and Vijayawada and Visakhapatnam in Andhra Pradesh. For Hyderabad, data was recorded from Bollaram Industrial Area station.

The dataset included pollutants such as PM2.5, PM10, NO, NO2, NOx, CO, SO2, ozone, NH3, benzene, toluene and xylene, along with meteorological variables such as temperature, humidity, visibility, wind speed and wind direction. Continuous data from 2025 was used to test the models’ real world accuracy.

According to the study, both modelling approaches achieved R2 values exceeding 0.96 for all three cities, indicating that the models could explain more than 96% of the variation in air quality data. For Hyderabad, the direct FNN model achieved an R2 of 0.97 and a Root Mean Square Error of 11.25.

In one validation sample, the model predicted an AQI of 118.45 for Jan 4, 2025, against the actual recorded value of 125.50, an error of 5.6%.

Winter spikes, monsoon relief

The study found that particulate matter, especially PM2.5 and PM10, was the primary driver of AQI across the three cities. It also found a clear seasonal pattern: pollution levels peaked in winter, particularly Dec and Jan and in the post-monsoon months, while air quality improved during the monsoon between July and Sept due to atmospheric cleansing.

Among the three cities, Visakhapatnam recorded the highest average AQI at about 112, followed by Hyderabad at 94 and Vijayawada at 78. The study found that Visakhapatnam’s average pollution was about 19% higher than Hyderabad’s and 43% higher than Vijayawada’s.

For Hyderabad, the study attributed winter pollution spikes to vehicular emissions, road dust and open burning, worsened by meteorological conditions that trap pollutants. As a sprawling industrial and technology hub, the city’s air quality is affected by a combination of emissions and local weather conditions, the study said.

It said such forecasting systems could be used for public health early warnings, air-quality advisories and regulatory responses. It said the models could help officials activate measures such as the graded response action plan, manage traffic or industrial activity before pollution levels worsen, and use long-term data for sustainable urban planning.

The indirect forecasting method, according to the study, could also help policymakers identify the specific pollutant driving a pollution spike, enabling targeted action against sources such as diesel vehicles or industrial plants.