AI Learns From History to Predict Ozone Pollution Across Chinese Cities

Ground-level ozone is one of the most stubborn air pollution problems facing the world’s megacities, and forecasting it accurately has long frustrated scientists. Unlike particulate matter, ozone is not emitted directly but forms in complex photochemical reactions involving nitrogen oxides, volatile compounds, and sunlight, all modulated by weather patterns that shift from hour to hour and year to year. Now, a team of researchers in China has unveiled a new artificial intelligence framework that takes an unusual approach: instead of relying solely on a neural network’s learned parameters, it actively retrieves historical episodes of pollution from a curated knowledge base and uses them to correct its forecasts. The system, called OzoneKBNet, is described in the journal Air Quality, Atmosphere & Health.

The research, led by Mingkun Zhu of the Department of Automation at Taiyuan University of Technology, together with Xiaoxia Han, Jinde Wu, Haonan Zhu, and Wenxin Chai, targets a problem that most data-driven forecasting models handle poorly. Conventional deep learning models encode historical patterns implicitly within their trained weights, which means that once training is complete, the model has no explicit memory of specific pollution episodes it can consult. If a city experiences a recurring type of ozone event, driven for example by a particular combination of stagnant air, high temperatures, and regional transport of precursors, a standard model must re-learn that pattern from scratch each time it appears in slightly different form. OzoneKBNet instead makes those recurring episodes searchable.

The framework works by forecasting station-level ozone concentrations over the next 48 hours based on the preceding 96 hours of observations. Its inputs include ozone itself, nitrogen dioxide, fine particulate matter with a diameter of 2.5 micrometers or less, temperature, and relative humidity, a set of variables chosen to capture both the chemical precursors of ozone formation and the meteorological conditions that govern it. When a forecast is needed, the system searches a knowledge base of historical situations for multi-scale analogs drawn from across the monitoring network, retrieving similar episodes not just from the station in question but from neighboring stations as well. This cross-station retrieval is crucial in urban agglomerations, where pollution transported from upwind cities can dominate local ozone behavior.

Retrieval alone is not enough, and the researchers built in a sophisticated filtering stage. Candidate analogs are re-ranked according to two criteria: representation similarity, which measures how closely the retrieved historical situation matches the current one in the model’s learned feature space, and local-trend consistency, which checks whether the historical episode’s near-term trajectory actually resembles what is happening now. Only analogs that pass both tests contribute to the forecast. Their known future trajectories are then used as bounded, reliability-aware corrections to a lightweight long short-term memory network, or LSTM, that provides the baseline prediction. The word bounded matters: rather than allowing retrieved evidence to override the neural forecast wholesale, the system applies corrections within limits calibrated to how trustworthy the retrieved analogs appear to be.

A particularly careful aspect of the study is its handling of temporal data leakage, a common pitfall in machine learning applications to time series. If a model is trained and tested on data drawn from overlapping periods, it can inadvertently peek at information from the future, inflating its apparent accuracy. The team partitioned their data strictly by year: observations from 2023 were used for normalization, for pretraining the retrieval encoder, and for constructing the knowledge base; data from 2024 were reserved for supervised training; and a fixed set of samples from 2025 served as the test set. This design means the model was evaluated on genuinely unseen future conditions, a far more demanding test than random splitting of the data.

The results span seven Chinese urban agglomerations, offering a broad regional assessment of the framework’s capabilities. Across all stations, OzoneKBNet achieved macro-average mean absolute error of 20.06 micrograms per cubic meter and root mean square error of 24.45 micrograms per cubic meter over the 48-hour forecast horizon. It posted the lowest regional mean absolute error in six of the seven regions evaluated, suggesting that the retrieval-based approach generalizes well across geographically and climatologically diverse areas rather than being tuned to a single city’s quirks. Compared with a station-specific LSTM baseline, the framework reduced mean absolute error by 1.6 percent and root mean square error by 1.2 percent.

Those percentage improvements may sound modest, but their significance lies in what they demonstrate about the architecture of forecasting systems. The baseline LSTM is already a strong performer, trained on the same data, and squeezing additional accuracy out of it through any independent mechanism is difficult. The gains from OzoneKBNet come from an entirely complementary source: explicit regional historical evidence that the parametric model cannot access on its own. In effect, the study shows that a model’s learned parameters and a searchable library of past episodes contain partially non-overlapping information, and combining the two yields forecasts more accurate than either alone. This principle echoes a broader trend in artificial intelligence, where retrieval-augmented generation has transformed language models by letting them consult external documents rather than relying purely on memorized knowledge.

The intellectual lineage of the analog approach stretches back decades in atmospheric science. Long-range weather forecasting based on analogs, matching current conditions to historical situations and borrowing their outcomes, was explored as early as the late 1980s, and modified k-nearest-neighbor methods have been used for real-time weather prediction. What is new is the marriage of this old idea with modern deep learning: an encoder trained to map multivariate pollution and meteorology sequences into a representation space where similarity search becomes meaningful, combined with a re-ranking scheme and a residual correction mechanism that treats retrieved evidence with appropriate caution. The framework also connects to recent work on retrieval-augmented time series forecasting in the machine learning community, where researchers have begun applying similar ideas to general forecasting foundation models.

Why does station-level forecasting matter so much for ozone? Regional averages can hide enormous local variation. Ozone concentrations at a single monitoring station depend on the interplay of local emissions, photochemistry, boundary-layer dynamics, and transport from surrounding areas, and public health warnings are ultimately issued for specific places where people live and breathe. A framework that can deliver accurate 48-hour forecasts at the station level, across an entire urban agglomeration, gives authorities the granularity needed to warn residents before pollution episodes peak, to time traffic restrictions or industrial curbs more effectively, and to study how ozone responds to control measures. The authors suggest their approach may support short-term ozone-pollution episode warning, and the 48-hour horizon is well matched to the timescale on which such interventions operate.

The team has also made an unusually strong commitment to openness. The source code of OzoneKBNet and the complete fixed 2025 test set are available in a public repository on GitHub, along with the model implementation, running scripts, data-format instructions, and fixed test samples that allow others to inspect the experimental pipeline and evaluation protocol. The raw air quality monitoring data and ERA5 reanalysis data analyzed in the study come from publicly available providers, although redistribution restrictions prevent the processed training datasets from being shared directly. This transparency matters in a field where evaluation practices vary widely and where, as the strict 2023-2024-2025 split shows, the details of how models are tested can determine whether reported accuracy reflects real forecasting skill or subtle leakage. For a problem as consequential as urban ozone, reproducible methods and honest evaluation are not luxuries but necessities, and this work offers both alongside a genuinely new architectural idea.

Subject of Research: Retrieval-augmented deep learning for station-level ground-level ozone forecasting in urban agglomerations

Article Title: A retrieval-augmented historical knowledge-base framework with residual correction for station-level ozone forecasting across urban agglomerations

Article References: Zhu, M., Han, X., Wu, J., Zhu, H., & Chai, W. (2026). A retrieval-augmented historical knowledge-base framework with residual correction for station-level ozone forecasting across urban agglomerations. Air Quality, Atmosphere & Health, 19(10), Article 215. https://doi.org/10.1007/s11869-026-02105-2

Image Credits: AI Generated

DOI: 10.1007/s11869-026-02105-2

Keywords: ozone forecasting, air quality, machine learning, retrieval-augmented forecasting, LSTM, urban agglomerations, knowledge base, time series, China, pollution warning, deep learning, meteorology