climate-weather/ai produced the result/Environmental Science & Technology 2023 · v3
Machine learning separates weather from policy in China's winter heating air pollution
Researchers estimated how much winter heating, and China's clean heating policies, changed air quality in 189 Chinese cities. Random forest models stripped out the influence of weather, and a statistical method built a comparison world in which heating never happened.
spectrum · one line per step, placed by what the step does · bright lines used AI
Attribution of Air Quality Benefits to Clean Winter Heating Policies in China: Combining Machine Learning with Causal Inference
Environmental Science & Technology, 2023
doi:10.1021/acs.est.2c06800 · record aix-00290 v3 · checked 2026-11-05
- AI was for
- Property prediction, Denoising, Attribution
- Model family
- Random forest, Linear model
- Checked by
- Held-out
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Northern Chinese cities run central heating through the winter, and burning coal and other fuels to supply it releases pollutants into the air. Working out how much of the winter smog is down to heating is harder than it sounds. Pollution levels swing wildly with the weather: a still, cold, windless week traps pollutants near the ground, while rain and wind clear them away. So a winter that looks cleaner than the last may simply have been breezier. Any attempt to judge whether a policy worked has to somehow set the weather aside, and also account for the fact that emissions from traffic, industry and power generation were changing over the same years.
The researchers set out to put a number on two things: how much winter heating raises pollutant levels in Chinese cities, and how much China's policies to make heating cleaner have reduced that effect. They assembled hourly readings of fine particles and several gases for 189 cities from 2014 to 2021, alongside airport weather observations and reanalysis data, which is a reconstruction of past atmospheric conditions. Cities were sorted into groups that receive central heating and a group of southern cities that do not.
Where AI came in
The machine learning part was a weather filter. For each pollutant in each city, a random forest, a model that makes predictions by combining the votes of many simple decision trees, was trained to predict the measured concentration from weather conditions and from time markers such as the hour of day and the date. Seventy per cent of the data was used for training and the remaining thirty per cent for testing. The trained model was then asked to predict concentrations repeatedly with weather conditions randomly drawn from across the whole period, and those predictions were averaged. The result is a "deweathered" series: pollution levels with the effect of day-to-day weather smoothed out.
Those deweathered series were then fed into a statistical technique that builds a synthetic counterfactual, a weighted blend of the southern cities without central heating standing in for what the heated cities would have looked like had heating never started. The gap between the two is the estimate attributed to heating. The researchers note that conclusions drawn from the raw, unfiltered measurements could be completely different, so the headline figures rest on these fitted models. The same pipeline was rerun pretending that control cities had been treated, as a check.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The study estimates how much winter heating and China's clean heating policies changed urban air quality in 189 Chinese cities between 2015 and 2021. Random forest models were fitted to each city's pollutant records to remove meteorological variability, and the resulting deweathered concentrations were fed into a Ridge Augmented Synthetic Control Method that built counterfactual trajectories from southern cities without central heating. Winter heating was estimated to raise annual PM2.5 by 4.6 μg m-3 and SO2 by 2.3 μg m-3 averaged over mainland China, with the heating impact on PM2.5 falling by 5.9 μg m-3 (41.3%) in the "2 + 26" cities against 1.2 μg m-3 (12.9%) in other northern cities. Clean heating policies were associated with a 1.9 μg m-3 reduction in annual PM2.5 across mainland China, which the authors translate into about 23,556 avoided premature deaths in 2021.
How AI was used
Hourly pollutant measurements for 189 cities were paired with airport surface meteorology and ERA5 reanalysis fields, and a random forest was fitted per pollutant per city using the rmweather R package, with meteorological predictors plus Unix time, Gregorian date, day of week, hour of day and Lunar date as time proxies, 300 trees, a minimum terminal node size of 5, and a 70/30 train-test split. Deweathered concentrations were then obtained by averaging 1,000 model predictions in which meteorological variables were randomly resampled from the 2015-2021 period while time variables were excluded, and these were aggregated to daily and weekly series plus daily maximum 8-h O3. The weekly deweathered series for the treatment groups and for 39 nonheating southern control cities were passed to a Ridge Augmented Synthetic Control Method (augsynth R package), which fits a ridge regression outcome model to correct the synthetic control's pretreatment imbalance and was run year by year with the nonheating season as the pretreatment window. The same model pipeline was rerun treating each control city, and separately nine partially heated alternative cities, as if it were a treatment unit, to produce placebo comparisons. Heating-attributable PM2.5 from the model outputs was then carried into a separate, non-learned health burden calculation.
The shape of the work
Structural · the record, drawn
Assemble air quality and meteorological records
Obtaining raw data, whether by measurement, download or retrieval.
were obtained from the Integrated Surface Database (ISD) of NOAA (National Oceanic and Atmospheric Administration) using the “worldmet” R package.where the paper describes this · verbatim
Define treatment and control city groups
Cleaning, filtering, normalising or labelling data already obtained.
Population-weighted average air pollutant concentrations in “2 + 26” cities and “northern” cities are defined as two separate treatment groupswhere the paper describes this · verbatim
Build explanatory variables for weather normalization
Encoding data into features, descriptors, embeddings or graphs.
we used Unix time as a linear trend component, Gregorian date (day of the year) as a seasonal componentwhere the paper describes this · verbatim
Train per-city random forest models
Fitting model parameters, including fine-tuning an existing model.
We use 70% of the original data set to train the model and the remaining 30% to test the model.where the paper describes this · verbatim
Produce deweathered concentrations
Running a trained model over new data to predict, classify or score.
the deweathered concentration at a particular hour is calculated by averaging 1,000 predictions from the meteorological variableswhere the paper describes this · verbatim
Estimate counterfactual concentrations with Ridge ASCM
Running a trained model over new data to predict, classify or score. The AI stood in for simulation.
The Ridge ASCM is used on both observed and deweathered air pollutant concentrations on a year-by-year basis to isolate the effects from winter heatingwhere the paper describes this · verbatim
Placebo tests on control and alternative cities
Testing outputs against ground truth.
We applied the same approach to the 39 control cities one by one.where the paper describes this · verbatim
Translate heating-attributable PM2.5 into health burden
Extracting understanding from model behaviour.
the attributable fraction for premature deaths dropped from ∼1.7% in 2015 to ∼1.5% in 2021where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The reported causal estimates are differences between random-forest deweathered concentrations and ASCM-constructed counterfactuals; the paper states conclusions based on observed rather than deweathered concentrations could be completely different, so the headline numbers rest on the fitted models.
A random forest (RF)-based weather normalization approach was used to decouple the effects of meteorology from observed air pollutant concentrationswhere the paper describes this · verbatim
The RF models for all the studied groups are with low bias (close to 0) and high correlation coefficients (>0.7)where the paper describes this · verbatim
Data and codes for the augmented synthetic control method can be retrieved from the open-access repositorywhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of Random forest weather-normalization model (rmweather R package)Which version of the model was used is not stated.
- Version of Ridge Augmented Synthetic Control Method (augsynth R package)Which version of the model was used is not stated.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
- What step 7 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00290, version 3, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Machine learning separates weather from policy in China's winter heating air pollution. Record aix-00290, version 3, checked 2026-11-05. https://aixsci.org/articles/aix-00290/