Short-term peaks are challenging to forecast given how quickly they can arise from rapidly changing local conditions. It could be just a few minutes’ congestion, the wind shifting, a line of vehicles stopping at a junction, or a lack of ventilation in a street canyon which can cause a momentary increase in exposure.
AI has a role to play in detecting and predicting certain short-term peaks, particularly if data are available from high-frequency monitoring networks. However, forecasting these peaks involves more than applying machine learning to an air quality dataset.
It is as much a physical, as a statistical, problem. A short-term pollution peak is not simply a pattern in the data. It is the result of emissions, atmospheric movement, street shape, turbulence and human activity all coming together in a certain place and at a certain time.
Why Forecasting Short-Term Pollution Peaks Is Difficult
Describing long-term air quality trends is simpler than predicting short localised events. Daily and annual averages average out the variation in concentration that can occur within individual streets. Short peaks do the reverse: they reveal the complexity.
The reason for the peak might be traffic queues at a road junction. It might be caused by a lack of wind dilution. It might be because a street canyon retains polluted air for a while. It might be because a pollution plume shifts from one road onto another as wind direction changes.
This is why peak exposure in urban air quality policy deserves its own evidence base. Averages have a role to play, but they do not always capture the reality on the ground when people wait, walk or cycle around a particular street.
AI can help to predict such events if it can also understand the conditions that lead to them.
How AI Can Help to Detect Short-Term Pollution Peaks in High-Frequency Data
AI and machine learning can help us spot patterns in data that are not obvious to human observers. Sensors generating high-frequency data record thousands of individual events over time and across space. Some of these events are random. Some are repeating patterns of condition that give rise to short-term peaks.
AI can help to find patterns including:
- pollution peaks tied to individual wind directions;
- short-term peaks associated with traffic queues or traffic signal timing;
- repeating concentration peaks at certain junctions or crossings;
- differences between pollution events detected by sensors that are local, and those that are wider spread;
- changes in concentration levels that flow through the monitoring network, as air mass moves along.
This has the potential to make urban pollution monitoring networks more valuable, especially in situations where the question is not whether pollution is high, but when and why it is high, and for only a short duration.
Prediction Does Not Equal Explanation
An AI model might discover that short-term pollution peaks occur under particular circumstances. Perhaps a particular monitoring point finds higher concentrations when wind speed drops and traffic volume is high.
That does not necessarily give an explanation for the peak.
The cause could be local recirculation of emissions, pollution plumes moving from another street, a nearby polluting source, or poor air mixing and ventilation. In fact, it might be a mix of all of these.
Explaining the peak is important to understanding policy response. If the problem is caused by recirculation in an urban street canyon due to the buildings, no amount of traffic management will solve it. It may be necessary to change the position of the monitor if it has been placed in a location where the pollution is not actually being measured in the most relevant way.
This is another common problem in AI-assisted urban dispersion modelling: machine learning can find relationships, but those relationships still have to be interpreted.
Short Peaks and Pollution Hotspots
Short-term peaks are likely to be found where they are also of policy interest, such as busy junctions, bus stops, the entrance to schools, narrow pavements, cycle lanes and street canyons.
But it does not follow that they will be at the highest emission location, because the air can take the pollution away from its source. Airflow can carry plumes down side streets, around the face of a building, or to a junction if conditions are right.
This is where AI-based hotspot detection can help with short peak analysis. A model can identify where a short peak may occur repeatedly, to suggest the need for additional monitoring or other investigation.
However, there is a risk that a hotspot map is over-confident in its claims. Any prediction of a short-term peak is not evidence but should be a candidate for evidence.
The Role of Sensor Networks
Short-term peaks are difficult to characterise using sparse monitoring. A fixed monitor may not pick up a short event in close proximity or may measure a short peak without knowing where it originated.
A sensor network might help as concentrations in one sensor location may be compared to others. If sensor locations are close to each other and the recording frequency is high enough, a short-time concentration plume might be measured travelling along the street network or a short-term peak may show in repeated conditions.
In this case, the model can assist with interpretation, such as comparing how the sensors read over time and across space. Is the short peak isolated? Is it in transit? Does it repeat? Does it vary with meteorology?
However, sensor networks must be carefully designed. Poor calibration, weak siting or different response times will limit the confidence of short-term prediction. Machine learning can provide a quick analysis of the sensor data but it cannot remove the inherent uncertainty from weak sensor measurements.
Using Field Evidence to Anchor AI Predictions
Field evidence is still important as AI models need examples of the conditions that they will predict, including variation in traffic, meteorology, local buildings, background concentrations and measured concentrations.
The process of building urban pollution evidence from field campaigns is important for AI as a learning process. Without field evidence, the AI may be looking for patterns in an incomplete or biased dataset.
This is especially the case for short peaks as rare events may not show up in a regular data set. An AI model may be trained using mostly normal conditions, so it will predict those conditions well but not the conditions where it is needed most for policy.
Why Controlled Experiments Are Still Needed
Short peaks are hard to study in a field experiment as the same conditions rarely repeat. Wind direction, traffic conditions and background air pollution concentrations will always change.
Controlled experiments can identify the physical processes involved. Wind tunnel evidence can be used to determine the relationship between the position of a source, the dimensions and layout of the street and wind direction with short-term concentration build-up or dispersal.
This is relevant to AI: controlled data can tell us if a pattern being predicted has a physical cause. When a machine learning model predicts repeated peaks close to a corner of a building, controlled data could help us to know if recirculation or plume channelling is causing those peaks.
That is, when AI is linked to this process evidence, it becomes a useful tool.
Modelling Peaks Without Creating a False Sense of Confidence
Short-term prediction might create a false sense of confidence. While a model may output a high-resolution estimate for a specific location and time period, the associated uncertainty may still be very high.
The uncertainty is important here: if a model is used to inform traffic management, space use planning, or emergency response plans, users want to know that a model can be relied upon to represent short-term peaks at the relevant spatial or temporal scale.
This is why model validation for policy is key when applying AI to this kind of problem. Models should not just be validated against how well they capture an overall average, but instead be validated against the kind of event they are actually trying to predict.
If the policy problem is that short exposure peaks are a problem, the validation needs to focus on those short time periods, that kind of spatial variation, and that kind of peak behaviour.
Emergency Response and Rapid Prediction
Short-term prediction has the most potential utility when thinking about emergency response. One example is how AI could bring together a range of data types, including sensor measurements, meteorological observations and dispersion model output, to quickly produce an estimate of where material is going.
However, urban airborne release planning relies upon fine-scale flow processes down streets, past junctions, around building edges, and into building canyons.
While a rapid estimate may be useful here, we still need to understand what its limitations are. There is a risk here: a missed pathway and a missed peak has real implications for the prioritisation of monitoring and how response efforts are deployed.
Therefore, AI should be supporting tools used for rapid emergency dispersion scenarios, rather than replacing validated models and interpretation from experts.
So Where Does AI Have Utility?
The short answer is that AI has utility as part of the short-term peak prediction workflow when used as a bounded tool.
It could be used to pick up early signals within a high-frequency measurement. It could be used to recognise conditions associated with recurring peaks. It could be used to assist the interpretation of a sensor network. It could be used to explore many potential contributing causes simultaneously. And it could be used to help identify locations where a peak may be occurring but more investigation is required.
AI is not a tool that should be treated as a replacement tool. A machine learning model that is used to estimate a peak, when the data is unvalidated, when the cause of the peak is not well understood and when there is no validation of peak estimates, just introduces uncertainty into the process.
Short-term pollution peaks are physical events with policy impacts. They require evidence to understand them: from monitoring, fieldwork experiments, controlled data sets and modelling. AI could support the curation and interpretation of that evidence, help to reveal patterns and allow faster investigation.
But it cannot be substituted for understanding why peaks occur, who is being exposed or how accurate the model prediction is.
So, does AI have a role in short-term peak prediction? Yes, but in practice it can do so responsibly only as part of a wider evidence system, in which data quality, physical understanding and validation are central.
