TL;DR
This paper evaluates Facebook's Prophet model for forecasting hourly NO2 concentrations across the UK, finding it performs well for diurnal trends and seasonality, especially when incorporating traffic data. The study highlights Prophet's utility for local-scale air quality predictions, particularly during the COVID-19 lockdown, and demonstrates its potential for informing urban planning and traffic management strategies.
What Problem Does This Paper Address?
Accurately forecasting air quality at a local level is crucial for public health and environmental management. Traditional chemical transport models can be computationally intensive and challenging to deploy at high resolutions. This paper addresses the need for simpler, more accessible forecasting methods that can capture local variations in air pollution, especially in response to dynamic events like the COVID-19 lockdown. The authors explore whether a readily available time-series forecasting tool, Facebook's Prophet, can effectively predict NO2 concentrations at single sites, considering factors like traffic and regional influences.
Methods
The study utilizes Facebook's Prophet model (v0.6) for time-series forecasting. Prophet is an additive model that fits non-linear trends with yearly, weekly, and daily periodicity. The researchers applied Prophet to hourly NO2 data from the UK's Automatic Urban and Rural Network (AURN) for the years 2018-2019. They evaluated its performance by comparing predicted values against measured data, focusing on absolute values, diurnal trends, and seasonality.
A key aspect of their methodology involved incorporating mobility and traffic volume data into the Prophet model's fitting process. This was done by fitting Prophet to the ratio of NO2 to traffic volume, allowing them to infer changes in NO2 based on traffic data. They also compared Prophet's performance to the EMEP (European Monitoring and Evaluation Programme) regional model. For a deeper dive into specific events, they analyzed air mass origins using the HYSPLIT model and investigated the impact of different vehicle types (HGVs vs. non-HGVs) on NO2 levels during the COVID-19 lockdown.
Datasets
The study primarily uses data from the Automatic Urban and Rural Network (AURN), which provides hourly measurements of pollutants including NO2, as well as meteorological data (wind speed, direction, temperature) from the UK Met Office.
Traffic data was sourced from Transport for Greater Manchester (TfGM), specifically from their Intelligent Transport Systems (ITS) estate, including Automatic Traffic Count (ATC) sites and Bluetooth Journey Time Passive Sensors. This data provided vehicle flow, speed, and classification.
For regional context and comparison, data from the EMEP (European Monitoring and Evaluation Programme) model (v4.33) was also utilized.
Key Insights and Ideas
- Prophet's Local Forecasting Power: Prophet demonstrates a promising ability to forecast NO2 concentrations at a local level, capturing diurnal trends and seasonality effectively, especially when compared to a regional model like EMEP.
- Traffic Data is Key: Incorporating traffic volume data significantly improves Prophet's ability to predict NO2, particularly during periods of substantial traffic reduction like the COVID-19 lockdown.
- Not All Traffic Data is Equal: While Prophet can capture overall traffic trends, relying solely on historical NO2-per-traffic volume can lead to errors when the nature of traffic changes (e.g., an increase in HGVs relative to other vehicles), as observed in Sharston during the lockdown.
- Regional Influences Matter: Discrepancies between Prophet's forecasts and actual measurements increase for sites with a larger contribution from regional sources, highlighting the need to consider broader atmospheric transport.
- COVID-19 as a Natural Experiment: The lockdown provided a unique opportunity to test forecasting models under extreme conditions, revealing how traffic reduction impacts NO2 and how models respond to these changes.
Key Results
- Prophet showed promising performance in predicting absolute NO2 values, diurnal trends, and seasonality across UK AURN sites for the 2018-2019 period.
- Discrepancies between Prophet forecasts and measurements were generally smaller for Urban Traffic sites compared to Rural Background sites.
- During the COVID-19 lockdown in Manchester, Prophet's forecasts incorporating traffic data captured the significant reduction in NO2 levels well, outperforming the "business as usual" forecast.
- At Manchester Piccadilly, Prophet's forecasts with traffic data closely matched observed diurnal profiles during the lockdown.
- At Sharston, reliance on historical NO2-per-traffic volume led to under-predictions during the lockdown, likely due to a shift in the vehicle fleet composition (increased HGVs).
- The EMEP model, while generally lower than measured values, sometimes captured periods of significant NO2 increases (potentially from regional sources) that Prophet missed.
Limitations and Future Work
- Traffic Data Granularity: The study acknowledges potential errors in inferring traffic volume from Bluetooth Journey Time sensors and the assumption that traffic data from nearby sites is representative of the AURN site's immediate vicinity.
- Vehicle Type Impact: The analysis at Sharston suggests that the relative contribution of different vehicle types (especially HGVs) is crucial and not fully captured by simple traffic volume metrics.
- Regional Source Identification: While Prophet can indicate periods of regional influence, further work is needed to explicitly incorporate regional background contributions into the forecasting model.
- Intervention Impact: Understanding the impact of previous local interventions on air quality could improve error interpretation.
- Ancillary Data: The authors note that incorporating additional data like boundary layer height or ancillary activity data could further enhance forecasting performance.
- Replication: The study suggests replicating this approach in other cities where traffic data is available.
Summary updated