A Novel AI Framework for PM Pollution Prediction Applied to a Greek Port City

Fotios K. Anagnostopoulos, Spyros Rigas, Michalis Papachristou, Ioannis Chaniotis, Ιoannis Anastasiou, Christos Tryfonopoulos, Paraskevi RaftopoulouAtmosphere 2023 · DOI

TL;DR

This paper introduces a novel AI framework for predicting particulate matter (PM) pollution, leveraging low-cost sensor data and interpretable deep learning. By combining temporal modeling with Long Short-Term Memory (LSTM) networks and spatial modeling using urban characteristics, the framework offers accurate forecasts and insights into pollution dynamics, making it a valuable tool for environmental policy and public health.

What Problem Does This Paper Address?

Particulate matter (PM) pollution is a significant global health concern, impacting respiratory and cardiovascular systems and even correlating with infectious disease spread. Effectively managing and mitigating PM pollution requires accurate forecasting systems. Traditional methods like transport and computational fluid dynamics (CFD) models are often computationally expensive, require detailed input data, and struggle with generalization across different urban environments. This paper addresses the need for a more accessible, scalable, and interpretable PM forecasting solution, particularly in complex urban settings with diverse pollution sources.

Methods

The core of this research is a novel AI framework built upon Interpretable Multi-Variable LSTM (IMV-LSTM) networks. This approach aims to balance high prediction accuracy with interpretability, addressing the "black box" criticism often leveled at deep learning models.

Key methodological aspects include:

  • Temporal Modeling: An IMV-LSTM network is employed to capture the temporal variability of PM concentrations. This architecture is designed to retain long-term dependencies in the data.
  • Spatial Modeling: The framework incorporates urban characteristics to model spatial dependence. Specifically, it uses Mean Population Density (MPD) and Mean Floor Area Ratio (MFR) as features. These metrics help capture the influence of urban structure on pollution distribution.
  • Feature Engineering: Beyond standard meteorological data (temperature, humidity, wind, pressure), the study engineers temporal features (sine/cosine transformations of hour, day, month) and spatial features (MPD, MFR). A novel feature, wind speed variance, is also introduced to capture turbulent flow dynamics.
  • Sensor Agnosticism: The framework is designed to be sensor-agnostic, meaning it can accommodate data from various types of sensors, including portable ones.
  • Data Pre-processing: A crucial step involves pre-processing data from dual-sensor PurpleAir devices, including a consistency check between sensor channels and a linear fit to derive final PM measurements. A calibration procedure is also applied.
  • Interpretability: The IMV-LSTM's attention mechanism allows for feature importance assessment, providing insights into which factors most influence PM predictions.

Datasets

The study utilizes publicly available data from PurpleAir PA-II sensors for PM2.5 concentrations. This dataset spans from December 1, 2018, to June 19, 2022. Meteorological data is sourced from WeatherUnderground meteorological stations. The authors highlight the use of open-source tools and data, contributing to the cost-efficiency and accessibility of their framework.

Key Insights and Ideas

  • Urban Structure Matters: The inclusion of urban characteristics like Mean Population Density and Mean Floor Area Ratio significantly contributes to modeling spatial PM pollution, demonstrating that urban form is a key determinant of air quality.
  • Interpretability Enhances Trust: By using an interpretable LSTM variant, the researchers can not only predict PM levels but also understand why certain predictions are made, which is crucial for building trust and enabling informed decision-making.
  • Low-Cost Sensors are Viable: The framework's success with data from low-cost PurpleAir sensors underscores their potential for widespread air quality monitoring and research, democratizing access to environmental data.
  • Temporal Dynamics are Complex: The importance of features like "month" and "day" for daily predictions, and "hour" for hourly predictions, highlights the significant influence of seasonal patterns, human activity cycles, and short-term emission variations on PM concentrations.
  • Wind Dynamics are Crucial: The introduction and significance of "wind speed variance" suggest that turbulent wind patterns, not just average speed or direction, play a vital role in PM dispersion and concentration.

Key Results

The proposed framework demonstrates strong performance in predicting PM2.5 concentrations in Patras, Greece.

  • Accuracy: The model achieves forecasting accuracy comparable to the resolution of the sensors. For daily predictions with a 24-day look-back window, the RMSE was as low as 12.1 ± 1.7 µg/m³. Hourly predictions achieved an RMSE of approximately 20 µg/m³ and an MAE of ~3 µg/m³.
  • Feature Importance: For daily predictions, "month" and "day" were found to be the most important features, followed by "Auto-regressive" and meteorological variables like "RH" and "tempAvg". For hourly predictions, "month," "Auto-regressive," and "windDir" were prominent, with "hour" also gaining importance.
  • Comparison: The IMV-LSTM framework showed comparable predictive power between its "full" and "tensor" implementations, with the "tensor" version slightly outperforming in RMSE. The MAE for hourly and daily forecasts was notably lower than some previous studies, suggesting improved predictive precision for longer forecasting spans and finer temporal resolutions.

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • MFR Limitation: The Mean Floor Ratio does not account for spaces between buildings, which could be addressed by incorporating features like road density or free space percentage.
  • MPD Data Granularity: Population density data is based on decadal surveys, which might not capture recent fluctuations.
  • Point Predictions: The current model provides point predictions, requiring past measurements at the specific location.
  • Sensor Network Dependency: The framework currently relies on the availability of a sensor network for both PM and meteorological data.
  • Future Work:
    • Transfer Learning: Exploring transfer learning approaches to reduce the dependency on local sensor networks.
  • Additional Data Sources: Integrating data from social networks and daily news to capture out-of-distribution events and human activity fluctuations.
  • Hybrid Models: Combining the AI framework with transport models (e.g., CAMx) to address large-scale phenomena like dust transport.
  • Broader Application: Applying the framework to diverse urban contexts and spatial scales to assess its generalization ability.
  • Real-time Data Integration: Leveraging real-time data for more dynamic pollution management strategies.

Summary updated

Abstract

Particulate matter (PM) pollution is a major global concern due to its negative impact on human health. To effectively address this issue, it is crucial to have a reliable and efficient forecasting system. In this study, we propose a framework for predicting particulate matter concentrations by utilizing publicly available data from low-cost sensors and deep learning. We model the temporal variability through a novel Long Short-Term Memory Neural Network that offers a level of interpretability. The spatial dependence of particulate matter pollution in urban areas is modeled by incorporating characteristics of the urban agglomeration, namely, mean population density and mean floor area ratio. Our approach is general and scalable, as it can be applied to any type of sensor. Moreover, our framework allows for portable sensors, either mounted on vehicles or used by people. We demonstrate its effectiveness through a case study in Greece, where dense urban environments combined with low cost sensor networks is a peculiarity. Specifically, we consider Patras, a Greek port city, where the net PM pollution comes from a variety of sources, including traffic, port activity and domestic heating. Our model achieves a forecasting accuracy comparable to the resolution of the sensors and provides meaningful insights into the results.
BibTeX
@inproceedings{anagnostopoulos2023novel,
  author = {Fotios K. Anagnostopoulos and Spyros Rigas and Michalis Papachristou and Ioannis Chaniotis and Ιoannis Anastasiou and Christos Tryfonopoulos and Paraskevi Raftopoulou},
  title = {{A Novel AI Framework for PM Pollution Prediction Applied to a Greek Port City}},
  booktitle = {Atmosphere},
  year = {2023},
  doi = {10.3390/atmos14091413},
}
Figure