Estimated contributions and uncertainties of PCA/MLR–CMB results: Source apportionment for synthetic and ambient datasets

Guoliang Shi, Fang Zeng, Xiang Li, Yinchang Feng, Yuqiu Wang, Guangxun Liu, Tan ZhuAtmospheric 2011 · DOI

TL;DR

This paper introduces and validates a two-stage PCA/MLR-CMB model designed to improve the accuracy and reduce uncertainties in source apportionment of air pollutants, particularly when dealing with complex mixtures of sources. The model effectively handles the "collinearity problem" where multiple sources have similar chemical signatures, a common issue in urban air quality studies.

What Problem Does This Paper Address?

Identifying the sources of particulate matter (PM) in urban environments is crucial for effective air quality management. Receptor models, such as the Chemical Mass Balance (CMB) model, are widely used for this purpose. However, traditional CMB models face challenges when multiple sources have similar chemical profiles (the "collinearity problem"). This can lead to unstable results, negative source contributions, or the inability to distinguish between similar sources. Furthermore, a large number of unknown or minor sources can also contribute to this instability by acting as "noise" that disrupts the mass balance. This paper addresses these limitations by proposing and evaluating a combined PCA/MLR-CMB model that aims to reduce noise and better handle collinearity.

Methods

The core of this paper is the development and evaluation of a two-stage PCA/MLR-CMB model.

  • Stage 1: Noise Reduction and Factor Identification (PCA/MLR)
    • Principal Component Analysis (PCA) is used to reduce the dimensionality of the receptor data and identify potential "factors" representing source categories. This stage aims to filter out "noise" from the original receptor data.
  • Multiple Linear Regression (MLR) is then applied to these factors to estimate their contributions.
  • The output of this stage is a "secondary receptor" dataset, which is essentially the original receptor data with the noise reduced and potentially grouped sources.
  • Stage 2: Source Apportionment (CMB)
    • The "secondary receptor" from Stage 1 is then analyzed using the Chemical Mass Balance (CMB) model.
  • This stage benefits from the reduced noise and potentially better-defined source profiles, leading to more stable and accurate source contribution estimates.

The paper also discusses the use of Condition Indexes (CI) and Variance-Decomposition Propositions (VDP) to quantitatively assess collinearity in source profiles.

Datasets

The study utilizes two types of datasets:

  • Synthetic Datasets:
    • Three synthetic receptor datasets were generated to rigorously test the PCA/MLR-CMB model's performance under controlled conditions, specifically with a "serious collinearity problem."
    • These datasets were created using actual PM10 source profiles from three Chinese cities (Taiyuan, Anyang, and Tianjin), representing seven source categories (URD, soil dust, coal combustion, cement dust, vehicle exhaust, secondary sulfate, and secondary nitrate). Noise was intentionally added to simulate real-world conditions.
  • Ambient Dataset:
    • Ambient PM10 concentration data from Chengdu, China, collected in 2007 (135 samples).
  • Source samples for five actual source categories were also collected in Chengdu (soil dust, coal combustion, cement dust, vehicle exhaust, and URD).

Key Insights and Ideas

  • PCA as a Noise Filter: The PCA stage effectively reduces the impact of unknown or minor sources ("noise") on the subsequent CMB analysis, leading to more robust results.
  • Handling Collinearity: The two-stage approach significantly mitigates the collinearity problem. By treating complex factors from PCA as "secondary receptors," the CMB model can more accurately apportion contributions from similar sources.
  • Uncertainty Correlation: The study found that sources with highly collinear profiles tend to have higher uncertainties in their estimated contributions. This is an important consideration for interpreting model outputs.
  • Model Performance: The PCA/MLR-CMB model demonstrated acceptable performance on both synthetic and ambient datasets, providing more stable and realistic source contributions compared to traditional methods when collinearity is present.
  • Comparison with Other Models: The paper compares the PCA/MLR-CMB results with a traditional CMB model and another advanced model (NCPCRCMB), highlighting the advantages of the proposed method in specific scenarios.

Key Results

  • Synthetic Data: The PCA/MLR-CMB model accurately estimated source contributions for synthetic datasets, even with significant collinearity and added noise. The estimated values were generally close to the true values.
  • Ambient Data (Chengdu):
    • PCA/MLR-CMB: Identified vehicle exhaust (28.71%), coal combustion (24.45%), resuspended dust (19.24%), and soil dust (16.53%) as major contributors to PM10.
  • NCPCRCMB: Showed similar dominant sources but with slightly different proportions, e.g., vehicle exhaust (30.94%) and cement (17.65%) being higher.
  • Traditional CMB: Struggled with collinear sources, yielding negative results for coal combustion.
  • Uncertainty: The study quantified uncertainties for estimated source contributions, noting that collinear sources generally exhibited higher uncertainty values.

Limitations and Future Work

  • Collinearity and Uncertainty: The paper acknowledges that while the PCA/MLR-CMB model improves handling of collinearity, sources with very similar profiles will still inherently have higher uncertainties.
  • Source Profile Quality: The accuracy of the model is still dependent on the quality and representativeness of the source profiles used.
  • Future Work: The authors suggest that the combined model can be further applied to other urban areas to test its broader applicability and to investigate the impact of different levels of collinearity on source apportionment results. The comparison with NCPCRCMB also suggests that exploring different advanced receptor models for complex scenarios is a valuable avenue.

Summary updated

BibTeX
@inproceedings{shi2011estimated,
  author = {Guoliang Shi and Fang Zeng and Xiang Li and Yinchang Feng and Yuqiu Wang and Guangxun Liu and Tan Zhu},
  title = {{Estimated contributions and uncertainties of PCA/MLR–CMB results: Source apportionment for synthetic and ambient datasets}},
  booktitle = {Atmospheric},
  year = {2011},
  doi = {10.1016/j.atmosenv.2011.03.007},
}
Figure