Prediction of Air Pollutants Concentration Based on an Extreme Learning Machine: The Case of Hong Kong

Jiangshe Zhang, Weifu DingInternational Journal of Environmental Research and Public Health 2017 · DOI

TL;DR

This paper introduces the Extreme Learning Machine (ELM) as a superior method for predicting air pollutant concentrations in Hong Kong, outperforming traditional methods like Multiple Linear Regression (MLR) and Feedforward Neural Networks with Backpropagation (FFANN-BP). ELM achieves faster training times and better predictive accuracy, offering a more efficient and reliable tool for environmental monitoring and management.

What Problem Does This Paper Address?

Metropolitan areas worldwide are grappling with rising levels of ground-level air pollutants, posing significant risks to public health and ecosystems. Accurately predicting these pollutant concentrations is crucial for environmental agencies to implement effective control strategies and manage public health emergencies. While traditional methods like Artificial Neural Networks (ANNs) have been employed, they often suffer from slow convergence rates and a tendency to get stuck in local minima, hindering their practical application. This paper aims to address these limitations by exploring a more efficient and robust machine learning approach for air pollution prediction.

Methods

The study compares three distinct modeling approaches for air pollutant prediction:

  • Multiple Linear Regression (MLR): A traditional statistical method that models the linear relationship between explanatory variables (meteorological conditions, lagged pollutant concentrations) and the response variable (air pollutant concentration).
  • Feedforward Neural Network with Backpropagation (FFANN-BP): A widely used ANNs approach that learns by iteratively adjusting weights and biases to minimize prediction errors. However, it is known for its slow learning speed and susceptibility to local minima.
  • Extreme Learning Machine (ELM): The core contribution of this paper. ELM is a single-hidden-layer feedforward neural network where the weights of the hidden layer are randomly assigned and fixed. Only the output layer weights are trained using a simple generalized inverse operation. This approach significantly speeds up training and avoids the local minima problem, leading to better generalization performance.

The researchers also employed a data preprocessing step of normalizing input and target data to a range of [-1, 1] to ensure fair comparison across models. They used 10-fold cross-validation to assess the generalization ability of the models.

Datasets

The study utilizes six years of daily air quality data (2010–2015) from two air quality monitoring stations in Hong Kong: Sham Shui Po and Tap Mun. The data includes measurements for five key air pollutants: nitrogen dioxide (NO2), nitrogen oxide (NOx), ozone (O3), particulate matter under 2.5 µm (PM2.5), and sulfur dioxide (SO2). These pollutant concentrations were averaged over 24-hour periods to represent daily mean concentrations.

In addition to pollutant data, the models were trained using eleven predictor variables, including meteorological parameters (daily average temperature, daily maximum/minimum temperature, relative humidity, wind speed, wind direction) and three time variables (day of the week, month of the year). Lagged air pollutant concentrations were also included as predictors.

Key Insights and Ideas

  • ELM's Speed Advantage: The most striking insight is ELM's dramatically faster training time compared to FFANN-BP and MLR. This is attributed to its direct calculation of output weights, bypassing the iterative gradient descent process that plagues FFANN-BP.
  • Overcoming Local Minima: ELM's random initialization of hidden layer weights and direct solution for output weights effectively sidesteps the problem of getting trapped in local minima, a common issue with gradient-based learning algorithms like FFANN-BP. This leads to more robust and reliable predictions.
  • Generalization Power: ELM demonstrates superior generalization ability, meaning it performs well on unseen data. This is crucial for real-world applications where models need to adapt to varying conditions.
  • Practical Implications for Environmental Monitoring: The findings suggest that ELM can provide a more efficient and accurate tool for predicting air pollutant concentrations, enabling environmental agencies to make better-informed decisions for pollution control and public health advisories.

Key Results

The study found that ELM consistently outperformed both MLR and FFANN-BP across various performance metrics:

  • Coefficient of Determination (R²): ELM achieved higher R² values for most pollutants and seasons, indicating a better fit to the data. For instance, for PM2.5, ELM's R² ranged from 0.70 to 0.82, significantly higher than MLR (0.50-0.64) and FFANN-BP (0.52-0.67).
  • Root Mean Square Error (RMSE): ELM consistently yielded lower RMSE values, signifying more accurate predictions with less deviation from observed values.
  • Speed: ELM was found to be significantly faster, running up to 25 times faster than FFANN-BP and eight times faster than MLR in some scenarios. For example, ELM took only 0.05 seconds for learning, while FFANN-BP took nearly 3 seconds.
  • Generalization: Figures 3-7 visually demonstrate ELM's tighter clustering of predicted versus observed values, indicating better generalization compared to the other models.
  • Episode Prediction: ELM also showed promising results in predicting high pollution episodes, achieving a high Probability of Detection (POD) of 0.73 and a reasonable False Alarm Rate (FAR) of 0.27 for PM2.5 exceedances.

Limitations and Future Work

  • Data Limitations: The authors acknowledge that the data sets are not entirely noise-free and that data incompleteness (NA values) can limit the study. Approximately 3% of the data for each pollutant was removed due to missing values or outliers.
  • Model Complexity: While ELM is faster, the paper doesn't delve deeply into the optimal number of hidden nodes for all scenarios, relying on cross-validation. Further research could explore more sophisticated methods for determining optimal ELM architectures.
  • Future Work: The authors suggest that ELM's performance could be further enhanced by exploring different combinations of input variables and activation functions. Additionally, investigating the application of ELM to other air pollutants and in different geographical regions would be valuable. The paper also hints at the potential for ELM in predicting ozone threshold exceedances, which could be a fruitful area for future research.

Summary updated

Abstract

With the development of the economy and society all over the world, most metropolitan cities are experiencing elevated concentrations of ground-level air pollutants. It is urgent to predict and evaluate the concentration of air pollutants for some local environmental or health agencies. Feed-forward artificial neural networks have been widely used in the prediction of air pollutants concentration. However, there are some drawbacks, such as the low convergence rate and the local minimum. The extreme learning machine for single hidden layer feed-forward neural networks tends to provide good generalization performance at an extremely fast learning speed. The major sources of air pollutants in Hong Kong are mobile, stationary, and from trans-boundary sources. We propose predicting the concentration of air pollutants by the use of trained extreme learning machines based on the data obtained from eight air quality parameters in two monitoring stations, including Sham Shui Po and Tap Mun in Hong Kong for six years. The experimental results show that our proposed algorithm performs better on the Hong Kong data both quantitatively and qualitatively. Particularly, our algorithm shows better predictive ability, with R 2 increased and root mean square error values decreased respectively.
BibTeX
@article{zhang2017prediction,
  author = {Jiangshe Zhang and Weifu Ding},
  title = {{Prediction of Air Pollutants Concentration Based on an Extreme Learning Machine: The Case of Hong Kong}},
  journal = {International Journal of Environmental Research and Public Health},
  year = {2017},
  doi = {10.3390/ijerph14020114},
}
Figure