A Novel Air Pollutant Concentration Prediction System Based on Decomposition-Ensemble Mode and Multi-Objective Optimization for Environmental System Management

Yan Hao, Yilin Zhou, Jialu Gao, Jianzhou WangSyst 2022 · DOI

TL;DR

This paper introduces a novel Air Pollutant Concentration Prediction (APCP) system designed to improve both accuracy and robustness in predicting air quality. By combining a decomposition-ensemble mode for data reconstruction with a multi-objective optimization approach for model weighting, the system offers a more reliable tool for environmental management decisions.

What Problem Does This Paper Address?

Air pollution, particularly from PM2.5, poses significant threats to human health and incurs substantial economic losses globally. Accurate and reliable prediction of air pollutant concentrations is crucial for timely environmental management and early warning systems. While existing methods often focus solely on prediction accuracy, they frequently overlook the critical aspect of model robustness, leading to systems that might perform well under ideal conditions but falter when faced with real-world complexities and noise. This paper aims to bridge this gap by developing a system that balances both prediction precision and stability.

Methods

The proposed APCP system is built upon four core modules:

  • Time Series Reconstruction (Decomposition-Ensemble Mode): This module addresses the inherent noise and fluctuations in air pollutant data. It employs a decomposition-ensemble (DE) mode to filter out high-frequency noise and reconstruct the original sequence, aiming to retain key information while reducing interference.
  • Submodel Simulation: Four individual prediction submodels are utilized: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Extreme Learning Machine (ELM), and Elman Neural Network (ENN). These models are chosen for their complementary strengths in capturing various sequence modes.
  • Weight Search (Multi-Objective Optimization): A novel weight search (WS) mechanism, based on a multi-objective grasshopper optimization algorithm (MOGOA), is developed. This mechanism aims to find optimal weights for the submodels, balancing the dual objectives of prediction precision and robustness, thereby overcoming limitations of single-objective optimization methods.
  • Integration: The final prediction is generated by integrating the outputs of the four submodels, weighted according to the optimal weights determined by the WS mechanism.

Datasets

The study utilizes hourly PM2.5 concentration data from three Chinese cities: Guangzhou, Shanghai, and Chengdu, for January 2015. This data was sourced from the UCI Machine Learning Repository. The datasets were split into training and testing sets for model evaluation.

Key Insights and Ideas

  • Balancing Accuracy and Robustness: The core innovation lies in explicitly addressing both prediction accuracy and robustness, moving beyond single-objective optimization common in many existing models.
  • Decomposition-Ensemble for Noise Reduction: The DE mode effectively filters out noise from historical air pollutant data, which is crucial for improving the performance of subsequent prediction models.
  • Ensemble of Diverse Models: Combining CNN, LSTM, ELM, and ENN allows the system to leverage different modeling capabilities, leading to a more comprehensive capture of data patterns.
  • Multi-Objective Weighting: The WS mechanism, by considering multiple objectives, provides a more nuanced approach to combining submodel predictions, leading to better overall performance.

Key Results

The APCP system demonstrated superior performance across all three cities and for multi-step predictions (1, 2, and 3 steps ahead).

  • Guangzhou: The APCP system achieved the best results across all measurement indicators for 1-step predictions, with a minimum MSE of 20.861. It also showed significant advantages in 3-step predictions, with MSE reduced by 58.46% to 76.50% compared to other models.
  • Shanghai: While the overall performance was slightly less pronounced than in Guangzhou, the APCP system consistently outperformed individual models, especially in longer prediction steps. For instance, in 3-step predictions, the MSE was 13.637 lower than the best individual model (CNN).
  • Chengdu: The system showed strong performance, with R² values indicating the historical data's strong interpretability for future predictions. The APCP system's MSE for 2-step predictions was 157.888, significantly lower than the average MSE of well-known models.
  • Robustness: Sensitivity analysis indicated that the APCP system's performance was not significantly affected by changes in key parameters (iteration number, population number, archive size), highlighting its excellent stability.

Limitations and Future Work

  • External Factors: The current model primarily relies on historical air pollutant concentration data. The authors acknowledge that many external factors, such as wind speed, temperature, and humidity, also significantly influence AP concentration changes.
  • Future Work: The authors suggest expanding the APCP system to incorporate these external factors to further improve its prediction performance and real-world applicability.

Summary updated

BibTeX
@article{hao2022novel,
  author = {Yan Hao and Yilin Zhou and Jialu Gao and Jianzhou Wang},
  title = {{A Novel Air Pollutant Concentration Prediction System Based on Decomposition-Ensemble Mode and Multi-Objective Optimization for Environmental System Management}},
  journal = {Syst.},
  year = {2022},
  doi = {10.3390/systems10050139},
}
Figure