TL;DR
This paper introduces a novel air quality prediction model, CC-SSBLS, which enhances the Broad Learning System (BLS) by incorporating a correlation entropy criterion to robustly handle noisy data with missing values and outliers. The CC-SSBLS model demonstrates superior prediction accuracy and stability compared to existing methods, making it a promising tool for real-time air quality monitoring.
What Problem Does This Paper Address?
Accurate air quality prediction is crucial for public health and environmental management. However, real-world air quality datasets are often plagued by missing values and outliers, which significantly degrade the performance and stability of prediction models. Traditional methods struggle to effectively address these data imperfections. While Broad Learning System (BLS) offers advantages in terms of speed and handling smaller datasets, its performance can be compromised by noisy data. This paper aims to tackle the challenge of data quality issues in air quality prediction by developing a more robust and accurate prediction model.
Methods
The core of this research is the development of the Correlation Entropy-based Semi-Supervised Broad Learning System (CC-SSBLS). This model builds upon the existing Broad Learning System (BLS) framework, which is known for its efficient incremental learning and ability to achieve high accuracy without deep architectures.
Here's a breakdown of the key methodological aspects:
- Broad Learning System (BLS): The foundation of the proposed model. BLS uses random vector functional link networks, mapping input data to a feature layer and then to an enhancement layer, followed by a pseudo-inverse for output. This structure allows for rapid training and easy extension.
- Semi-Supervised Learning (SSBLS): To address the issue of missing data, the authors integrate a semi-supervised approach. This leverages both labeled and unlabeled data, improving the model's ability to learn from incomplete datasets.
- Correlation Entropy Criterion: This is the novel contribution. Instead of the standard mean square error (MSE) criterion used in BLS and SSBLS for optimization, CC-SSBLS employs a correlation entropy criterion. Correlation entropy is known to be more robust to outliers than MSE, making it better suited for handling noisy data. The paper details how this criterion is integrated into the BLS framework, particularly in the pseudo-inverse calculation.
- Comparative Analysis: The CC-SSBLS model is rigorously compared against several established models, including Random Forest (RF), Support Vector Regression (V-SVR), the base BLS, and the semi-supervised BLS (SSBLS), as well as CART-BLS.
Datasets
The study utilizes four regional air quality datasets from China:
- Xuzhou
- Nanjing
- Beijing
- Changchun
These datasets were collected at one-hour intervals from January 1, 2018, to December 31, 2021, and include air pollutants and meteorological parameters. The datasets are noted to have significant issues with missing values (up to 70% in Xuzhou) and outliers (up to 40% in Xuzhou). The data was split into 80% for training and 20% for testing. No new benchmark datasets were introduced; rather, existing real-world datasets with known data quality issues were used to demonstrate the model's effectiveness.
Key Insights and Ideas
- Outliers are a major hurdle for air quality prediction: The paper strongly emphasizes that data outliers significantly degrade the stability and accuracy of prediction models, a common problem in real-world air quality data.
- Correlation entropy is a robust alternative to MSE for noisy data: The core insight is that replacing the mean square error criterion with correlation entropy in the learning system's objective function makes it significantly more resilient to outliers. This is because correlation entropy operates in a kernel space, offering a more generalized measure of similarity that is less sensitive to extreme values.
- Semi-supervised learning is crucial for missing data: The integration of semi-supervised learning effectively addresses the challenge of missing data, which is prevalent in air quality datasets, thereby improving overall model performance.
- Broad Learning System's potential for real-world applications: The research highlights that by adapting BLS with robust criteria like correlation entropy, it can be effectively applied to complex, real-world problems like air quality prediction, overcoming limitations of simpler BLS implementations.
Key Results
The CC-SSBLS model demonstrated superior performance across all evaluated metrics and datasets:
- Improved Accuracy: The CC-SSBLS model achieved the lowest Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) among all tested models. For instance, in the Nanjing dataset, the RMSE was 8.68 µg·m⁻³ and the MAPE was 0.24%.
- Enhanced Stability: Compared to BLS and SSBLS, CC-SSBLS exhibited significantly less variation in evaluation metrics across different iteration windows, indicating more stable predictions.
- Outperformance of Baseline Models: CC-SSBLS consistently outperformed RF, V-SVR, BLS, SSBLS, ANN, and RNN models in terms of RMSE, MAPE, and MAE, while achieving higher R² values. For example, on the Xuzhou dataset, CC-SSBLS achieved an RMSE of 11.94 µg·m⁻³, which was 3.31 µg·m⁻³ lower than SSBLS, and an MAE of 6.89 µg·m⁻³, 2.30 µg·m⁻³ lower than SSBLS.
Limitations and Future Work
The authors acknowledge a few limitations and suggest avenues for future research:
- Focus on Prediction Effect: The current study primarily focuses on the prediction accuracy of the CC-SSBLS model. The training time of the model was not analyzed in detail. Future work could investigate methods to reduce training time and improve overall model efficiency.
- Further Model Enhancement: While CC-SSBLS shows significant promise, there is scope for further strengthening its prediction capabilities. Future research could explore advancements in the feature layer, enhancement layer, and the pseudo-inverse solving principle to further boost prediction accuracy and efficiency.
- Broader Application: The study focused on air quality prediction. Extending the CC-SSBLS model to other domains with similar data quality challenges could be a valuable future direction.
Summary updated