Abstract
Predicting urban nitrogen dioxide ($NO_2$) levels is a complex task due to the non-linear nature of pollutant dynamics and the dependency on often unavailable meteorological data. This paper presents a Cyclostationary Neural Network (CNN) model that leverages the periodic nature of urban traffic patterns. By decomposing the daily cycle into 24 independent sub-networks, the model achieves high prediction accuracy using only historical pollutant data. Experimental results on datasets from Northern Italy demonstrate that the CNN significantly outperforms traditional linear models, particularly during high-traffic periods, without requiring exogenous weather variables.
Introduction
Urban air pollution prediction is critical for public health but remains challenging. Traditional approaches typically fall into two categories: atmospheric diffusion models, which require precise but often unavailable meteorological data, and statistical models, which often struggle with the non-linear dynamics of urban environments. This study introduces a Cyclostationary Neural Network (CNN) designed to capture these non-linearities using only historical pollutant data, effectively treating meteorological influences as implicit in the pollutant time series.
Methods
The proposed CNN architecture exploits the cyclostationary nature of urban pollution, where statistical properties vary periodically over a 24-hour cycle driven by traffic. Instead of a single complex network, the system comprises 24 independent Multi-Layer Perceptrons (MLPs), each dedicated to predicting the $NO_2$ concentration for a specific hour of the day.
Each MLP uses a single hidden layer with five neurons and is trained using the Levenberg-Marquardt quasi-Newton method. The inputs are restricted to the current $NO_2$ level and the $NO$ level from 24 hours prior, making the model independent of external weather data.
Experimental Setup
The model was evaluated using hourly $NO$ and $NO_2$ data from Station 649 in Brescia-Broletto, Northern Italy. Two evaluation sets were used: a small set (1994-1995) and a larger set (2003-2004) to test scalability and robustness.
Results and Discussion
The CNN was compared against a standard AutoRegressive eXogenous (ARX) model. While both models performed similarly on the smaller dataset, the CNN demonstrated superior scalability, significantly outperforming the ARX model when trained on the full 12-month dataset.
Key insights from the results include:
- Implicit Weather Modeling: The pollutant time series effectively "bakes in" meteorological influences, allowing the model to remain robust without explicit weather inputs.
- Efficiency through Decomposition: Splitting the task into 24 hourly blocks simplifies the learning process, as each sub-network focuses on the specific dynamics of its assigned hour (e.g., morning rush hour).
- Non-linear Advantage: The CNN's advantage is most pronounced during peak traffic hours, where pollution dynamics are most complex and non-linear.
Conclusion
The Cyclostationary Neural Network offers a robust and efficient alternative for real-time urban air quality forecasting. By focusing on the periodic nature of human activity and utilizing a modular architecture, it provides high-accuracy predictions without the need for complex external data feeds.
Summary updated