syN-BEATS for robust pollutant forecasting in data-limited context

Area of Science:

• Environmental Science
• Artificial Intelligence
• Data Science

Background:

• Accurate air pollutant forecasting is crucial for public health.
• Limited monitoring data poses a significant challenge for traditional forecasting models.
• Developing robust models for data-scarce environments is essential for environmental management.

Purpose of the Study:

• To introduce syN-BEATS, an ensemble deep learning model for pollutant forecasting with limited data.
• To evaluate syN-BEATS' performance against standard models under data-constrained conditions.
• To demonstrate the model's utility in supporting health alert systems in under-resourced areas.

Main Methods:

• Developed syN-BEATS, an ensemble deep learning model based on the N-BEATS architecture.
• Integrated diverse configurations of stacks and blocks, combining weak and strong learning.
• Utilized Bayesian optimization for fine-tuning ensemble weights.
• Simulated limited data environments using single meteorological and air quality monitoring stations per region.

Main Results:

• syN-BEATS demonstrated superior performance compared to standard models.
• The model achieved consistently low relative root mean square errors, indicating precise forecasting.
• Bayesian optimization significantly enhanced ensemble weight tuning and model accuracy.
• The model proved flexible and resilient across diverse climatic and air quality conditions.

Conclusions:

• syN-BEATS offers effective pollutant forecasting capabilities, particularly in data-limited scenarios.
• The model's performance supports the development of public health alert systems in resource-limited regions.
• This research advances environmental monitoring and public health management strategies for areas with restricted infrastructure.