Monitoring and estimation of urban emissions with low-cost sensor networks and deep learning

Abstract

Sustainable development in cities requires advanced technologies for monitoring and estimating air pollution emissions, which directly affect the health of local inhabitants and residents in the neighborhoods. For this, low-cost sensors and information technologies are increasingly used to provide accurate air quality forecasts. They are, however, subject to data constraints. This paper presents new techniques for accurate, reliable monitoring and forecasting of air pollution at various scales using data from IoT-enabled sensors along with state-run air-quality monitoring stations. Here, we develop an extended deep-learning model based on neural networks and algorithms for optimization of the model hyperparameters and network dropout rates. These can yield a significant improvement of over 31% in prediction accuracy while maintaining a prediction coverage of approximately 80% for forecasting air-particle levels over a 24-h period. The advantages and effectiveness of our model are validated and verified in two real-world scenarios, a suburban construction site and a civil infrastructure development project. Comparison analysis is conducted to indicate the outperformance of the proposed method over two recent techniques for probabilistic time series forecasting used for air pollution estimation on regular days and extreme events.

Keywords

• Urban emissions
• Data constraint
• Prediction interval
• Bayesian optimization
• Deep learning