Monitoring and modelling of PM2.5 concentration at subway station construction based on IoT and LSTM algorithm optimization

Abstract

Fine particulate matter (PM2.5) generated during construction not only has great influences on the health of the workers, but also has a negative impact on the health of the surrounding public. In-depth study of PM2.5 during construction is of great significance to the environment and human health. Traditional prediction methods have limited consideration of influencing factors and low prediction accuracy. Therefore, this study proposed a Bayesian optimization-based Long Short-Time Method (LSTM) algorithm which can achieve effective PM2.5 prediction at construction sites. In this study, The Internet of Things (IoT) technology was used to conduct real-time monitoring of a subway station construction site in China for 6 months. The PM2.5 concentrations generated by different construction processes were obtained through data cleaning and analysis. The average PM2.5 increments for the excavation and support works, reinforcement works and concrete works were 17 μg/m 3 , 20 μg/m 3 , and 21 μg/m 3 , respectively. After that, feature extraction was performed by the gray correlation algorithm to eliminate redundant information. Then, a Bayesian optimization algorithm was introduced to tune the hyperparameters to optimize the LSTM model performance. Finally, the validity of the proposed model was verified using multiple base models. The results showed that the accuracy of the LSTM model was improved by 6% (R 2 from 0.88 to 0.94), after the Bayesian optimization. And the optimized model also has better evaluation indicators , with RMSE = 13.06μg/m 3 , MAE = 8.61μg/m 3 . It was proved that the method can effectively predict air pollution. • Using IoT technology to monitor construction site PM2.5. • Considering the impact of construction machinery activities on PM2.5 at construction sites. • Proposing a Bayesian optimization-based LSTM machine learning algorithm. • Validating the performance of the proposed algorithm against multiple underlying neural network models.