Deep learning for construction emission monitoring with low-cost sensor network

Emissions from construction activities, particularly in metropolitan areas, are carefully monitored to prevent health problems and environmental degradation. The data quality of low-cost wireless sensors in construction sites remains a challenge for pollution predictive models due to uncertainties of measurement and volatile environment. In this study, we propose a hybrid model using a Long short-term memory integrated with a Bayesian neural network to infer the probabilistic forecasts of particulate matters (i.e., ??1.0, ??2.5, and ??10) emitted from construction activities. The training data are fused by two sources: (1) our developed low-cost wireless sensor network (LWSN) monitoring at a construction site located in Melrose Park, Sydney, Australia, and (2) air-quality stations (AQSs) in four suburbs nearby that monitoring site. The proposed model (LSTM-BNN) is compared with other deep learning methods, namely Gated recurrent unit (GRU), Bidirectional long short-term memory (BiLSTM) and One-dimension convolution neural network (1D-CNN), commonly used for time-series forecast. The experimental results indicate the outperformance of our model to all benchmark models and display a significant improvement at 56.3%, 27.9% and 37.9% in MAEs forecast for all three types of particles compared to a deterministic LSTM model.