Subway air quality modeling using improved deep learning framework

Soft-sensing modeling of indoor air quality in subways is critical for public health. For the purpose of reducing monitoring costs and building health risk assessment models, a new deep learning forecasting model based on empirical mode decomposition, long short-term memory (LSTM) block and squeeze and excitation networks (SENet) is proposed. To begin, the original PM2.5 data is decomposed into multiple sub-series with varying frequencies using empirical mode decomposition. Then, an LSTM neural network is built to forecast the new sub series. Finally, squeeze and excitation networks were constructed and coupled to automatically pick informative weights to obtain the real-time forecasting result. The proposed model is compared to other commonly used models such as convolutional neural network and LSTM for its ability to forecast PM2.5 on an hourly experiment. The proposed model outperforms reference models in terms of forecasting performance, owing to its ability to capture informative characteristics and temporal patterns from varying PM2.5 dataset. The mean square error is improved by 38.29% and 29.21% compared with convolutional neural network and LSTM, respectively. When compared to convolutional neural networks and LSTM, the mean absolute error is reduced by 22.93% and 13.38%, respectively. Moreover, the proposed model also performs best in health risk warning assessment.

Automated summary

A new deep learning forecasting model is proposed for soft-sensing modeling of indoor air quality in subways. The model combines empirical mode decomposition, long short-term memory (LSTM) block, and squeeze and excitation networks (SENet) to achieve excellent forecasting performance and health risk assessment.