An optimized TCN-LSTM model for predicting PM2.5 in metro systems.

Abstract

Metro has become one of the main transportation modes for people's daily travel, and a good indoor air environment helps ensure people's health. This study aims to develop a data-driven, soft-measurement-based model for predicting and optimizing key metrics of metro air quality. In order to capture the key features in the indoor air quality data, a new model combining a temporal convolutional network (TCN) and long short-term memory (LSTM) is introduced in this study. As an example, subway air quality data from Seoul City Hall Station in South Korea are unified to reduce the complexity of the subsequent process. The TCN and LSTM, which perform better as single models, are chosen to build a hybrid model to capture more detailed features in it, and an attention mechanism is introduced to predict PM_2.5, which is the most important metric in indoor air quality data. In addition, experiments are conducted to compare the size of the residual modules and convolution kernels, which are critical parameters in the TCN model. Finally, the proposed TCN-LSTM model achieves a coefficient of determination of 0.88 on the test set, demonstrating superior prediction performance relative to other baseline models.