Gas concentration prediction based on temporal attention mechanism in temporal convolutional networks

Abstract

To predict indoor air quality, we can analyze the concentrations of gas present in the air. Accurate prediction of gas concentrations can help individuals identify the presence of harmful gas in the environment, thereby preventing potential accidents. Previous studies have utilized electronic nose (E-nose) in conjunction with traditional neural networks to obtain gas concentration information; However, these neural networks often experience a decline in prediction accuracy when handling longer time series data, failing to meet expected outcomes. To enhance prediction accuracy, this study introduces an innovative Temporal Convolutional Network (AGT-TCN), designed for the prediction of mixed gas concentrations in E-nose. AGT-TCN comprises a temporal convolutional network (TCN), gated recurrent units (GRU), a temporal attention mechanism and residual convolutions. In the experiments, we employed data from carbon monoxide and ethylene mixed gas (CO-ethylene) for concentration predictions and compared the results with baseline models including TCN, GRU, Long and short-term memory network (LSTM), convolutional neural network - long short-term memory network (CNN-LSTM) and Long Short-Term Memory-Transformer (LSTM-Transformer). The results demonstrate that AGT-TCN outperforms baseline in terms of prediction accuracy. This further confirms the applicability of the AGT-TCN in the early prediction of CO-ethylene concentrations.