As one of the harmful substances in the atmosphere, carbon monoxide (CO) is harmful to human beings. With the wide application of gas sensors and machine learning algorithms, the accuracy of concentration predicting of various gas such as CO is constantly improving. Now we apply TrellisNet, a network utilizing both recurrent and convolutional techniques, to gas concentration prediction, which is also a time-series prediction task, with the aim of improving its performance. To enhance TrellisNet’s ability to retain long time series information, we replaced the activation units in each layer of the model with gated recurrent unit (GRU). Compared to using long short-term memory (LSTM) as the activation unit, our approach has lower computational complexity and offers a more stable model. Additionally, we introduced dilated convolutions in each layer, allowing the model to establish connections with as many past time steps as possible at a given time point, even with fewer layers. This further enhances the preservation of long time series information. We named our improved technique trellis convolutional dilated network (TrelliSense). Due to the injection of the same input values in each layer, TrelliSense also exhibits superior training stability. Experimental results demonstrate that TrelliSense outperforms other time prediction networks, including temporal convolutional network (TCN), LSTM, GRU, Gaussian-TCN and bidrectional lstm (Bi-LSTM) in terms of all error metrics (MAE, RMSE, SMAPE). Therefore, we argue that TrelliSense is a better method for predicting CO concentration.