With the rapid expansion of urban utility tunnels, the safety of their environments has become an important research area. We collect diverse environmental sensor data for multivariate time series forecasting, ensuring tunnel safety. However, there is an inherent spatial dependency among variables from different sensor locations. Static sensor networks usually utilize fixed distances to reflect the spatial correlation, and it is difficult to capture the dynamic spatial correlation among them. In addition, different environmental factors in the utility tunnels exhibit temporal features of different periods, and traditional methods are incapable of capturing the temporal features of different periods that exist among these different environmental factors. To overcome these challenges, this paper proposes a spatio-temporal model that combines a graph convolution module and a temporal convolution module. The model adaptively learns graph structures to capture dynamic spatial correlations and utilizes the temporal convolution module to capture temporal features of different periods among different environmental factors in utility tunnels. To better capture the trend and seasonal variations among environmental factors in utility tunnels, we also integrate the linear capture module with a nonlinear neural network in parallel, resulting in our proposed graph and temporal convolution-linear capture networks (GTCLNs). We conduct experiments on the real-world datasets collected in Suzhou utility tunnels. The experimental results surpass the performance of existing baseline methods, proving the effectiveness of our model.