Adaptive spatio-temporal graph network for environmental factors forecasting in utility tunnels

Multivariate time series forecasting is crucial for constructing digital twin systems for urban utility tunnels. We collect diverse environmental sensor data for multivariate time series forecasting, ensuring tunnel safety. However, static sensor networks usually utilize fixed distances to reflect the spatial correlation, making it struggle to capture the dynamic spatial correlations among them. In addition, different environmental factors in utility tunnels exhibit temporal features of different periods. To overcome these challenges, we propose a spatio-temporal model, which can adaptively learn graph structures to grasp dynamic spatial correlations among different environmental factors and employ a temporal convolutional module to capture various temporal features among different environmental factors. Additionally, we integrate a linear capture module with a nonlinear feature in parallel, forming the graph and temporal convolution-linear capture networks (GTCLNs). We carry out experiments utilizing real-world datasets from Suzhou utility tunnels, demonstrating that our model outperforms baseline models and provides support for the precise construction of digital twin systems.