Ada-STGMAT: An adaptive spatio-temporal graph multi-attention network for intelligent time series forecasting in smart cities

The intelligent city is an exceedingly cognizant urban configuration propelled by artificial intelligence and big data technology. Anticipating chronologically arranged data amassed by numerous sensors and equipment within the ingenious metropolis can heighten the intelligence and efficacy of urban governance. However, it is challenging to accurately predict these time series data due to their prominent spatio-temporal and complex nonlinear characteristics. In order to tackle this issue, the paper presents an innovative adaptive spatio-temporal graph multi-attention network (Ada-STGMAT) aimed at achieving intelligent forecasting of time series data characterized by intricate spatio-temporal features. Comprising three distinct modules, Ada-STGMAT includes the adaptive graph learning module which adaptively characterizes the spatial relationships among nodes. The Graph Multi-Attention Network and Time Convolution modules uncover the latent spatial-temporal dependencies within the time series. The empirical findings demonstrate that, in a 24-step prediction experiment, our model has significantly reduced the metrics of Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Squared Log Error (MSLE), and Symmetric Mean Absolute Percentage Error (SMAPE) by 23%, 21%, 41%, and 24% respectively, thereby offering an efficient approach for urban system analysis and prediction.