TASGCN: A Temporal Attention and Spatial Graph Convolution Network for Multivariate Time Series Prediction

Multivariate time series prediction is extremely important but challenging in fields such as finance, energy, transportation, etc. Due to the highly complex dependencies between pairs of variables and the non-stationarity of time series, existing works fail to obtain satisfactory results. We are committed to modeling complex spatial-temporal relationships in multivariate time series, proposing TASGCN, which adapts the attention mechanism to extract temporal relations and graph convolution to model spatial dependencies. Specifically we introduce a graph generation layer to dynamically extract directed relations among variables. TASGCN demonstrates excellent prediction results in our datasets originating from real-world data from data center computer room environments. In particular, TASGCN outperforms other advanced baseline models by up to 26% in terms of the MAE evaluation metric.