Figure 2. .

Block-level architecture of the transformer model. Left: The architecture of the spatial transformer component, where T_w is a temporal window (time series segment) within which the input data are derived, and y^s is the output of this transformer. The output of the positional embedding is supplied to the graph convolution network and the attention in parallel. The output of these two components is then fused through a gate mechanism to generate the features. Right: The architecture of the temporal transformer block. The input to this block is the output of the spatial block combined with the input to the spatial block by a residual connection (also see Figure 1).