Algorithm 1: The training stage of the proposed method

Require: Original dataset MS; Number epochs N Initialized embedding vector u, Gat $Gat(·)$,Gru $Gru(·)$, short time series forecasting $SF(·)$, Encoder $E(·)$, Decoder $D(·)$, and fully connected layer $f_{\theta }(·)$ $W\leftarrow u$ $n\leftarrow 1$ Repeat    for each $S_i\in \mathrm{MS}$ do:       $\mathrm{X}\leftarrow Gru\left(Gat(S_i,W)\right)$       ${\widehat{S}}_K\leftarrow SF\left(\mathrm{X}\right)$       $Los{s}_S=\Vert {\widehat{S}}_k-S_k{\Vert }^2$       ${\mathrm{X}}_{\mathrm{token}}\leftarrow E\left(\mathrm{X}\right)$       ${\mathrm{X}}_{\mathrm{de}}={\mathrm{Concat}(\mathrm{X}}_{\mathrm{token}}{,\mathrm{X}}_0)$       ${\widehat{S}}_j\leftarrow D\left({\mathrm{X}}_{\mathrm{de}}\right)$       $Los{s}_L=\Vert {\widehat{S}}_j-S_j{\Vert }^2$       ${\widehat{S}}_i=f_{\theta }\left(Concat({\widehat{S}}_k,{\widehat{S}}_j)\right)$       $Los{s}_z=\Vert {\widehat{S}}_i-S_i{\Vert }^2$       $Loss=\lambda Los{s}_s+\kappa Los{s}_L+\tau Los{s}_z$       $u,Gat(·),Gru(·),E(·),D(·)\leftarrow$update weights using $Loss$    end for $n\leftarrow n+1$ until  $n=N$