Adding Entity ID about patchtst

[yuqinie98]

Hi @ikvision, sorry for the late reply. We haven't tried that yet, but this is a novel idea.

[namctin]

Hi @ikvision, that is a good suggestion. We did try pretraining the model with multiple datasets (ETT, electricity, traffic,...) and then finetune on a single dataset (ETT11). In order to encode the station id, you can have a separate linear layer to encode the station id and then concatenate it with each input patch. Since the station id is not time series, input patches will have the same id encoded in it. Certainly that is not the only way. You can try different approaches and please let us know which one is the best. Thank you!

[ikvision]

@namctin in the same topic of introducing id to enable the joint adaptation during training.
Did you try to encode the different metrics with metric_id and then concatenate it with each input patch?