bdonnot / leap_net Goto Github PK

The current implementation of given_list_topo_encode does not allow to detect a topology present in list_topos if a line is disconnected at this substation.

For instance if we have in list_topo this sub_topology_in_dic=[1,2,2,2,1] and that the first line is disconnected giving this sub_topology_in_obs=[1,-1,2,2,1], a match is not found

The import of PandaPowerBackend does not match the last version of grid2op.

Today, the "proxy" models are trained "on the fly" and we cannot train them for more epochs. It would be nice, once the data are generated, to be able to continue the training.

It would also be nice to be able to train them if data are loaded from somewhere instead of being generated.

It's me again🤣🤣🤣

In your paper, you mentioned that there are 11558 possible unary topologies(case118), how can I understand the number 11558?

The operation "merging" or "splitting" on single node are obviously not the same with the operation "disconnect" or "connect" on pair of nodes. If I pick "disconnect" or "connect" on pair or nodes, I think I may have 118*117/2=6903 possible topologies. And 118 possible topologies for "merging" or "splitting"? I am not sure about it.

For "merging" and "splitting", will the system case118 become case117 or case119?

Here is your description:

There are 11,558 possible unary actions (corresponding to single node splitting or merging, compared
to the reference topology)

and:

(a) 50,000 of them are sampled with injections x and the reference topology τ = τ ∅ (all substations in the reference topology), (b) then 100 unary actions are sampled randomly among
the 11,558 unary topologies, as well as 1000 injections x, which
generates the other 100 000 samples.

Especially for the topology encoding when "topo_vect_to_tau" is not "raw"