State space issue

[Kacjer]

Hi, Hou.

Thanks for sharing. Just one question regarding the state space limits. It seems that the net_load could be negative. In that case, shouldn't the normalized state space defined as [-1,1]?

Thanks.

[EnergyQuantResearch]

This doesn't impact the performance according to my experience, state normalization and denormalization is to stable the training by scaling the state into a magnitude that agents can easily handle. But you can try to do some ablation cases using my framework. I did several methods for normalization when running the model, but didnt see any big performance derivations.

[ShengrenHOU]

Thanks, you were right.

net_load can indeed become negative on the dataset, so constraining the normalized state space to [0, 1] was too restrictive. This has now been corrected in the current main branch by allowing the net_load dimension to use a negative lower bound.

I’m closing this issue because the state-space bound has been fixed in main.