I don’t belive the current literature solves it with SDDP
Ah… 
I don’t think you should put effort into attempting to solve this model with SDDP.jl, and I’m not aware of any other algorithm that you can use.
It fundamentally can’t be decomposed by time because the decisions to be made in a node depend (via a constraint, not the objective) on the decisions that are made for every possible realization of the uncertainty in the next time-step. This also hints at which the literature solves this problem with much smaller problems…
I would instead consider alternative formulations that you could investigate that yield similar behavior and yet are more amenable to computation.
As one example, you might add \eta as a state variable, and penalize or constrain ||\eta_m - \eta_n||. This would work because it requires adding only one state variable (instead of 8*8) and it doesn’t require knowing the full distribution of the noise in the next time-step (since \eta_m happened in the past and you know exactly what it is).