Hi @odow,
We are using SDDP.jl to develop IARA.jl (https://github.com/psrenergy/IARA.jl).
For some reason, we haven't updated the version of SDDP since 1.8.1. This week in tries to accelerate the simulation cases I have updated to SDDP 1.13.0 and many market clearing cases were broken.
The main reason is that in previous versions you added the function _get_incoming_domain
|
domain = _get_incoming_domain(policy_graph) |
to get the bounds of state variables in all nodes of the problem.
The problem here is that in order to do this query correctly you call
|
parameterize(node, noise.term) |
not during the optimization but during the construction of the problems of each node.
In our problems, we are using the structures of SDDP.jl to build a simulation that runs period by period and import the cuts trained in the policy. In that case this simulation calls parametrize and reads results of the simulation performed in the previous period. The problem is that we dont have all information needed to parametrize the problem ahead of time, making the call to parametrize when building the policy graph impossible in our case.'
I have seen that these bounds are only necessary when we are dealing with integer sub problems and we need LagrangianDuality. Would it be possible to skip this part and give an error if we do need LagrangianDuality and the bounds were not saved?
Hi @odow,
We are using SDDP.jl to develop IARA.jl (https://github.com/psrenergy/IARA.jl).
For some reason, we haven't updated the version of SDDP since 1.8.1. This week in tries to accelerate the simulation cases I have updated to SDDP 1.13.0 and many market clearing cases were broken.
The main reason is that in previous versions you added the function
_get_incoming_domainSDDP.jl/src/user_interface.jl
Line 1038 in a8be9aa
The problem here is that in order to do this query correctly you call
SDDP.jl/src/user_interface.jl
Line 1072 in a8be9aa
In our problems, we are using the structures of SDDP.jl to build a simulation that runs period by period and import the cuts trained in the policy. In that case this simulation calls
parametrizeand reads results of the simulation performed in the previous period. The problem is that we dont have all information needed to parametrize the problem ahead of time, making the call toparametrizewhen building the policy graph impossible in our case.'I have seen that these bounds are only necessary when we are dealing with integer sub problems and we need
LagrangianDuality. Would it be possible to skip this part and give an error if we do need LagrangianDuality and the bounds were not saved?