stavrosorf/poweflownet

PoweFlowNet: Leveraging Message Passing GNNs for Improved Power Flow Approximation

poweflownet's Issues

Config file for best performance model

I see that in the pull request the arguments are written in a bash file. It'd be better to write a json file in the configs/ folder. Neater this way

Result Plots and Tables

Revamp the plots of the paper.

Since we generate Datasetv3 and need to rerun experiments we can also add results for more cases (need to discuss this). Cases between 118 and 6470 might be of interest (the greater the number of nodes the better our architecture look).

We show transferability is not great with our architecture, but what if we did not train for 1000 epochs (->overfit) and try to find a way to make it more generalizable :

1. Do few-shot training on the new grid case
2. Less initial training epochs

Tasks

• Create new script for Error Per Feature Distribution Plots
• Revamp the Interpretability plots to make the text visible
• Find the exact error of the DC Power Flow method

Prepare repository for publication

Tasks

• Tidy up the repository branches and files.
• Write a complete ReadMe with execution and setup instructions.

argparser possibly incorrectly defined

See: argument_parser.py, line 13 and 15.
@sentient-codebot will fix later.

Data Generation v4

Plan for Data Generation v3

1. Use entries in the admittance matrix as edge features.
   1. In this way, we do not need to put extra care on the shunt capacitance of the lines anymore and we do not need to remove them; the generated data are also more realistic.
   2. We should also be able to allow non-zero $G_s$ and $B_s$ because they will also be integrated into the the admittance matrix. (We can allow this only when this does not influence our output format (P, Q, V, theta)
2. The admittance matrix elements are non-zero only where two nodes are directly connected and in the diagonal line. To view them as edge features, we also need to adjust our formulation of the edges. That is, we need to add self-loops at every node to assign the diagonal elements in the admittance matrix.
3. The admittance matrix is symmetric, meaning the according edges are undirected. So we might as well make our edge_index also containing undirected edges, which is different from our v1 and v2 data generation. We perhaps need to adjust our network slightly.
4. Since the meaning of edge features changed, we need to adjust our implementation of the physical loss.

Action Items

• change the edge feature generation in the script dataset_generator_pandapower_nan.py -> get from nodal admittance matrix AND undirected.
• make edges undirected and add self-loops in the edge_index variable.
• remove the "unrealistic" modifications on the C_nf per line.
• allow non-zero $G_s$ and $B_s$
• [optional] adjust our sampling method of $P_d$ and $P_g$, as when they are too imbalanced, the power flow would not converge.
• double check our network inference steps to make sure it align with the new edge formulations.
• revise the PowerImbalance (physical loss function) to make it align with the new edge features. do sanity checks (with target data). the result should work perfectly, since the admittance matrix should be mathematically sufficient and strict for calculating the power flow.

wonder if there is description about pypower.networks.case118?

I'm confused about the columns in the default test case in pandapower.networks.case118 as I'm a fish in this elec and DL crossing field.

I can only follow your code but don't know their meanning when debug in this repo file: dataset_generator_alt.py
I'd appreciate it if you can offer help for identifying the specific column names in networks.case118['bus'] and networks.case118['branch']

Dataset and models link is not public

Task

• Move the Dataset and models to publicly available storage