Quick Start Guide
With OPFLearn installed and a network data file (e.g. "pglib_opf_case5_pjm.m") available in the current directory, an AC OPF dataset with N samples can be created with,
using OPFLearn
N = 100
results = create_samples("pglib_opf_case5_pjm.m", N)When creating datasets for networks with greater than around 50 buses, runtimes can be improved by reducing the initial sampling space through either the pd_max, pd_min, or pf_min arguments. OPFLearn has been tested on networks up to 300 buses with pd_max specified as 2 times the nominal load at each bus. For the 'pglib_opf_case300_ieee' test network a dataset of 20,000 samples took approximately a day to create using 40 distributed processors on an HPC node.
Getting Results
OPFLearn's create_samples functions return the resulting sample data as a dictionary. This dictionary contains sub-dictionaries containing feasible AC OPF sample input, output, and lagrangian dual values. Each of these three primary data category's dictionaries map variable key values to arrays containing the corresponding data for relevent elements in the network. These arrays are two dimensional with columns corresponding to elements in the system and rows corresponding to different AC OPF samples.
The input data, queried with results["inputs"], is a dictionary containing information for input variables to the AC OPF problem. The output data, queried with results["outputs"], is a dictionary of AC OPF solutions corresponding to the input data load profiles. The dual data, queried with results["duals"], is a dictionary containing information about the lagrangian dual values found when solving the AC OPF problem for each sample. Nonzero values (In OPFLearn values greater than 1e-5) indicate that the contraint associated with a dual value is active.
For example, the following dictionary query can be used to find the active power, pd, at each load in the network for the first saved AC OPF sample,
results["inputs"]["pd"][1,:]Note that input and output data is in per unit. The base MVA can be found from the PowerModels network data dictionary.
By default input, output, and dual result data for all variables are saved to the results object. To reduce the size of the resulting data an array of the desired variables can be provided in the create samples call. For example, if you are only interested in saving the generator active power, p_gen, and generator bus voltage magnitudes, vm_gen, the following call can be made,
outputs = ["p_gen", "vm_gen"]
results = create_samples("pglib_opf_case5_pjm.m", N, output_vars=outputs)This results object will only contain output data for these two specified output varaibles. Note, the input and dual results will still contain all the default variables, unless also specified.
Additional Results
More information from each run can be saved by specifying additional parameters when calling create samples. This information includes iteration statistics, found maximum load bus demands, infeasible AC OPF samples, and the sampling space polytope definition. See the Additional Results section for instructions on how to enable saving these additional results.
Distributed Processing
If there are multiple processors available, the runtime to create samples can be reduced by using distributed processing. OPFLearn has addition functions for creating AC OPF datasets that utilize distributed processing, which can be read about in the Distributed Processing section.
Issues & Questions
If you have any issues while using OPFLearn or think of any features that would be useful to you, please submit a new issue here. Any other questions or comments can be directed to Trager Joswig-Jones and/or Ahmed S. Zamzam.