In this repository there is the necessary to reproduce the figures of the neurocomputing paper LEAP Nets for System Identification and Application to Power Systems
The first figures are easily reproducible using the provided notebooks. Unfortunately, to train neural network, a variant of Stochastic gradient descent algorithm is used. This makes it pretty hard to reproduce exactly the figures showed in the paper. The detail results may vary depending on the initialization of neural network, order of minibatch etc. Though the results might slightly vary, the conclusion holds.
- Figures_3_5 details the code to produce the code necessary to get the figures 3 and 5. This implements the LEAP Net architecture, the ResNet architecture using the tensorflow keras framework. It takes roughly 10-15 mins to run it.
- Figures4_part1 contains the code to make the figure 4. This systematic experiments requires to train roughly 4000 (small) neural networks to get the exact figure of the paper. It takes approximately on my laptop. This notebook is used to train the neural networks and regularly save the results to make the figure on the hard drive in json format.
- Figure4_part2 shows the code to make the figure from the json generated by the previous notebook. You can use it either with the data provided in the package if you want to rapidly remake the figure. If however you started the previous notebook, you can use it to update the figure during the training.
- generate_data is the base script used to generate the data to reproduce the figures 10, 11 and 12. As real confidential data have been used for Figure 13, nothing will be provided to replicate the results.
- train will be the main script to train the model. Note that at this time, this script will allow easily to implement again the code of the paper. But the parameters do not match the one in the paper. Neural networks don't have the right shape nor inputs / outputs. This part is still under development.
NB at the time of writing, a proprietary powerflow was used to generate the data necessary to train a LEAP neural network on powersystem related issues (showed in the paper in figures 10 to 13). We will make our best effort to port the code used at that time to leverage the new possibility offered by the Grid2Op framework. Figures will be really close, but unfortunately we cannot open source the exact code that use the proprietary software used to generate the exact figures of the paper.