This is a repository managing report files and experiment codes implementing the contents covered in [1].
If possible, further studies will be covered.
- Complete the proposal.
- Build a DC-OPF solver.
- Pre-process the given datasets.
- Build a Neural Nets Classifier.
- Tune the hyper-parameters.
- Visualize the training process & test result.
- Prepare for the presentation.
- Finalize the report.
Optimal power flow is used in power system operational planning to estimate the most economical efficiency solution while satisfying demand and safety margins.
- Due to increasing uncertainty and variability in energy sources and demand, the optimal solution needs to be updated near real-time to respond to observed uncertainty realizations.
- The existing methods, such as affine control policy [2][3][4] and ensemble control policy [6][7], could not cope with frequent updating due to the high computational complexity.
- Propose the use of neural network as a classifier to enable extremely low computational complexity compared to the traditional method.
- Choose to learn the mapping from uncertainty realization to active constraints set at optimality instead of directly map to the adjustment in the generation, which will guarantee the satisfactory performance of the neural net classifier.
Fig. 1: Traditional Methods vs. The Proposed Method.
Instead of focusing on active constraints classification (softmax output layer), we set up a model that determines which of the given constraints is active (multi-label binary output layer). This change allows us to predict the status of individual constraints separately, which will be an approach to develop a deeper understanding of various operational patterns, such as clustering of constraints.
For following numerical experiments steps, we will use dataset from the IEEE PES PGLib-OPF benchmark library [8].
- Set up OPF test-cases for the learning.
- Check the distribution of active constraints of the OPF test-cases.
- Visualize the active constraints distribution.
- Extending the proposed method to AC OPF with non-linear variations.
- Trying to get weight mapping between uncertainty realization and constraints through an attention model, such as Transformer [9]. This will help us to develop tools for operational planning and real-time control [1].
[1] D. Deka and S. Misra. “Learning for DC-OPF: Classifying active sets using neural nets,” 2019 IEEE Milan PowerTech,2019.
[2] B. Borkowska, “Probabilistic load flow,” IEEE Transactions on Power App. Syst., vol. PAS-93, no. 3, pp. 752–759, 1974.
[3] M. Vrakopoulou, K. Margellos, J. Lygeros, and G. Andersson, “Probabilistic Guarantees for the N-1 Security of Systems with Wind Power Generation,” in Probabilistic Methods Applied to Power Systems (PMAPS), Istanbul, Turkey, 2012.
[4] L. Roald, S. Misra, T. Krause, and G. Andersson, “Corrective control to handle forecast uncertainty: A chance constrained optimal power flow,” IEEE Trans. Power Systems, vol. 32, no. 2, pp. 1626–1637, 2017.
[5] L. Roald, S. Misra, M. Chertkov, and G. Andersson, “Optimal power flow with weighted chance constraints and general policies for generation control,” in IEEE Conference on Decision and Control (CDC). IEEE, 2015, pp. 6927–6933.
[6] Y. Ng, S. Misra, L. A. Roald, and S. Backhaus, “Statistical learning for DC optimal power flow,” Jan. 2018.
[7] S. Misra, L. Roald, and Y. Ng, “Learning for convex optimization,” arXiv preprint arXiv:1802.09639, 2018.
[8] The IEEE PES Task Force on Benchmarks for Validation of Emerging Power System Algorithms, “PGLib Optimal Power Flow Bench-marks,” Published online at https://github.com/power-grid-lib/pglib-opf, accessed: April 3, 2020.
[9] Vaswani, Ashish, Shazeer, Noam, Parmar, Niki, Jakob, Jones, Gomez, A. N., Kaiser, Lukasz, Polosukhin, and Illia, “Attention Is All You Need,” arXiv.org, Dec. 2017.