Developing Neural Network Schemes for Optimal Power Flow Problems: Universal Solver and Unsupervised Training

Integrating renewable energy sources at large scale has been a prioritized focus in the recent development of power systems, as a prime effort to fight climate change. In 2020, renewable generation represented 29% of the global electricity generation, exceeding 7,468 terawatthours [1]. Further, these numbers are still multiplying. With increasing uncertainty from renewablegeneration and load variation due to demand-side innovation, optimal grid operating points, i.e., generator dispatch decisions, change rapidly and irregularly. As such, grid operators need to solve the essential optimal power flow (OPF) problems more frequently than before, to track the operating points for reliable and efficient real-time grid operation. However, the OPF problem with a full AC power flow formulation (AC-OPF) is non-convex and NP-hard, making it difficult to solve efficiently by iterative solvers, especially for large-scale instances. Recently, there have been promising endeavors in employing DNN to directly solve OPF problems, in a fraction of the time used by iterative solvers. The idea is to leverage the approximation capability of DNNs [2–4] to learn the load-solution mapping of the OPF problem. Then one can feed the load to the DNN to instantly obtain a solution. To date, various studies haveapplied DNNs to generate feasible and near-optimal solutions for popular OPF formulations with a few orders of magnitude speedup as compared to iterative solvers [5–16]. Despite these exciting progresses, two critical issues hinder DNN’s practical applications.• Restricted DNN applicability. Existing methods train one DNN for each OPF problem under a particular system configuration, which specifies the topology, line admittance, generator on/off and various limits. When the configuration, e.g., topology, changes,one needs to retrain the DNN to solve the updated OPF problem. Retraining DNNs in real-time [17, 18], or pre-training a large number of DNNs offline for all possible configurations, incurs significant computational/data complexity and is not practical.• Explosive labeled-data volume. Most existing DNN schemes are based on supervised learning and require a large amount of labeled data to achieve considerable accuracy, which is non-trivial and computationally expensive, especially for large-scale systems.Project Aim and Description We will tackle the two issues and further develop practical DNN schemes for solving OPF problemsefficiently. Our plan is three-fold. First, we will develop an embedded-training method to train one universal DNN to solve multiple OPF problems under flexible configurations, e.g., line admittance variation or generator failure. The idea is to embed discrete configuration representations, e.g., generator on/off, into continuous parameter spaces, e.g., line admittance, and train a DNN to learn the mapping from (load, system parameter) to the corresponding OPF solution. Second, we will develop unsupervised learning strategies to train DNNs, including the universal one, for solving OPF problems, with no or few labeled data. Finally, teaming up with a power system R&D center of a Fortune 500 conglomerate, we plan to carry out simulations based on real-world data and testbed trials to evaluate the performance of our DNN schemes. Preliminary Result and Significance of ProjectWe have developed one DNN for solving multiple AC-OPF problems with flexible topology [19]. We have also designed DNNs for solving individual AC-OPF problems by unsupervised learning [20]. We will build upon the initial success to fully develop theoretical foundation and universal DNNs for solving OPF problems under flexible configurations efficiently with no or few labeled data, so as to facilitate the development of carbon-neutral power systems.