Tractable Awareness of DC Power Networks from Limited Data
Abstract
This research effort intends to develop scalable, near real-time solution strategies for state estimation and topology identificatio,n in DC power electronics networks using limited sensory data. The proposed approach is capable of identifying state and topology in, cases where it is not feasible to equip all nodes and lines of the power network with sensors (due to, e.g., cost, space, or privac,y concerns), which could lead to low-observability conditions and limited available data. Furthermore, this methodology is quite imm,une to untrustworthy measurements (e.g., due to noise or faulty measurements). The joint consideration of both state estimation and,topology identification tasks would typically lead to a non-convex mixed-binary optimization problem. This estimation and identifica,tion problem could become computationally intractable due to the presence of binary decision variables that account for distribution, lines status (and, hence, the network topology identification) and nonlinear relations between sensor measurements and state varia,bles. Lastly, the presence of dynamical loads or power electronics devices necessitates near real-time network observability, partic,ularly to accommodate time-critical tasks. We intend to develop joint estimation and identification techniques for partially observa,ble dc networks and then evaluate the resulting methods in a Controller/Hardware-in-the-Loop (C-HIL) environment. We will develop sc,alable convex relaxation techniques to alleviate the computational challenge associated with this problem. To this end, the joint es,timation and identification problem is first formulated as a non-convex quadratically-constrained quadratic program (QCQP). It is, t,hen, relaxed to a sequence of convex QCQPs whose solutions are guaranteed to converge to the ground truth under minimal statistical,assumptions. Each convex QCQP is polynomial-time solvable and, therefore, leads to a scalable solution. Next, we plan to train a dee,p neural network toact as a surrogate for the proposed optimization framework. This is because optimization algorithms often rely on, an iterative procedure which makes them computationally intensive. Our neural network learns the high-order relation between the se,nsory input and the ground truth state and topology. We will seek a hybrid approach between the neural network and the original opti,mizer. Finally, the resulting relaxed and accelerated solution will be relayed to local controllers of dc power electronics devices,in the power network. The resulting system will be validated on a notional DC network emulated in real time using a C-HIL setup.The,research outcome offers insights into tractable awareness of dc networks from limited available data. DC networks avoid certain issu,es afflicting their AC counterparts and have gained prominence for their distribution efficiency and cable ampacity. Medium Voltage,DC (MVDC) networks are discussed in the 2019 Naval Power and Energy Systems Technology Development Roadmap as well as the IEEE 1709-,22018 standard for MVDC on ships. A power network s state and topology data are conventional prerequisites to the analysis, supervis,ory or machinery control, mission planning, resource allocation, or diagnostic efforts under normal or restorative operations. Line,switching could potentially help with the localization and clearing of faults to improve the survivability and resiliency of MVDC ne,tworks. We stand to fill a fundamental research gap in scalable, near real-time, topology/state monitoring in DC networks with limit,ed data.Approved for Public Release.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 05, 2022
- Source ID
- N000142212555
Entities
People
- Ali Davoudi
Organizations
- Office of Naval Research
- United States Navy
- University of Texas at Arlington