Distributed Energy System Design including Unbalanced AC Power Flow for Large LV Networks with ADMM

TL;DR

This paper develops a distributed optimization approach using ADMM for designing large low-voltage networks with unbalanced AC power flow and DERs, improving computational efficiency.

Contribution

It introduces a novel decomposition strategy combining MILP, NLP, and ADMM to efficiently solve complex DES design problems with unbalanced power flow.

Findings

ADMM achieves up to 13x speed-up with parallel computation.

The method handles networks with up to 55 loads and 120 timepoints.

Maximum optimality gap observed is 0.61%.

Abstract

With the addition of large numbers of distributed energy resources (DERs) to distribution networks comes the increasing risk that their operation may violate the safety constraints of these networks. The problem considered in this paper is that of combined siting, sizing and dispatch of these DERs, also known as distributed energy system (DES) design, to help meet electrical and heat loads within the network. Here, the operation of these DERs is modelled, along with the unbalanced three-phase alternating current (AC) power flow in the network. When this network power flow is considered, this admits a non-convex mixed-integer nonlinear program (MINLP) model formulation which scales poorly with network size in terms of solve time. To address this, the problem is decomposed into a series of algorithmic steps, starting with a mixed-integer linear program (MILP) formulation that does not consider network constraints, then fixing binary variables, adding power flow constraints and solving as a nonlinear program (NLP) and finally removing operational binary variables and replacing them with a complementarity reformulation. As the main contributors to the overall solve time, the NLP and Complementarity steps are solved using a hybrid spatial/temporal decomposition strategy and the alternating direction method of multipliers (ADMM) distributed optimisation method. Results are presented for networks based on the European low voltage test feeder with up to 55 loads and 120 timepoints, with the ADMM approach showing speed-ups of up to 13x when considering parallel computation of the subproblems, for a maximum observed optimality gap of 0.61%.