Fully Distributed Optimal Power Flow for Low Voltage DC Grids

Abstract

Multiple solutions for solving optimal power flow (OPF) have been employed, most of them using a centralised approach. However, there is another approach where every node is responsible for the local problem in order to reach a solution: the decentralised optimal power flow (D-OPF). In this thesis, the aim was to improve the speed and flexibility of a D-OPF algorithm based on the Consensus and Innovation (C+I) method using physical measurements from a direct current (DC) system. While in previous implementations, the system would work towards finding the solution for one point in time, the suggested implementation works by performing online optimisation, meaning that there is less time required in propagating information around the system and faster solutions are reached. A possible interaction between the physical and optimisation layers was suggested, to make online control feasible. Using current droop control for a DC system, it was possible to react to sudden changes the system and, in the long term, optimise the electrical resources. The improvements in speed were then demonstrated by the simulations results, where the time to reach the optimal solution was reduced, when compared to previous implementations. In order to increase the flexibility of the system, adaptive behaviour for the critical optimisation variables was suggested. To reduce the oscillatory behaviour of the system, some gains were made proportional to rate of change of said variables, meaning that the system didn't have to rely on user determine values in order to converge. It was also implemented a solution to calculate the line resistance between two nodes, further reducing the need for external inputs. These implementations were tested and it was concluded that it improved convergence speeds, while increasing the flexibility pf the system. Finally, a test case, based on a real existing lighting grid, was designed in order to test the algorithm under larger networks. The results showed that for a 25% increase in the size of the network there was no significant increase in the time required to reach a solution, indicating that the system can be scaled further, and might be dependent mainly on the network structure, and not its size.