Asynchronous, distributed optimisation for cooperative agents in a smart grid

This thesis addresses a number of issues present in the modern and future power distribution system where high penetration of distributed generation (DG) and smart sensors change the environment in which power flow must be managed. The shift in balance of power supply from the centralized to the distributed can lead to network constraint breaches such as voltage and frequency limits, fault ride through capability, system security, reliability and stability. Common regulation approaches may be inadequate in addressing these network regulation problems, leading to inefficient use of DG and unnecessary high voltage (HV) grid imports. Furthermore, the increase in intelligent Smart Grid components will lead to the transmission and processing of large volumes of data making the optimal control of a network a more challenging problem. Optimisation methods must take into consideration the increase and distributed nature of data, and account for data synchronization, latency, and privacy issues. However, if these challenges can be overcome, then the increase in the controllability and observability of smart grid components, such as distributed generators, storage and controllable demands, offers great potential for the improvement of network optimality and stability. In this thesis, a set of innovative distributed algorithms are presented that solve the optimal power flow problem of a distribution network featuring advanced nodal monitoring and control of DG, storage and loads. These algorithms exploit the network structure to produce iterative solutions to solve the global optimisation problem. They are carefully developed taking into account the realistic limitations where each node can only exchange information with adjacent neighbours but does not have sufficient information about the other nodes in a large scale system. Two partitioning strategies are considered which aim to improve the structure of the communication and control subsystem in order to better facilitate optimal control. The first strategy measures subnet optimality according to the minimisation of mismatch between DG power and local demand, therefore maximising DG utilisation, minimising line loss, and minimising HV grid imports. The second strategy is based on sets of strongly coupled buses, where the coupling of buses is characterised according to the potential for a change in power at one bus to impact the state estimation error at another, therefore improving solution optimality. Subsequently, a distributed predictive optimal control algorithm is proposed, through the method of approximate dynamic programming, that utilises a central coordinator to improve network state estimation and control sequence optimality. The centrally coordinated solution has the benefit of a near optimal solution without burdening controllers with the high dimensional state of the entire distribution network, but rather utilising only a summary of global information. Improvements to the centrally coordinated scheme are then developed through a fully distributed optimal power flow algorithm that requires no central coordination. The fully distributed approach maintains the reduced computational requirements of controllers but improves on the centrally coordinated configuration by restricting data communication to local neighbourhoods. Three variants of the distributed OPF solution are suggested for application to three distinct scenarios: Optimal DG control in a distribution network, DG optimal control in an islanded distribution network, and optimal power management in a home energy management system. These three approaches address significant issues associated with distributed control. An optimal solution to the global problem is achievable in each case, and iterations of the algorithm are shown to be stable and convergent through the use of an augmented Lagrange formulation. Global information necessary for a feasible solution is shared through the development of a new asynchronous consensus protocol, and a communication protocol is presented to enable instantiation, execution and conclusion of fully distributed optimisation sessions. For each studied approach, algorithms are carefully developed that concisely define the method of application. For each presented algorithm, a reasonable amount of computer simulation is applied to verify their applicability to a range of relevant scenarios. The simulations study the algorithms' convergence and scalability, solutions' optimality in a local sense, and solution feasibility. In each case the simulations successfully demonstrate the presented methods' practicality for the studied scenarios.