Neural-network simulation of complex mini- grids
Electrical grids are characterised as systems having many sources, massive numbers of loads, and a complex, non-uniform network connecting the two. Thus, simulation of such a system is computationally intensive. Fast simulation of such systems is a key enabler for optimising them, reducing their cost, and for avoiding catastrophic system failure.
While neural networks have primarily been developed for machine learning, it is also possible to use them to emulate dynamic non-linear systems. Such an approach has been shown to be computationally efficient and can be made even more so by making use of specialised hardware, such as GPUs or TPUs. Furthermore, autoencoder NNs can reduce dimensionality, which is particularly important when dealing with systems with large numbers of degrees of freedom. The use of NNs also makes it straightforward to learn a model from past data. This is particularly important when the detailed characteristics of the components of the system are unknown. For many mini-grids, infrastructure and available expertise is minimal. Thus, expert simulation and optimisation may not be feasible.
This PhD program is intended to bring together ideas from machine learning, dimensionality reduction, physical system simulation, and mini-grid design to develop a method to allow a mini-grid to self-learn, to adjust to a changing physical environment, and possibly to self-repair and thus enable and reduce the cost of this technology for a wide audience of isolated and semi-isolated communities.