Abstract
Most renewable energy sources are dependent on unpredictable weather conditions, which have considerable variation over space and time. The intermittent nature of this production means that any renewable energy prosumer may sometimes produce an amount of energy in excess of its local consumption needs and sometimes in deficiency. This thesis is concerned with developing methods that can improve the effectiveness and widespread adoption of renewable energy usage. In order for renewable energy to be more economically viable, there needs to be a scheme for sharing energy among the prosumers so that those with excess energy can give their excess amounts to those in energy deficiency. That is the task addressed in this thesis.
The way to deal with this problem is to setup an optimal arrangement of local coalitions of renewable energy prosumers such that energy is shared within the coalitions in an optimally efficient manner. As is formally explained early on in this work, finding such an optimal coalition arrangement is an example of a Coalition Structure Generation (CSG) problem. The most straightforward way to find an optimal solution for a given pool of prosumer agents in these circumstances is to examine every possible coalition partition (coalition structure) and evaluate its comparative utility. This is known as ``exhaustive search'' (ES) and can be computationally expensive. As has been shown earlier, the number of such evaluations in ES even for a pool of twenty agents can be in the tens of trillions.
The problem for us in the renewable energy domain is that, because of the constantly changing weather conditions among the scattered prosumers, the CSG optimization calculation must be carried out every hour of the day. This means that the ES approach in the CSG optimization calculation for a reasonable number of prosumer agents is computationally intractable. So a more computationally feasible stochastic optimization method must be used, which searches through the coalition structure search space in order to find a reasonably good solution even if it is not the global optimum.
To this end, a number of stochastic optimization search methods have been investigated in this thesis, including some of our own novel extensions to existing approaches. These search methods have been examined with respect to two different connection arrangements with respect to the outside world – (1) when the local prosumer networks have a connection to a public utility power grid and can therefore buy needed energy (at a high price) from the grid and sell excess energy (at a low price) to the grid and (2) when the local prosumer networks are isolated from any public utility, which is referred to as ``island mode''. The overall goal of these investigations has been to find an optimization approach that arrives at a near-optimal (near the global optimum of the given search space) that is computationally efficient (i.e. it does not require a vast amount of computer memory or running time).
Based on these empirical examinations, which have employed realistic parameters drawn from existing consumption and renewable energy data sets, the following conclusions concerning renewable energy can be drawn from this study:
• It is feasible to employ ordinary computer resources to obtain on an
hourly basis near-optimal energy-sharing coalition structures that will
lead to the more effective and economical use of renewable energy.
• This energy-sharing approach will contribute to more rapid adoption
and proliferation of existing renewable energy equipment and
infrastructure.
The principal contributions towards these end that this thesis work has
made are as follows:
• A modelling framework has been set up that can be used for extensive
empirical determinations of near-optimal energy-sharing coalition
structures.
• A detailed empirical study has been carried out that has examined the
relative capabilities in this context of various optimal coalition structure
search methods, including genetic algorithms (GA), dynamic
programming (DP), particle swarm optimization (PSO), population-based
incremental learning (PBIL), and several variants to PBIL.
• The novel extensions to basic PBIL optimization have included Top-k
Merit Weighting PBIL (PBIL-MW), Set-ID Encoding Schemes, and
Hierarchical PBIL-MW.