Markov chain optimisation for energy systems (Ext.)

This is an extension of the Fellowship: 'Optimal Prediction in Local Electricity Markets'. In this project we will develop novel approaches to the optimisation of energy systems under uncertainty. Our approach, based on methods of computationally intensive statistics, offers significant advances on multiple fronts relative to the state of the art. Firstly more detailed and appropriate representations of random variations will be made possible, to address the increasingly important question of the integration of renewable power generation. Secondly we will apply cutting-edge approaches in computationally intensive statistics to reduce the computational time required for the optimisation of energy systems under detailed models of uncertainty, and to develop methods capable of scaling up to large power systems. We will work together with both established and start-up energy companies in the UK to maximise the potential impact of our work. The developed methods will be general in their applicability across energy systems and this research will also support the technical development in the UK of heat networks, a potentially efficient method of delivering water and space heating to multiple buildings. Our research therefore offers multiple contributions to the 'Energy trilemma' of delivering affordable, clean and reliable energy and to the COP21 agenda.

Electrical power systems are said to be the largest industrial systems created by humankind. The necessary equilibrium between generation and demand is maintained by a complicated control system, whose modelling and optimisation is a high-dimensional problem, and the UK spends approximately GBP 14 billion per year on power production. While the scale of potential economic impacts is therefore clear, there are also potential environmental impacts. Current industrial approaches to power system optimisation, based for example on mixed integer programming, are appropriate for traditional highly predictable thermal generators and predictable, passive demand. When uncertainty is significant due for example to economic feedback or the complex spatial and temporal properties of wind generation forecast errors, however, deterministic formulations can lead to solutions having both higher real-world operational cost and a greater requirement for capital-intensive and potentially highly polluting balancing reserve such as diesel farms, when compared to stochastic formulations. Such uncertainty is set to increase as UK market incentives sharpen and storage technologies evolve and as power systems incorporate further renewable generation following the recent COP21 Paris Agreement governing carbon dioxide reduction measures from 2020. More powerful approaches to the stochastic optimisation of power systems, based for example on modern computationally intensive statistics, are therefore increasingly important.

The extent of potential impacts from research in mathematical optimisation under uncertainty has been highlighted by a 2011 report of the Department of Energy (DOE, see Alexander (2013), Case for Support). As I understand it this is due to be re-asserted in the coming 2016 report "Analytical Research Foundations for the Next-Generation Electric Grid" of the US National Academies. The DOE report highlights the following research priorities:
* the modelling of non-Gaussian noise processes,
* the development of scalable algorithms to address the interaction of uncertainty with integrality constraints,
* development of methods capable of exploiting problem structures such as networks, and constraints of multiple types.
All of the above priorities are addressed in the present proposal.

There are a large number of particular power systems optimisation problems, with a range of timescales (near real-time security, short-term efficiency, mid-term maintenance and natural resource planning and long-term capital planning) and objective functions (average cost, reliability, emissions). We will provide a methodology based on MCMC potentially applicable in all the above contexts. In this way our work will be directly applicable to questions including the value of flexibility, adequacy of generation reserve levels, and suitability of the generation mix, in power systems and in heat networks.

A key part of delivering these impacts will relate to implementation and software. We will therefore take care to perform algorithm development using standard statistical libraries for MCMC as far as possible, and otherwise to make code available openly on our web pages. Further there are a number of initiatives worldwide promoting both open energy data and modelling (including OpenEI, Open Power System Data, and openmod), and we will pursue opportunities to contribute to these forums. In industrial collaborations we will put in place appropriate symmetric agreements relating to intellectual property (IP), ensuring that outputs from our own research are owned by our institutions while enabling partners to retain their proprietary IP. In this way we will ensure that our own research outputs remain available to be used widely in order to maximise their impact.