Long-term electricity market agent based model validation using genetic algorithm based optimization

Kell, AJM; Forshaw, M; McGough, AS

doi:10.1145/3396851.3397682

Long-term electricity market agent based model validation using genetic algorithm based optimization

Lookup NU author(s): Alex Kell, Dr Matthew Forshaw ORCiD, Dr Stephen McGough

Downloads

Full text for this publication is not currently held within this repository. Alternative links are provided below where available.

Abstract

Impacts on natural and human systems due to climate change have already been observed, with many land and ocean ecosystems having changed. A rise in carbon emissions increases the risk of severe impacts on the world, such as rising sea levels and heat waves \cite{Masson-Delmotte2018}. A study by Cook \textit{et al.} demonstrated that 97\% of scientific literature concurred that recent global warming was anthropogenic \cite{Cook2013}, thus limiting global warming requires limiting the total cumulative global anthropogenic emissions of \ce{CO2} \cite{Masson-Delmotte2018}. Global carbon emissions from fossil fuels have increased since 1900 \cite{boden2017global}. Fossil-fuel based electricity generation sources such as coal and natural gas currently provide 65\% of global electricity. Low-carbon sources such as solar, wind and nuclear provide 35\% \cite{BP2018}. To halt this increase in \ce{CO2} emissions, a transition of the energy system towards a renewable energy system is required. Such a transition needs to be performed in a gradual and non-disruptive manner. This ensures that there are no electricity shortages that would cause damage to businesses, consumers and the economy. To ensure such a transition, energy modelling is often used by governments, industry and agencies to explore possible scenarios under different variants of government policy, future electricity generation costs and energy demand. These energy modelling tools aim to mimic the behaviour of energy systems through different sets of equations and data sets to determine the energy interactions between different actors and the economy \cite{Machado2019}.Optimization based solutions are the dominant approach for understanding long-term energy policy \cite{Chappin2017}. However, the results of these models should be interpreted in a normative manner. For example, how investment and policy choices should be carried out, under certain assumptions and scenarios. The processes which emerge from an equilibrium model remain a black-box, making it difficult to fully understand the underlying dynamics of the model \cite{Chappin2017}. In addition to this, optimization models do not allow for endogenous behaviour to emerge from typical market movements, such as investment cycles \cite{Chappin2017, Gross2007}. By modelling these naturally occurring behaviours, policy can be designed that is robust against movements away from the optimum/equilibrium. Thus, helping policy to become more effective in the real world. The work presented in this paper builds on the agent-based model (ABM), ElecSim, developed by Kell \textit{et al.} \cite{Kell}. ABMs differ from optimization models by the fact that they can explore `\textit{what-if}' questions regarding how a sector could develop under different prospective policies, as opposed to determining optimal trajectories. ABMs are particularly pertinent in decentralized electricity markets, where a centralized actor does not dictate investments made within the electricity sector. ABMs can closely mimic the real world by, for example, modelling irrational agents, in this case, Generation Companies (GenCos) with incomplete information in uncertain situations \cite{Ghorbani2014}. There is a desire to validate the ability of energy-models to make long-term predictions. Validation increases confidence in the outputs of a model and leads to an increase in trust amongst both the public and policymakers. Energy models, however, are frequently criticized for being insufficiently validated, with the performance of models rarely checked against historical outcomes \cite{Beckman2011}.In answer to this, we postulate that ABMs can provide accurate information to decision-makers in the context of electricity markets. We increase the temporal granularity of our prior work \cite{Kell} and use genetic algorithms (GAs) to tune the model to observed data enabling us to perform validation. This enables us to understand the parameters required to observe certain phenomena, as well as use these fitted parameters to make inferences about the future. We use a GA approach to find an optimal set of price curves predicted by generation companies (GenCos) that adequately model observed investment behaviour in the real-life electricity market in the United Kingdom. Similar techniques can be employed for other countries of various sizes \cite{Kell}. Similarly to Nahmmacher \textit{et al.} we demonstrate how clustering of multiple relevant time series such as electricity demand, solar irradiance and wind speed can reduce computational time by selecting representative days ~\cite{Nahmmacher2016}. In this context, representative days are a subset of days that have been chosen due to their ability to approximate the weather and electricity demand in an entire year. Distinct to Nahmacher \textit{et al.} we use a $k$-means clustering approach \cite{forgy65} as opposed to a hierarchical clustering algorithm described by Ward \cite{doi:10.1080/01621459.1963.10500845}. We chose the $k$-means clustering approach due to the previous success of this technique in clustering time series \cite{Kell2018a}. We measure the accuracy of projections for our improved ABM with those of the UK Government's Department for Business, Energy and Industrial Strategy (BEIS) for the UK electricity market between 2013 and 2018. In addition to this, we compare our projections from 2018 to 2035 to those made by BEIS in 2018 \cite{DBEIS2019}.We can model the transitional dynamics of the electricity mix in the UK between 2013 and 2018. During this time there was an ${\sim}88\%$ drop in coal use, ${\sim}44\%$ increase in Combined Cycle Gas Turbines (CCGT), ${\sim}111\% $ increase in wind energy and increase in solar from near zero to ${\sim}1250$MW. We are therefore able to test our model in a transition of sufficient magnitude.