Model in model: Electricity price forecasts in agent-based energy system simulations

Kurzfassung

Energy system models are powerful tools to evaluate current and possible future electricity systems. Especially agent-based modelling (ABM) opens up the possibility of assessing the profitability of technologies within a future energy system while taking into account actor behaviour and uncertainty. The latter two aspects are particularly relevant for the assessment of flexibility options, e.g. energy storage. Since future energy systems will likely comprise rising capacities of flexibility technologies, it is increasingly important to depict their dispatch in energy system models. In order to optimize their dispatch, however, flexibility options need (reasonably accurate) price forecasts. In addition, scientific evaluations of monetary potentials of flexibility options and their effect on the total system require inducing a controlled amount of uncertainty into otherwise accurate price forecasts [1]. In the real world the challenges to obtain a good price forecast are uncertainties due to short-term weather fluctuations, unexpected changes in load or conventional power generation. In simulated energy systems, however, knowledge on load, renewable power generation and availability of power plants can be perfect. Thus, a perfect price forecast could be given to a single flexibility option. For an energy system with multiple competing flexibility options, on the other hand, their interplay cannot be modelled easily and thus accurate electricity price forecasts are hard to obtain. A common approach to model the competition among the market participants is using game theoretic methods. In most cases, however, both analytic and iterative solutions to such game theoretic problems require a significant computational effort. We propose an alternative approach that bypasses the explicit modelling of competition. Instead, a machine learning model based on Long Short-Term Memory (LSTM) networks aims to predict the influence of the competing actors on the electricity prices. This price-prediction model is integrated in the energy system ABM and influences the behaviour of the competing actors in turn. The 'model in model' concept is applied to an energy system ABM. The prediction model receives information about the state of the energy system (e.g. the previous and future residual load and previous electricity prices) and applies its LSTM network in order to provide forecasts to agents managing flexibility options. These can use the provided forecast for their schedule optimization based on their technological constraints and market strategy. In order to minimize the errors, the LSTM needs to predict the impact of the flexibility option's dispatch onto the spot market prices and thus generates a self-fulfilling prophecy. This concept will be integrated into the existing ABM AMIRIS [2], which has its current focus on the German electricity spot market. The extension will allow modelling the market participation of multiple flexibility options and their competition with reduced computational effort. References: [1] C. Schimeczek and F. Nitsch, "Modelling forecast errors for day-ahead electricity market prices," in 8th INREC 2019 - Uncertainties in Energy Markets, Essen, 2019. [2] M. Deissenroth, M. Klein, K. Nienhaus, and M. Reeg, "Assessing the Plurality of Actors and Policy Interactions: Agent-Based Modelling of Renewable Energy Market Integration," Complexity, vol. 2017, 2017.