Modelling forecast errors for day-ahead electricity market prices

Kurzfassung

How profitable can flexibility providers, e.g. storage plants, operate in the future electricity system? To answer this question it is crucial to understand future electricity market prices. Profit potentials of flexibility providers may, on the other hand, also depend on the short-term forecasting quality of electricity prices. It is therefore important to include electricity price forecasting errors when modelling profit estimates for flexibility options. However, creating artificial errors is not an easy task, since those do not match common distributions, e.g. Gaussian distributions. We aim to fill this gap and thereby provide an approach to generate sophisticated artificial price forecast errors. We present a statistical analysis performed on several years of forecasting data for the German day-ahead wholesale electricity market. Based on this analysis we derive a model to explain the major impacts on the price forecast quality. We extract a statistical distribution of forecast quality changes and finally provide a method to create artificial forecasting errors that can be applied on top of perfect foresight prices. Such can often be found in optimisation models [1], [2] or even simulation models [3] of energy systems. Thus, prices from such models can be enhanced to resemble typical uncertainties with respect to short-term forecasts. The employed price uncertainty model considers the shape of the merit-order curve, a 24 hour cycle for the estimate of residual load and an auto-correlation. We show that the forecasting uncertainty can be explained satisfactorily using these three components. Following a histogram analysis, an exponential distribution for the change of forecasting errors is derived. In order to construct artificial errors, random values are drawn from this distribution and correlated following an observed autocorrelation function. The artificial forecasting errors can be scaled to represent any degree of uncertainty for the actor under investigation. Figure 1 shows an example of such artificially generated forecasting errors. Although the line plot of the artificial and actual errors do not overlap precisely the degree of uncertainty, depending on the depicted hour, is similar in both graphs. Furthermore, their qualitative characteristics match. Descriptive statistics in Table 1 show a very high agreement for different measures of artificial and actual errors. The values for the artificial errors were not perfectly stable although several thousand data points of artificial errors were created. This is expected since artificial errors are created based on random numbers. Therefore, connected values in Table 1 are shown with reduced precision. [1] Y. Scholz, Renewable Energy Based Electricity Supply at Low Costs: Development of the REMix Model and Application for Europe, Universität Stuttgart, 2012. [2] N. Sun, Modellgestützte Untersuchung des Elektrizitätsmarktes: Kraftwerkseinsatzplanung und -Investitionen, Universität Stuttgart, 2013. [3] 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, pp. 1-24, 2017.