Forecasting price spikes in European day-ahead electricity markets using decision trees

Author(s):  
Anna Fragkioudaki ◽  
Adamantios Marinakis ◽  
Rachid Cherkaoui
2011 ◽  
Vol 60 (1) ◽  
Author(s):  
Gert Brunekreeft ◽  
Roland Meyer

AbstractThe increasing share of renewables in generation leads to a structural change in electricity markets. As the utilization of conventional generators decreases in the short run, fixed costs of those power plants have to be recovered through high price spikes in times of scarce capacities. Moreover, for the longer run, given the technological peculiarities of power supply, the price spikes may not be strong enough to induce adequate capacity investments to ensure an efficient level of supply security. A solution may be to complement the energy-only markets by capacity markets. The latter provide revenues not only for the electricity actually sold but also for available capacity in order to reduce the utilisation risk, in particular for peak load plants that may be dispatched only for a small fraction of the time. Such an approach requires substantial reform of market design though.


2014 ◽  
pp. 195-221
Author(s):  
Rangga Handika ◽  
Chi Truong ◽  
Stefan Trück ◽  
Rafał Weron

2021 ◽  
Vol 13 (1) ◽  
pp. 65-87
Author(s):  
Efthymios Stathakis ◽  
Theophilos Papadimitriou ◽  
Periklis Gogas

Electricity markets are considered to be the most volatile amongst commodity markets. The non-storability of electricity and the need for instantaneous balancing of demand and supply can often cause extreme short-lived fluctuations in electricity prices. These fluctuations are termed price spikes. In this paper, we employ a multiclass Support Vector Machine (SVM) model to forecast the occurrence of price spikes in the German intraday electricity market. As price spikes, we define the prices that lie above the 95th quantile estimated by fitting a Generalized Pareto distribution in the innovation distribution of an AR-EGARCH model. The generalization ability of the model is tested in an out-of-the-sample dataset consisting of 4080 hours. Furthermore, we compare the performance of our best SVM model against Neural Networks (NNs) and Gradient Boosted Machines (GBMs).


Sign in / Sign up

Export Citation Format

Share Document