In English

Time series prediction for the Nordic electricity market

Gabriel Asphult
Göteborg : Chalmers tekniska högskola, 2013. Diploma work - Department of Applied Mechanics, Chalmers University of Technology, Göteborg, Sweden, ISSN 1652-8557; 2013:43, 2013.
[Examensarbete på avancerad nivå]

Trading on the Nordic electricity market depends on accurate predictions about future events of many di erent characteristics. This type of forecasting, with some satisfactory probability of being correct, can be very hard to manage and hence large computational models need to be used. Electricity is a resource of power that is dicult to store which makes it necessary to produce exactly the same amount as will be consumed at all times. This is why certain companies have an obligation to keep all production and consumption under active control by trading the future power on di erent markets. Each model used to estimate future production of various types needs to be optimized to take speci c information into account. Price trends are also very complex to tell in advance which is why accurate models can give an objective view complementing a personal view. Stochastic optimization is a great tool for creating self-learning models that are able to constantly adapt to changes which is needed when trading electricity. The work was carried out by literature studies and implementation that created an arti cial neural network and trained it on historical data about both physical power and nancial contract prices. To make the results easier to use open source software was used during the thesis. Randomly generated time series used as test objects show that some transformation before modeling increases speed and accuracy of forecasting by multiple factors. The best model for hydro power production forecasting reduced the average error by almost 40% from the reference case. When forecasting prices for nancial contracts the best models give correct direction in 62% of the cases studied. Using stochastic optimization to create various models for the Nordic electricity market is helpful and such models may compete with standard regression models that are mostly used today. One big advantage is the possibility to adapt fast when changes occur. The conclusion of this project is that it is possible for a model to learn typical characteristics of the complex behavior of the electricity market. However, such learning requires knowledge about the market in order to guide the process, as well as a great deal of computational power.

Nyckelord: Time series prediction, stochastic optimization, electricity prices, Nord Pool Spot, DinEl, power trading

Publikationen registrerades 2014-01-14.

CPL ID: 192329

Detta är en tjänst från Chalmers bibliotek