Predicting Electrical Power Consumption on Yearly Events for Substations: Algorithm Design and Performance Evaluations
Master thesis
Permanent lenke
https://hdl.handle.net/11250/2683504Utgivelsesdato
2020Metadata
Vis full innførselSamlinger
Originalversjon
Langemyr, A. (2020) Predicting Electrical Power Consumption on Yearly Events for Substations: Algorithm Design and Performance Evaluations (Master's thesis). University of Agder, GrimstadSammendrag
Accurate prediction of electricity usage is critical for grid companies in or-der to ensure reliable power supply for their customers. Many factors in-fluence usage patterns, but generally they consist of yearly-, weekly- and daily trends in addition to stochastic noise due to random user behaviour. Besides the above-mentioned cyclic trends, certain yearly events, i.e. events that take place once per year, can affect usage patterns significantly and thus may cause abnormally high or -low power consumption. Therefore, it is in the interest of grid companies to predict the consumption on such events so they can take measures in advance, if necessary. Much effort has been put into developing methods of improving forecasting accuracy through the use of time series clustering in conjunction with the actual prediction algorithm, but the methods’ ability to specifically improve the prediction of power consumption on yearly events has not yet been evaluated. In this the-sis, we are going to utilize machine learning algorithms to cluster electricity usage patterns and predict power consumption for yearly events based on real operational data at the substation level. More specifically, groups of similar usage profiles are formed by a clustering algorithm, and rather than training a prediction model on a single time series, a similar series from the same cluster is appended. In order to extend the prediction model’s training set in this manner, the appended time series is transformed to fit the scale of the initial time series. Our experiments reveal that combining similar time series, thereby introducing additional yearly events to the prediction model’s training set, can improve the accuracy of the load forecast on the event. This approach is also capable of compensating for missing events in the initial time series, when present in the appended-, similarly behaving time series.
Beskrivelse
Master's thesis in Information- and communication technology (IKT590)