Estimation of Load Profiles for Secondary Substations

Abstract: The power system today is facing a transformation from fossil and nuclear energy sources to renewable energy sources such as solar- and wind power. At the same time electric vehicles are becoming more common every year. To analyse how these two changes affect the power system it is crucial to understand how loaded the power system is today from traditional loads on a local and regional level. The purpose of this thesis has been to create the best possible representation of today’s electricity consumption. This is done by generalising and aggregating hourly collected samples of load from energy meters used at low voltage customers of Kraftringen in Lund. The estimation is done on medium voltage level substations by collecting publicly available features which include information about the expected electricity consumption. The resulting model has combined the strengths of linear regression and artificial feed forward neural networks. The model has a mean absolute percentage error of only 10% when evaluated on unseen data from stations used in training the model and 16% when evaluated on an entirely unseen station. The model has been compared to different implementations of the load curve method (typkurvor) which it outperforms with a mean absolute error 54% smaller than the best load curve implementation. The results are based on data from residential districts only and therefore the accuracy of the methods are limited to residential districts. One of the major advantages of this model is that it should be able to predict the electricity consumption from unseen residential districts with only feature data, no data of the electricity consumption from unseen areas is needed. Besides modelling hourly values of load from substations, an extreme value theory model has been used to model the expected maximum loads that occur for one station. This is done with a combination of block maxima of two week-period blocks and by using the Generalised Extreme Value distribution. The resulting model covers all observed load maxima when including confidence intervals of 95% and can be used to predict the expected maximum load for a given time period. The models shown in this thesis can be used by researchers and utility companies to generate expected load of substations and also to model extreme values of load.