Bachelor's Thesis Harmanvir Singh
Development of machine learning based forecasting methods for grid state estimation.Copyright: EBC
As the production of renewable energies advances, new challenges arise in the energy supply sector.
Distribution network operators are using load forecasts to tackle the increasing uncertainties and the shortening of planning horizons. The aim is to model a forecast using a machine-learning algorithm. The model should provide a load prognosis for the following day in an hourly resolution. The data of the distribution network in Skane will be used.
Suitable algorithms were identified and selected in the research. This showed that regression models are particularly suitable for load forecasting. SARIMAX is a promising model due to the requirements of the prognosis.
The implementation included a comprehensive data analysis and a subsequent parameter identification for a SARIMAX model. The data analysis consisted of a cluster analysis, feature development and feature selection.
The cluster analysis of the data showed that the load profiles in summer differ from those in winter. In summer, the energy demand of the entire region is lower. For this reason, it is obvious to separate the modelling of the summer and winter months. The feature development was carried out with regard to a critical network condition. It became clear that critical conditions only occur during the day on working days. It follows from the feature selection that the energy production from a nearby offshore wind farm has the greatest influence on the residual load. In addition, temperature, humidity and wind speed proved to be useful features. Parameter identification was used to identify parameters for winter and summer data and to implement individual models.
The validation of the models showed that they were suitable for forecasting working days. Weekend predictions did not yield satisfactory results. The forecasts of the models for weekend dynamics should be improved.