Pattern-based prediction of islanded power grid frequency

As part of achieving the climate goals from the Paris Agreement set by a united world community, the transition from non-renewable energy sources and electrification of transport and industrial sectors is central. The power grid faces new challenges as emerging renewable energy sources such as solar and wind fluctuate more than the more controllable traditional coal, gas, oil, and hydropower. In addition to more fluctuations on the supply side, an increase in emerging consumers like electric cars and data centers strengthens the need for good tools to maintain grid stability. Therefore, forecasting models and analysis to investigate both large and decentralized power systems’ dynamics are necessary to maintain control and reliability. Expanding upon a pattern-based prediction model called the Weighted-nearest neighbors (WNN) predictor, this thesis investigates the predictability of the power grid frequency for different European islanded power grids through several approaches. The selected islands include Ireland, the Balearic Islands, Iceland, and the Faroe Islands, with the Nordic region as a basis for comparison. The WNN predictor is successfully applied to all regions and performs 60 min predictions better than average daily profiles except for Iceland with its stochastic behavior. The Balearic Islands are the most deterministic region with precise predictions, while Ireland performs slightly worse than the Nordic region. The Faroe Islands exhibit similar performance to Iceland, but with significantly less data available. With varying population and geographical size, the regions cover the range of possible future grids, consisting of larger synchronous areas to small, isolated island systems and microgrids. All regions exclusively show that predictions improve with more data available. The predictor outperforms daily profiles with about a month of data available and with even less for more predictable regions. When electricity generation time series are included in an extended model approach, the performance slightly increases for parts of the predicted hour for several features, despite low time resolution and partially poor quality of the additional data. This suggests that there is further information to be gleaned from other power grid time series to improve the prediction of the power grid frequency.