Learning-based short-time prediction of photovoltaic resources for pre-emptive excursion cancellation

There is a growing interest in using renewable energy resources (RES) such as wind, solar, geothermal and biomass in power systems. The main incentives for using renewable energy resources include the growing interest in sustainable and clean generation as well as reduced fuel cost. However, the challenge with using wind and solar resources is their indeterminacy which leads to voltage and frequency excursions. In this dissertation, first, the economic dispatch (ED) problem for a community microgrid is studied which explores a community energy market. As a result of this work, the importance of modeling and predicting renewable resources is understood. Hence, a new algorithm based on dictionary learning for prediction of solar production is introduced. In this method, a dictionary is trained to carry various behaviors of the system. Prediction is performed by reconstructing the tail of the upcoming signal using this dictionary. To improve the accuracy of prediction, a new approach based on a novel clustering-based Markov Switched Autoregressive Model is proposed that is capable of predicting short-term solar production. This method extracts autoregressive features of the training data and partitions them into multiple clusters. Later, it uses the representative feature of each cluster to predict the upcoming solar production level. Additionally, a Markov jump chain is added to improve the robustness of this scheme to noise. Lastly, a method to utilize these prediction mechanisms in a preemptive model predictive control is explored. By incorporating the expected production levels, a model predictive controller is designed to preemptively cancel the upcoming excursions --Abstract, page iv