Doctoral Degrees (Electrical and Electronic Engineering)

Permanent URI for this collection

https://scholar.sun.ac.za/handle/10019.1/327

Browse

Browsing Doctoral Degrees (Electrical and Electronic Engineering) by browse.metadata.advisor "Bekker, Bernard"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Characterizing variable renewable energy generation uncertainty towards improved forecasting and operational decision making
    (Stellenbosch : Stellenbosch University, 2022-12) Mararakanye, Ndamulelo; Bekker, Bernard; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: As its first novel contribution, this dissertation investigates the broad challenge of understanding the impacts of integrating a high share of variable renewable energy (VRE) generators into power systems. These generators are variable, uncertain, non-synchronous and location constrained, introducing a wide range of impacts that are unique to a specific region and require different input data, models, and simulation tools to study. It is resource intensive to study all these impacts and therefore important to effectively identify and study only those impacts that are most relevant to the region under consideration. In addressing this challenge, this dissertation firstly identifies three main factors that influence the VRE integration impacts in different regions: available resources, penetration level, and grid characteristics. Thereafter, the international experience is used to understand how these three factors contributed to the issues that were experienced in different regions. The outcome of this investigation is a framework for comprehending VRE integration issues based on available renewable resources, penetration level, and grid characteristics of a region under consideration. This framework can be used by network planners, policymakers, and grid operators to prioritize VRE integration issues of concern to their region prior to conducting detailed studies, thereby reducing the resources required. This dissertation identifies wind power forecasting as key in mitigating some of the impacts introduced by high share of VREs. Within the context of wind power forecasting, this dissertation investigates the challenge of aggregating decentralized forecasts. These forecasts are typically optimized for local conditions because the individual wind farms do not have access to power data from other wind farms. Simply adding these decentralized forecasts together at the point where these forecasts are received (typically the system operator) may not capture some of the common spatial and temporal correlations of wind power, thereby lowering the potential accuracy of the aggregated wind power forecast. In response to this challenge, this dissertation proposes explanatory variables that are used to train the machine learning models to derive aggregated point and probabilistic wind power forecasts from decentralized forecasts. The proposed explanatory variables include clusters of point forecasts (to account for spatial correlations between wind farms), hour of day (to account for diurnal cycles), month of year (to account for seasonal cycles) and, atmospheric states (to account for correlations due to large-scale atmospheric circulations). Training machine learning models using these explanatory variables results in a significant improvement in the accuracy of aggregated forecasts, becoming the second novel contribution of this dissertation. This is particularly important in regions where individual wind farms generate their own forecasts. This dissertation also acknowledges the fact that wind power forecasts are not always perfect, giving rise to the need to understand and estimate wind power forecasting uncertainty. One of the challenges concerning the characterization of forecasting uncertainty is that some of the parametric distributions (normal, beta, Weibull, etc.) commonly used for modeling forecast errors may be inappropriate in representing extreme errors. While non-parametric approaches can be accurate, extreme errors often do not occur frequently enough to make accurate non-parametric inferences. There remains a need to find a parametric model that best represents the extreme errors. To address these challenges, this dissertation identifies a suitable parametric distribution for representing extreme errors, investigates some of the factors that may influence extreme errors, and proposes a suitable model for representing spatial correlations of extreme errors between wind farms. Therefore, the third novel contribution of this dissertation is to propose modeling approaches for improving the estimation and understanding of extreme errors. This is an important step toward better allocation of operating reserves to account for forecasting uncertainty. Continuing within the context of forecasting uncertainty, it is known that the conditional forecast error distributions change with the wind power forecast mostly due to the slope of wind to power conversion curves. The variance is often small at low and high power forecasts but large at mid-range power forecasts. The forecast error distribution is skewed right at low power forecasts, symmetric at mid-range power forecasts, and skewed left at high power forecasts. As a result, some of the commonly used distributions for modeling forecast errors may lack the flexibility required to represent conditional forecast error distributions at different wind power forecasts. As a fourth novel contribution, this dissertation proposed and evaluated an approach for deriving the conditional forecast error distribution for a given wind power forecast. These conditional distributions typically contain more probabilistic information (as compared to unconditional distribution), which can be used to improve reserve allocation in grids with high share of wind generators.