The forecasting of future energy consumption and generation is now an essential part ofpower system operation. In networks with high renewable power penetration, forecastsare used to help maintain security of supply and to operate the system efficiently.Historically, uncertainties have always been present in the demand side of the network,they are now also present in the generation side with the growth of weather dependentrenewables. Here, we focus on forecasting for wind energy applications at the day(s)-ahead scale. Most of the work developed is for power forecasting, although we alsoidentify an emerging opportunity in access forecasting for offshore operations. Powerforecasts are used by traders, power system operators, and asset owners to optimisedecision making based on future generation. Several novel methodologies are presented based on post–processing Numerical Weather Predictions (NWP) with measured data, using modern statistical learning techniques; they are linked with the increasingly relevant challenge of dealing with high-dimensional data. The term ‘high-dimensional’ means different things to different people, depending on their background. To statisticians high dimensionaility occurs when the dimensions of the problem are greater than the number of observations, i.e. the classic p >> n problem, an example of which can be found in Chapter 7. In this work we take the more general view that a high dimensional dataset is one with a high number of attributes or features. In wind energy forecasting applications, this can occur in the input and/or output variable space. For example, multivariate forecasting of spatially distributed wind farms can be a potentially very-high dimensional problem, but so is feature engineering using ultra-high resolution NWP in this framework.Most of the work in this thesis is based on various forms of probabilistic forecastingProbabilistic forecasts are essential for risk-management, but also to risk-neutral participants in asymmetrically penalised electricity markets. Uncertainty is always present, it is merely hidden in deterministic, i.e. point, forecasts. This aspect of forecasting has been the subject of a concerted research effort over the last few years in the energy forecasting literature. However, we identify and address gaps in the literature related to dealing with high dimensional data in both the input and output side of the modelling chain. It is not necessarily given that increasing the resolution of the weather forecast increases the skill, and therefore reduces errors associated with the forecast. In fact and when regarding typical average scoring rules, they often perform worse than smoother forecasts from lower-resolution models due to spatial and/or temporal displacement errors. Here, we evaluate the potential of using ultra high resolution weather models for offshore power forecasting, using feature engineering and modern statistical learning techniques. Two methods for creating improved probabilistic wind power forecasts through the use of turbine-level data are proposed. Although standard resolution NWP data is used, high dimensionality is now present in the output variable space; the two methods scale by the number of turbines present in the wind farm, although to a different extent. A methodology for regime-switching multivariate wind power forecasting is also elaborated, with a case study demonstrated on 92 wind balancing mechanism units connected to the GB network. Finally, we look at an emerging topic in energy forecasting: offshore access forecasting. Improving access is a priority in the offshore wind sector, driven by the opportunity to increase revenues, reduce costs, and improve safety at operational wind farms. We describe a novel methodology for producing probabilistic forecasts of access conditions during crew transfers.
Date of Award | 28 Jun 2021 |
---|
Original language | English |
---|
Awarding Institution | - University Of Strathclyde
|
---|
Sponsors | EPSRC (Engineering and Physical Sciences Research Council) |
---|
Supervisor | Jethro Browell (Supervisor) & David McMillan (Supervisor) |
---|