Quantification of the forecasting uncertainty in smart grid

Abstract

In the last two decades, the surging proliferation of renewable generations and the inception of competitive electricity markets worldwide have forced the decision makers to reconsider the planning, operation and trading mechanisms in modern power system. For renewable generations, such as wind power and solar photovoltaic power, the power output is characterized by variability and intermittence due the nature of chaotic weather conditions. While for electricity prices in competitive markets, the fundamental reasons behind are much more complex, the net load variability, system congestions, fuel prices and CO2 allowances are always considered as the major contributors to the uncertainty of electricity price. All these factors drove the grid operators and energy traders to seek a powerful forecasting product to aid their decision-making processes. Over the years, extensive works have been carried out on point (or deterministic) forecasts, which only gives one plausible estimate of the future. However, such forecasts are limited as they fail to inform the inevitable error information involved, which is fairly crucial for sagacious decision makings considering diversified uncertainties. This boosts the shift towards a more informative forecast tool under a probabilistic framework in recent years. In a nutshell, the uncertainty needs to be properly quantified as inputs fed into the specific applications of interest in one of the popular forms: quantiles, prediction intervals, PDF/CDF and scenarios. This thesis concerns three types of them, i.e., prediction intervals, PDFs and scenarios, with respect to two vital forecasting tasks in Smart Grid, i.e., prognosis of solar irradiance and market clearing prices. The research background and purpose are presented in Chapter 1. Chapter 2 gives a comprehensive review of the state-of-the-art techniques for the main forecasting activities in Smart Grid (e.g. wind power, solar photovoltaic power and electricity price). Subsequently, inspired by the fundamentals of information granules, a reliable prediction interval construction framework based on temporal granules is proposed for very short-term solar irradiance forecasts in Chapter 3. In Chapter 4, an effective density forecast approach based on ensemble extreme learning machines and a parametric post-processing technique is presented, which gives a full description of the underlying uncertainty involved in the day-ahead forecasts of Swedish market clearing prices. To further facilitate the generation of time trajectories, an efficient covariance structure determination method is developed to model the temporal dependency in the latter part of this chapter. Chapter 5 concludes the whole thesis and indicates the related aspects that can be enhanced and extended in the future.