Short-term wind power forecasting: standardization, evaluation and optimization of prediction uncertainty

Abstract

This thesis reports on research aimed at enhancing our understanding on wind power forecasting by investigating and proposing standards to evaluate such forecasts, quantifying the main sources of forecast error, and optimizing the currently available data and cutting-edge technologies to reduce the impact and energy consumption of the developed algorithms. These research objectives have been achieved by means of state-of-the-art statistical models, evaluated using high resolution data collected at Irish wind farms. Improved wind power forecasts play an essential role in reducing the operation \& maintenance (O\&M) cost of wind farms, and subsequently increasing the competitiveness of wind energy with respect to more polluting non-renewable sources of energy. The importance of this tool has lead to very active and fruitful research efforts over the years in terms of modelling. However, the impact of this research has been partially diminished due to the lack of standardization in terms of model evaluation and uncertainty quantification. Thus, measures and guidelines are to be developed to unlock the potential of wind power forecasting as a tool to reach a low-carbon future by reducing the cost of wind energy.

Firstly, lack of benchmarks for assessing the performance of wind power forecasting models has lead to a scenario where many methodologies have been applied to the modelling of wind power forecasts, but evaluated under diverse criteria with datasets of different nature and quality. Thus, our first goal is to explore the use of performance evaluation metrics to quantify model performance, and broaden their original purpose to evaluate aspects often disregarded such as the robustness of any wind power forecasting model over varied wind power generation scenarios. These concepts are later extended to propose guidelines on how to evaluate statistical wind power forecasting models.

Another key aspect often overlooked is our limited understanding on wind power forecast errors. Even if we do know the main existing sources of errors, we do not have the tools to evaluate and quantify them during the model development stage. To solve that, we have developed a simulation-based statistical framework which allow us to effectively determine such quantities.

One last contribution of this thesis stems from the fact that not only we pursue to improve the forecasting skill of our algorithms, but to use the data and the current technology in such a way that we reduce our carbon footprint. Thus, this contribution is developed following two lines of thought: first, leveraging the use of high resolution turbine-level data with the use of clustering algorithms, aiming to find a middle ground between forecasting accuracy and computational cost, and second, developing algorithms tailored for cutting-edge platforms such as the case of neuromorphic devices, a brand new technology inspired by the energy-efficient nature of the brain.