Machine Learning-Based Renewable and Load Forecasting in Power and Energy Systems

In the past decades, the power electricity industry has undergone tremendous evolution, which ends up with the development and establishment of electricity markets. This transformation breaks up generation services into a separate, more competitive part of the industry, and facilitates advanced techniques, such as the smart grid techniques and the integration of high-penetration levels of renewable energies, which introduce more uncertainty into the systems. To balance the electricity supply and demand at every moment, power system load and renewable energy forecasting have emerged as one of the major research fields in power and energy engineering. The development of the smart grid provides opportunities for accurate forecasting, such as the utilization of machine learning. However, the current machine learning-based forecasting techniques have several nonnegligible deficiencies, such as the over-reliance on single-algorithm machine learning models, the lack of concern of weather effects, and the neglect of heterogeneity between macroscopic superiority and local performance. This dissertation proposes to mitigate power system uncertainty by improving power system forecasting accuracy utilizing advanced machine learning techniques that are capable of providing robust, weather-aware, and widely applicable forecasting services to power system operators. Considering the unique characteristics of wind, solar, and load forecasting, this research develops advanced machine learning-based forecasting methodologies for the three forecasting tasks from different perspectives. Specifically, we first improve short-term wind forecasting accuracy by adaptively ensembling multiple machine learning models (M3) by another machine learning model, and assess the forecastability of wind sites in the United States by this enhanced M3 method. Then, short-term and very short-term solar forecasting methodologies that are aware of different weather conditions and embraces state-of-theart deep learning techniques based on sky imagery are developed. At last, we compare different aggregate strategies in short-term load forecasting and develop a reinforcement learning based dynamic model selection (QMS) methodology that is able to select the best forecasting models at every single forecasting step from a deterministic forecasting model pool or probabilistic forecasting model pool. Numerical simulations show that the developed forecasting models significantly improve forecasting accuracy, which brings benefits to various power system individuals.