Probabilistic model-based predictive control of wind-focused gas-renewable energy systems

Abstract

As we head towards a future with increased energy consumption, a larger penetration of green energy is required in the world’s transition to net-zero emissions. In this transition, it is believed that gas turbines (GTs) and energy storage solutions will be central to support hybrid renewable energy systems (HRESs) in periods with little wind, sun, or rain, resulting in less renewable energy production from wind-, solar-, and hydropower sources. From an investment point of view, integrating more green energy decreases greenhouse gas (GHG) emissions. However, from an operations point of view, integrating more green renewable energy increases complexity due to increased modelling, measurement, and forecast uncertainty. To address these challenges, this dissertation proposes several methods for stochastic dynamically optimising these gas-renewable energy systems based on model predictive control (MPC) and probabilistic models for handling the increased complexity and uncertainty from these renewable energy sources in addition to investigating their impacts. The probabilistic models are used either directly in the MPC for predicting the system or indirectly as a tool for auto-tuning MPC schemes. Though this dissertation focuses specifically on wind energy sources, most of these methods are generalisable to other renewable energy technologies such as solar- and hydropower. The first part of this dissertation focuses on scheduling and controlling these gas-renewable energy systems. Based on data-driven stochastic grey-box models, first-principle probabilistic models of intermittent renewables can be derived to handle model, measurement, and forecast uncertainty. These methods are general and applicable to many different renewable energy technologies, such as solar- and hydropower, including the electricity prices if these energy systems are connected to a grid. Based on these probabilistic models and together with complementarity constraints, simultaneously optimal balancing of exogenous energy demands and accelerating system efficiency can be achieved to minimise total GHG emissions in gas-renewable energy systems with stochastic nonlinear model predictive control (SNMPC). The second part investigates specifically the operations of the wind farms (WFs) at a lower level, given the references from the upper level, which the first part of this dissertation focuses on. This part generally applies to the low-level control of other renewable energy technologies as they rely on black-box methods, though we focus specifically on wind. Similar to the first part, the second part utilises MPC where the focus lies in the idea that there will always be model mismatch no matter the model in the MPC, even for the models used in the first part. Thus, fine-tuning the employed dynamic optimiser is always required. To this end, this dissertation investigates data-efficient black-box Bayesian optimisation (BO) for improving both single- and multi-objective (MO) closed-loop WF control performance by auto-tuning MPCs by means of probabilistic Gaussian process (GP) models. Together, these two parts contribute towards the increased integration of renewable energy sources in today’s energy infrastructure by alleviating some of the complexity and uncertainty arising from including these energy sources. A key aspect of the proposed methods from this dissertation is the emphasis on developing methods suitable for real-life applications outside of simulations. Examples can be seen in the focus on computationally efficient approaches that can be solved in real-time during operations or data-efficient learning since gathering closed-loop data may be prohibitively expensive. Another critical aspect of the proposed methods lies within their interpretability. With the rise of machine learning, the need for interpretability and prediction outside of the training data may prohibit their implementation in the industry. This dissertation has thus only focused on methods based on grey-box first-principle models.