Content area
In an era defined by digital transformation and the exponential growth of big data, industries are increasingly leveraging advanced computational tools to address complex challenges. The energy sector, in particular, has embraced machine learning and optimization techniques to enhance decision-making processes under uncertainty. These data-driven approaches utilize vast datasets and computational power to model, predict, and optimize systems more effectively than traditional methods. This thesis investigates the integration of machine learning into mathematical optimization frameworks, with a focus on constraint learning and surrogate modeling methodologies for addressing stochastic and complex problems in energy systems. By embedding predictive models directly into optimization processes, the proposed solutions aim to improve efficiency while effectively managing uncertainty. Traditionally, optimization under uncertainty has been addressed through frameworks such as stochastic optimization and robust optimization. Stochastic optimization incorporates randomness by modeling uncertain parameters as random variables with predefined probability distributions, supporting risk-neutral or risk-averse decision-making. Robust optimization, on the contrary, ensures feasibility in worst-case scenarios by defining uncertainty sets without relying on probabilistic assumptions. Although these classical approaches have been instrumental, they often adopt restrictive assumptions about uncertainty distributions and can become computationally prohibitive for large-scale problems. The emergence of data-driven optimization marks a paradigm shift, utilizing machine learning models to predict uncertain parameters and embed these predictions directly into optimization problems. We build on this transition by bridging classical frameworks with machine learning techniques to develop data-informed adaptive methodologies. A significant portion of this thesis focuses on addressing endogenous uncertainty, i.e., uncertainty arising from the relationship between decision variables and response variables that must be accounted for within the optimization problem. This is achieved through trained surrogate machine learning models, which fall under the broader field of constraint learning. Additionally, we explore the application of surrogate models in scenarios involving exogenous uncertainty, such as those encountered in classical stochastic optimization. The thesis contributes to several methodological advancements. It extends constraint learning frameworks by embedding quantile-based and distributional predictive models directly into optimization problems to approximate unknown constraints or objectives. These techniques are implemented using linear models, tree-based models, and neural networks with ReLU activation functions, ensuring computational tractability while maintaining predictive accuracy. Furthermore, it generalizes surrogate modeling approaches for two-stage stochastic optimization, enabling decision-makers to incorporate risk measures such as conditional value-at-risk into their analyses. These methodologies are applied to real-world energy challenges to demonstrate their practical utility in enhancing decision-making under uncertainty. The thesis begins by contextualizing the limitations of traditional optimization methods in uncertain environments while highlighting the strong potential of machine learning-based approaches in Chapter 1. It introduces the concept of constraint learning and surrogate modeling and establishes the primary objectives of the thesis. Chapter 2 explores how machine learning models can be embedded into optimization problems to approximate unknown constraints or objectives. It discusses embedding techniques for linear models, tree-based models, and neural networks with ReLU activation functions, while introducing strategies to ensure robustness through trust regions and ensemble embedding. Chapter 3 develops the first energy application, focusing on optimizing heliostat field aiming strategies in concentrating solar power tower plants. A neural network-based surrogate model is employed to balance energy capture and flux uniformity on the receiver while addressing thermal stress issues. Results based on a real plant demonstrate a significant reduction in thermal stress indicators while maintaining high energy capture efficiency. Chapter 4 extends constraint learning methodologies to uncertain environments by introducing chance constraint learning and distributional constraint learning approaches. Chance constraint learning employs quantile regression for probabilistic constraint satisfaction, while distributional constraint learning uses neural networks to model full response distributions for stochastic optimization problems. These innovations are applied in optimizing day-ahead market participation strategies for large energy producers by combining price response learning with distributional constraints to effectively handle market uncertainties in Chapter 5. Another methodological innovation involves a quantile neural network framework for risk-averse two-stage stochastic optimization in Chapter 6. This framework incorporates conditional value-at-risk measures into surrogate modeling, allowing us to obtain heuristic solutions that account for risk aversion. The final application implements this framework to optimize risk-averse operations in smart grids by addressing renewable generation uncertainty through surrogate modeling with quantile neural networks in Chapter 7. The thesis concludes in Chapter 8 by synthesizing its key findings and emphasizing its contributions to data-driven optimization under uncertainty. It discusses limitations such as computational complexity in large-scale applications and identifies future research directions, including exploring the relationship between predictive accuracy and prescriptive performance of machine learning models. By integrating machine learning with mathematical optimization frameworks, this work highlights the transformative prospects of data-driven approaches to enable more resilient and efficient decision-making systems in uncertain environments.