Built on research at TU Delft.

Wake models come in different varieties: fast, analytical models that can be used for wind farm optimization and high-fidelity models that can compute wake effects in great detail but are too slow to be used in optimization loops. Graph Neural Networks, trained on high-fidelity model data, offer a way out of this impasse.

Energy system optimization models can be used to model electricity prices, but are known to produce “too flat” price curves. This thesis explores various techniques to get more realistic prices.

A power system model coupled to a grid-connected electrolyser and storage, with a new method for mapping near-optimal regions. The results show how flexible hydrogen consumption and storage investment complement one another.

Wind farm wake effects are usually ignored in energy system optimization models. Embedding them in the optimization model yields a powerful tool to investigate the value of energy system aware wind farm control, letting the wind transport the energy to where it is needed.

This work includes the effect of curtailment on the lifetime of the wind turbines in an optimization model. The model therefor seeks to balance day-ahead income, negative prices and the marginal costs of curtailment.

This work proposes a method to combine representative days in optimization models for energy system investment. The benefit is that one can work with less periods and thus less computation time, while preserving accuracy.

This thesis proposes new methods for selecting and combining representative periods in energy system investment models. Aspects like feasibility and robustness are discussed and a method for the targeted selection of extremes is shown to remedy some issues that arise for European scale models.

This thesis proposes a framework to accelerate energy system optimization models with AI. Neural networks are used to predict feasible solutions of the primals and duals, which are integrated into the Benders decomposition algorithm.

An energy system optimization model in which bilateral transactions and indirect costs are represented. This allows one to model non-technical factors that play a large role in how markets operate.

A wind farm control simulation environment is developed, in which different reinforcement learning strategies are proposed and evaluated. This shows the potential of reinforcement learning for decision making in a complex and dynamic environment.

This work demonstrates how energy system optimization models can be used to model electricity prices and analyse investments. One of the contributions is to use the combination of MGA and stochastic weather years to get electricity price distributions.