Power-producing units (e.g., hydro power plants) are typically operated in discrete modes—either turned off or running at a minimum level—each with associated costs. Determining the optimal combination of units to operate at any given time is known as unit commitment planning, a key problem in energy systems optimization.
Motivation
In real-world scenarios, the speed of existing optimization algorithms can become a bottleneck, limiting the quality and responsiveness of solutions. This thesis explores how machine learning (ML) can be used to accelerate the optimization process by learning proxy objective functions that approximate the true optimization landscape.
Thesis Goals
- Investigate the use of ML to learn proxy objective functions from data generated by standard optimization algorithms.
- Evaluate how well these proxies can guide faster, approximate solutions to the unit commitment problem.
- Quantify the trade-off between speed and accuracy, including bounds on deviation from optimality for given parameter ranges.
Methodology
- Set up a simplified version of the unit commitment problem suitable for experimentation.
- Use a standard optimization algorithm (e.g., mixed-integer programming) to generate training data.
- Train ML models (e.g., neural networks, gradient boosting) to approximate the objective function.
- Benchmark the ML-based approach against the original optimization algorithm.
Possibilities for extension: Compare to reinforcement learning, explore multiple ML models or architectures, apply the method to other combinatorial optimization problems (e.g., scheduling, routing).
Recommended Resources
Lectures by Pascal Van Hentenryck: YouTube
Paper: Learning to Optimize Combinatorial Problems
Practical information
This thesis can be adapted to either 30 or 60 credits, depending on the scope.
This project is offered in collaboration with the Norwegian Center on AI for Decisions and includes co-supervision from SINTEF.