AlphaZero-Inspired Reinforcement Learning with MCTS for Onshore Wind Farm Layout Optimisation

Sammanfattning

This thesis applies wind farm layout optimisation as a Markov Decision Process using a reinforcement learning (RL) model with Monte Carlo Tree Search and dynamic layer updates. The modelling approach is inspired by Google DeepMind’s AlphaZero architecture and is developed to investigate whether AI models can support early-stage onshore wind farm siting under complex terrain constraints. To account for wake and turbulence, the model updates dynamic wake and turbulence layers after each sequential turbine placement. Three prototypes were developed and compared against expert industry layouts and a Genetic Algorithm baseline using wind park production estimates, wind speed metrics, turbulence indicators, constraint validity, and runtime. Prototype 0 shows that simpler spatial optimisation can already generate useful first guess layouts. Prototype 1 demonstrates feasible sequential placement with RL-MCTS, while Prototype 2 adds dynamic wake and turbulence modelling. The results show promising prototype-level performance, with several evaluated layouts achieving more than 80% of the industry-based wind park production reference in the selected test and inference cases. However, the more advanced RL-MCTS models remain limited by training depth, computational budget, and generalisation challenges. The thesis concludes that AlphaZero-inspired RL-MCTS is feasible as a decision-support approach, but not yet sufficiently developed to replace existing engineering workflows.