On the Use of Pareto-Based Features in Meta-learning for Multi-objective Offshore Wind Farm Layout Optimization

• Gustavo J. N. Silva UFBA
• João G. L. S. G. Silva UFBA
• Islame F. C. Fernandes UFBA

Resumo

The Multi-objective Wind Farm Layout Optimization Problem (MoWFLOP) seeks the optimal turbine placement in offshore wind farms, where construct cost minimization and power generation maximization are simultaneously optimized. As MoWFLOP belongs to NP-hard, previous work has focused on metaheuristic approaches to compute high-quality solutions. However, metaheuristics’ results can vary across wind farm instances, making manual algorithm selection impractical. This paper investigates the use of Pareto-based meta-features to guide a landscape-aware method based on meta-learning for automated metaheuristic selection applied to MoWFLOP. To our knowledge, no such investigation exists for this problem. We explore 30 different sampling configurations to extract Pareto-based meta-features from 300 wind farms, balancing computational cost with feature informativeness. We build a dataset for each configuration and train Random Forest models to predict the best-performing algorithm for unseen instances. The analyses comprised the metaheuristics’ results regarding Pareto-compliant indicators, the impact of sampling configurations on feature values, and the meta-learning performance for automated metaheuristic selection.