The integration of renewable energy sources into modern power systems is a global priority aimed at reducing carbon emissions, decreasing fossil fuel dependency, and advancing sustainable development goals. In this context, wind energy stands out for its efficiency, cost-effectiveness, and suitability across diverse geographic conditions. However, identifying sites for wind farm deployment remains a complex challenge, as it requires integrating spatial, environmental, and socio-economic factors. Traditional methods often face limitations, such as dependence on subjective parameters, high cognitive effort for decision makers, and limited interpretability of results, which can lead to biases and inconsistencies. This study proposes a novel framework based on the Dominance-Based Rough Set Approach – Preference Learning to overcome these limitations. The approach learns from past decisions using dominance principles, generating interpretable decision rules for evaluating new potential sites. It eliminates the need for predefined preference parameters and enables the integration of multiple decision makers through an interactive process, thereby enhancing consistency and reducing cognitive demands. A case study conducted in the Brazilian state of Rio Grande do Norte is presented, producing reliable and explainable knowledge by integrating preferences from three decision makers. Across five iterations, the framework improved collective sorting performance by reducing unsorted areas from 8% to 4% and increasing class discrimination while preserving fully interpretable decision rules. The tool is adaptable and can be replicated in other territories. Future work will extend the methodology to include additional renewable sources and develop a user-friendly decision support system for broader application. • Preference-learning framework supports wind farm site assessment • Applied to 216 wind projects in Rio Grande do Norte, Brazil • Three decision makers built collective decision rules across five iterations • Unsorted areas fell from 8% to 4%, improving classification coverage • Results are interpretable and reproducible
Ferreira et al. (Wed,) studied this question.