Spatial assessment of utility-scale solar photovoltaic siting potential using machine learning approaches: A case study in Aichi prefecture, Japan

Abstract

Optimal spatial planning is crucial for utility-scale photovoltaic (PV) development for efficient energy utilization and the mitigation of land-use conflicts and environmental disruptions. While traditional multi-criteria decision-making approaches often suffer from inherent subjectivity and poor transferability, machine learning (ML) techniques provide data-driven insights into siting potential assessment. However, the localization of model predictions remains a concern when accommodating complex site-specific circumstances. To address these issues, this study proposes an ML-based comparative framework for siting potential assessment while integrating hierarchical regulatory restrictions. Initially, a real-world dataset was established comprising a digitalized inventory of PV locations and sixteen siting criteria from topographic, climatic, environmental, and socioeconomic perspectives. Maximum Entropy (MaxEnt) and random forest models were employed for assessing criteria impact and siting potential through mutual validation. The original model predictions were refined by incorporating three-level restrictive zones retrieved from local legislation. Both ML models showed great predictive accuracy (AUC >0.8) for the test dataset. A varied range of criteria importance was revealed, among which distance to residential areas, annual sunshine duration, slope, land price, and distance to conservation forests showed consistent dominant contributions and nonlinear impacts. Moreover, incorporating local restrictions improved the consistency and interpretability of model predictions, whereas the MaxEnt output exhibited more conservative predictions. Insights from this study enhance the practicability of regional siting potential assessment, contributing to the resource-environment-economy balance for solar energy promotion and facilitating sustainable energy transition. • A machine learning framework is developed for mutual validation of siting potential. • Consistent dominant siting criteria are identified with non-linear impact. • Local restrictions are incorporated and refine the consistency of model predictions. • Siting potential distribution provides a feasible spatial plan for development.

Topics & Concepts