Data-Driven Machine Learning for Performance Ratio Assessment in Sustainable Solar PV Planning on Tropical Highland Terrain

Abstract

Accurate prediction of the Performance Ratio (PR) of photovoltaic (PV) systems in complex mountainous terrain is critical for bankable project planning. However, fixed-PR assumptions systematically underestimate the influence of local topography and tropical microclimates. This study integrates drone-captured micro-topographical data, including slope, elevation, and aspect, with two years of NASA POWER climate records, comprising 730 daily observations, to construct engineered interaction features that capture physically meaningful phenomena. Three machine learning regressors: Artificial Neural Network (ANN), Support Vector Machine (SVM), and Gradient Boosting, were trained and validated on these features to predict dynamic PR values for a 550 Wp monocrystalline module planned for the UNNES Ledek Mountain Science and Techno Park. The ANN model achieved the highest predictive accuracy (R² = 0.97546, MAE = 0.00119), significantly outperforming SVM (R² = 0.94669) and Gradient Boosting (R² = 0.84025). Feature importance analysis confirmed that the GHI × Slope interaction term was the dominant predictor, underscoring the critical yet often neglected role of surface orientation in tropical highland PV assessments. These findings demonstrate that micro-topography-informed machine learning constitutes a robust methodological advance over static PR benchmarking and provides a replicable framework for site-specific PV planning in equatorial highland environments, directly supporting SDG 7 (Affordable and Clean Energy).