Levelized Cost of Electricity (LCOE) Assessment of Bifacial PV Systems in Uribia, Colombia: Integrating Stochastic Simulation and Machine Learning Under Fiscal Incentives

Engineering notes

Engineering notes will be added by the Power for Solar editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为光伏效率、钙钛矿太阳能电池、储能技术、太阳能热利用、BIPV、并网技术等高价值论文补充中文说明。

Original abstract

This study evaluates the techno-economic viability of bifacial photovoltaic (PV) systems in Uribia, La Guajira—the department that holds Colombia’s strongest solar resource, with a mean global horizontal irradiance near 5.6 kWh/m2/day and seasonal peaks above 6.0. Despite this endowment, the country’s installed solar capacity remains far below its potential, largely because developers lack the site-specific financial risk analyses that investment decisions require. To address this, we pair stochastic Monte Carlo simulation with a set of machine learning surrogate models and quantify how Colombia’s Law 1715 fiscal incentives—VAT exclusion and accelerated depreciation—reshape the Levelized Cost of Electricity (LCOE) of bifacial PV systems under realistic climatic variability. Drawing on six years of daily meteorological data, we model bifacial PERC performance under two ground-albedo conditions: the natural site value (α=0.125) and an optimized surface (α=0.30). The results are consistent and encouraging. Under the Law 1715 tax shields, the mean LCOE settles at 0.0588 USD/kWh, and even the 95% Value-at-Risk (VaR) of 0.0638 USD/kWh stays below the prevailing Colombian industrial tariff across every climatic realization evaluated—evidence that the fiscal framework does as much to compress downside risk as it does to lower the average cost. Ground-albedo optimization proved to be the decisive lever: raising α from 0.125 to 0.30 through low-cost surface preparation shortens the payback period to roughly four years and lets the bifacial configuration overtake the cumulative net present value of the monofacial baseline before year seven. The surrogate models tell a complementary story about the structure of the problem. The non-linear algorithms—Support Vector Regression, Gradient Boosting, a Multi-Layer Perceptron and Gaussian Process Regression—reproduce the Monte Carlo response surface almost exactly (R2≥0.99, 0.994–0.997), whereas linear models trail at R2≈0.90–0.94, a gap that quantifies just how strongly the techno-economic drivers of LCOE interact.

Links and sources

• PDF from original source

Comments