Cascade Predictive Model for Photovoltaic Solar Energy Generation Based on Machine Learning With Integration of Meteorological and Operational Variables: A Case Study in Lima, Peru

Huamani J.A.R.; Palacios J.A.V.; Sanchez-Ccoyllo O.; Ventocilla C.L.R.; Sequera J.L.C.; Pulido J.M.G. · IEEE Access (2026) · DOI: 10.1109/access.2026.3670286

Resumen

This study presents a cascaded predictive model for estimating photovoltaic solar energy generation, structured in two sequential Machine Learning (ML) stages. In the first stage, supervised algorithms are trained to predict global solar radiation from 13 meteorological and temporal variables. In the second stage, the estimated radiation is used as input to predict the power and energy generated by a monocrystalline solar panel using a Multilayer Perceptron (MLP) model. Meteorological data were collected during 2024 at a station located south of Lima, Peru. After rigorous preprocessing (outlier detection, variable transformation, and scaling), the MLP model - optimized using GridSearchCV and K-Fold cross-validation - achieved outstanding performance, with an R2 of 0.947 and a MAE of 44.85. The energy was estimated hourly and validated against actual measurements. The results demonstrate that this cascaded approach enables reliable estimation even in environments without direct irradiance sensors, making it an effective tool for energy planning and photovoltaic system optimization in urban contexts.

Instituciones: Universidad Nacional de Ingeniería; Universidad Nacional Tecnológica de Lima Sur