Scholarlink

<p>Accurate active power prediction in photovoltaic (PV) systems installed at extreme altitudes above 3800&#xa0;m a.s.l. faces critical challenges due to non-stationary climate variability, monitoring equipment failure, and limitations in conventional prediction models. This study developed a hybrid stacking metamodel to overcome these barriers through a four-stage process: adaptive preprocessing that reconstructs shifted time series, sequential feature selection (SFS), dimensionality reduction with PCA, and ensemble that integrates Lasso/Ridge mathematical regularization with the nonlinear capabilities of LightGBM/CatBoost. The results demonstrate exceptional accuracy: the LightGBM metamodel achieved R² = 99.9858%, MAE 6.76 and MSE 13.66, outperforming CatBoost and OLS models, with stable convergence in &lt; 50 iterations and minimal training-validation discrepancy evidencing the synergy between robust dimensionality reduction (PCA/SFS) and dual stacking architecture (linear + boosting). Concluding that the LightGBM model is presented as the best option to solve the dual challenge of data fragmentation and environmental complexity in mountainous areas, future research could evaluate the generalizability of this model in other geographical contexts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing photovoltaic power prediction at extreme altitudes using stacking metamodels and dimensionality reduction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月3日

查看全文
  • Saul Huaquipaco,
  • Wilson Mamani,
  • Norman Beltran,
  • Jose Ramos,
  • Vilma Sarmiento,
  • Pedro Puma,
  • Henry Pizarro,
  • Victor Yana-Mamani,
  • Jose Cruz,
  • Reynaldo Condori

摘要

Accurate active power prediction in photovoltaic (PV) systems installed at extreme altitudes above 3800 m a.s.l. faces critical challenges due to non-stationary climate variability, monitoring equipment failure, and limitations in conventional prediction models. This study developed a hybrid stacking metamodel to overcome these barriers through a four-stage process: adaptive preprocessing that reconstructs shifted time series, sequential feature selection (SFS), dimensionality reduction with PCA, and ensemble that integrates Lasso/Ridge mathematical regularization with the nonlinear capabilities of LightGBM/CatBoost. The results demonstrate exceptional accuracy: the LightGBM metamodel achieved R² = 99.9858%, MAE 6.76 and MSE 13.66, outperforming CatBoost and OLS models, with stable convergence in < 50 iterations and minimal training-validation discrepancy evidencing the synergy between robust dimensionality reduction (PCA/SFS) and dual stacking architecture (linear + boosting). Concluding that the LightGBM model is presented as the best option to solve the dual challenge of data fragmentation and environmental complexity in mountainous areas, future research could evaluate the generalizability of this model in other geographical contexts.