Scholarlink

<p>This work presents a data-efficient framework for solar irradiance forecasting based on multi-source early fusion of high-resolution satellite and ground meteorological data. The methodology integrates multiplicative seasonal normalization, quantile-based bias correction tuned via sensitivity analysis and temporal encoding to accurately capture diurnal and annual patterns. A comprehensive benchmark across major regression families including linear, tree-based, support-vector, Gaussian-process, kernel and ensemble models demonstrates the limitations of static feature learners in this setting. The best regressor (linear model) achieves RMSE ≈ 162.5&#xa0;W/m<sup>2</sup> (<i>R</i><sup>2</sup> ≈ 0.76) whereas the proposed neural network time-series architecture reaches RMSE ≈ 50.4&#xa0;W/m<sup>2</sup> and <i>R</i><sup>2</sup> = 0.98 corresponding to ~ 69% error reduction relative to the strongest regressor. A Shapley interpretability assessment confirms the physical relevance of the dominant temporal and atmospheric predictors. Overall, the results establish a lightweight, interpretable and scalable forecasting framework tailored to sparse-data semi-arid environments and suitable for operational smart-grid planning.</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

Transforming Solar Energy Forecasting with Machine Learning and Multi-Source Meteorological Data Fusion for Sustainable and Intelligent Energy Systems

  • 期刊论文

  • 发表日期：

    2026年6月24日

查看全文
  • Leila Aissaoui Ferhi,
  • Balkis Bettoumi,
  • Mounir Ferhi,
  • Ridha Bouallegue

摘要

This work presents a data-efficient framework for solar irradiance forecasting based on multi-source early fusion of high-resolution satellite and ground meteorological data. The methodology integrates multiplicative seasonal normalization, quantile-based bias correction tuned via sensitivity analysis and temporal encoding to accurately capture diurnal and annual patterns. A comprehensive benchmark across major regression families including linear, tree-based, support-vector, Gaussian-process, kernel and ensemble models demonstrates the limitations of static feature learners in this setting. The best regressor (linear model) achieves RMSE ≈ 162.5 W/m2 (R2 ≈ 0.76) whereas the proposed neural network time-series architecture reaches RMSE ≈ 50.4 W/m2 and R2 = 0.98 corresponding to ~ 69% error reduction relative to the strongest regressor. A Shapley interpretability assessment confirms the physical relevance of the dominant temporal and atmospheric predictors. Overall, the results establish a lightweight, interpretable and scalable forecasting framework tailored to sparse-data semi-arid environments and suitable for operational smart-grid planning.